From: Tobias Burnus <Tobias_Burnus@mentor.com>
To: gcc-patches <gcc-patches@gcc.gnu.org>,
fortran <fortran@gcc.gnu.org>, Jakub Jelinek <jakub@redhat.com>
Subject: [Patch] OpenMP/Fortran: Use firstprivat not alloc for ptr attach for arrays
Date: Fri, 13 May 2022 19:21:02 +0200 [thread overview]
Message-ID: <e2f799ad-338e-9c38-e1c9-2a4516b07586@mentor.com> (raw)
[-- Attachment #1: Type: text/plain, Size: 1253 bytes --]
Based on sollve_vv's tests/4.5/target_teams_distribute/test_target_teams_distribute_nowait.F90
As discussed, for simple pointers – like here with nondescriptor array,
instead of alloc:a + pointer assign, a firstprivate + pointer assign makes
more sense.
It also avoids the race exposed by the sollve_vv testcase in some constellations.
(The testcase, both as attached and currently in sollve_vv [→ Issue #532], is
invalid if run with 'nowait' as there is a race related to the array - as it
only (un)mapped as (disjunct) array sections via interleaved, concurrently
running target constructs with nowait clause.)
There might be more places where this should/could be done – and in principle,
firstprivate could also be useful for an array descriptor (but not its data);
this could also be explored. (Including whether it should then not be privatized
with shared memory.)
OK?
Tobias
PS: OpenACC is excluded as it does its own firstprivate handling.
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
[-- Attachment #2: omp-firstprivate-ptr.diff --]
[-- Type: text/x-patch, Size: 4169 bytes --]
OpenMP/Fortran: Use firstprivat not alloc for ptr attach for arrays
For a non-descriptor array, map(A(n:m)) was mapped as
map(tofrom:A[n-1] [len: ...]) map(alloc:A [pointer assign, bias: ...])
with this patch, it is changed to
map(tofrom:A[n-1] [len: ...]) map(firstprivate:A [pointer assign, bias: ...])
The latter avoids an alloc - and also avoids the race condition with
nowait in the enclosed testcase. (Note: predantically, the testcase is
invalid since OpenMP 5.1, violating the map clause restriction at [354:10-13].
gcc/fortran/ChangeLog:
* trans-openmp.cc (gfc_trans_omp_clauses): When mapping nondescriptor
array sections, use GOMP_MAP_FIRSTPRIVATE_POINTER instead of
GOMP_MAP_POINTER for the pointer attachment.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/target-nowait-array-section.f90: New test.
gcc/fortran/trans-openmp.cc | 12 +++--
.../target-nowait-array-section.f90 | 56 ++++++++++++++++++++++
2 files changed, 65 insertions(+), 3 deletions(-)
diff --git a/gcc/fortran/trans-openmp.cc b/gcc/fortran/trans-openmp.cc
index baa45f78a0e..eb5870c3bc5 100644
--- a/gcc/fortran/trans-openmp.cc
+++ b/gcc/fortran/trans-openmp.cc
@@ -3312,9 +3312,15 @@ gfc_trans_omp_clauses (stmtblock_t *block, gfc_omp_clauses *clauses,
/* An array element or array section which is not part of a
derived type, etc. */
bool element = n->expr->ref->u.ar.type == AR_ELEMENT;
- gfc_trans_omp_array_section (block, n, decl, element,
- GOMP_MAP_POINTER, node, node2,
- node3, node4);
+ tree type = TREE_TYPE (decl);
+ gomp_map_kind k = GOMP_MAP_POINTER;
+ if (!openacc
+ && !GFC_DESCRIPTOR_TYPE_P (type)
+ && !(POINTER_TYPE_P (type)
+ && GFC_DESCRIPTOR_TYPE_P (TREE_TYPE (type))))
+ k = GOMP_MAP_FIRSTPRIVATE_POINTER;
+ gfc_trans_omp_array_section (block, n, decl, element, k,
+ node, node2, node3, node4);
}
else if (n->expr
&& n->expr->expr_type == EXPR_VARIABLE
diff --git a/libgomp/testsuite/libgomp.fortran/target-nowait-array-section.f90 b/libgomp/testsuite/libgomp.fortran/target-nowait-array-section.f90
new file mode 100644
index 00000000000..7560cff746b
--- /dev/null
+++ b/libgomp/testsuite/libgomp.fortran/target-nowait-array-section.f90
@@ -0,0 +1,56 @@
+! Runs the the target region asynchrolously and checks for it
+!
+! Note that map(alloc: work(:, i)) + nowait should be save
+! given that a nondescriptor array is used. However, it still
+! violates a map clause restriction, added in OpenMP 5.1 [354:10-13].
+
+PROGRAM test_target_teams_distribute_nowait
+ USE ISO_Fortran_env, only: INT64
+ implicit none
+ INTEGER, parameter :: N = 1024, N_TASKS = 16
+ INTEGER :: i, j, k, my_ticket
+ INTEGER :: order(n_tasks)
+ INTEGER(INT64) :: work(n, n_tasks)
+ INTEGER :: ticket
+ logical :: async
+
+ ticket = 0
+
+ !$omp target enter data map(to: ticket, order)
+
+ !$omp parallel do num_threads(n_tasks)
+ DO i = 1, n_tasks
+ !$omp target map(alloc: work(:, i), ticket) private(my_ticket) nowait
+ !!$omp target teams distribute map(alloc: work(:, i), ticket) private(my_ticket) nowait
+ DO j = 1, n
+ ! Waste cyles
+! work(j, i) = 0
+! DO k = 1, n*(n_tasks - i)
+! work(j, i) = work(j, i) + i*j*k
+! END DO
+ my_ticket = 0
+ !$omp atomic capture
+ ticket = ticket + 1
+ my_ticket = ticket
+ !$omp end atomic
+ !$omp atomic write
+ order(i) = my_ticket
+ END DO
+ !$omp end target !teams distribute
+ END DO
+ !$omp end parallel do
+
+ !$omp target exit data map(from:ticket, order)
+
+ IF (ticket .ne. n_tasks*n) stop 1
+ if (maxval(order) /= n_tasks*n) stop 2
+ ! order(i) == n*i if synchronous and between n and n*n_tasks if run concurrently
+ do i = 1, n_tasks
+ if (order(i) < n .or. order(i) > n*n_tasks) stop 3
+ end do
+ async = .false.
+ do i = 1, n_tasks
+ if (order(i) /= n*i) async = .true.
+ end do
+ if (.not. async) stop 4 ! Did not run asynchronously
+end
next reply other threads:[~2022-05-13 17:21 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-05-13 17:21 Tobias Burnus [this message]
2022-05-13 17:44 ` Jakub Jelinek
2022-11-12 9:19 ` Thomas Schwinge
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=e2f799ad-338e-9c38-e1c9-2a4516b07586@mentor.com \
--to=tobias_burnus@mentor.com \
--cc=fortran@gcc.gnu.org \
--cc=gcc-patches@gcc.gnu.org \
--cc=jakub@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).