From: Tobias Burnus <tobias@codesourcery.com>
To: Jakub Jelinek <jakub@redhat.com>
Cc: Hafiz Abid Qadeer <abidh@codesourcery.com>,
<gcc-patches@gcc.gnu.org>, <fortran@gcc.gnu.org>
Subject: Re: [PATCH] [gfortran] Set omp_requires_mask for dynamic_allocators.
Date: Mon, 21 Feb 2022 19:47:22 +0100 [thread overview]
Message-ID: <a7dc28fe-9bc9-9cff-1077-7116c6a09eb3@codesourcery.com> (raw)
In-Reply-To: <YhPQQ1XtgsB3bjND@tucnak>
Hi Jakub,
On 21.02.22 18:47, Jakub Jelinek wrote:
> Where ME is involved is
> !$omp requires atomic_default_mem_order(whatever) vs.
> !$omp declare variant ...atomic_default_mem_order(whatever).
Ups, missed that case. (Also because there wasn't 'declare variant' when
implementing 'requires' in Fortran.)
Disclaimer to all of the following remarks: I do not understand context
selectors and their fineprint. Thus, my comments my be completely off:
> subroutine baz
> ...
> interface
> subroutine bar
> end subroutine
> !$omp declare variant (foo) &
> !$omp & match (implementation={atomic_default_mem_order(seq_cst)})
> end interface
> call bar
> end subroutine baz
I concur that in this case, it needs to know the 'atomic_default_mem_order'
of baz. — But that seems to be not implementable using a module as
module m_foo
!$omp requires atomic_default_mem_order(...)
contains
subroutine foo
...
end
end module m_bar
...
subroutine baz
use m_foo, only: foo
...
end
seems to make the 'requires' available - such that it cannot be overridden
via a local 'require atomic_default_mem_order'. And having a 'use m_bar'
then has conflicting declarations. Similar probably with C++ modules,
unless the 'requires' does not propagate. (Does it?)
I find it odd to have only code which works when not using modules.
(Explicitly using the mem_order on 'omp atomic' still works.)
And for the other requires in context selectors, I do not really understand how they
are supposed to get used, either. If any 'unified_shared_memory' or 'dynamic_allocators'
appears (in linked-in code), it is in principle callable – the the run-time library
should then remove all devices which do not support it, possibly only keeping the
host device; for USM, it even has to be present in all compilation units. Thus, just
directly calling the dynamic_allocators/unified_shared_memory should have the same
effect at the end, shouldn't it?
Tobias
-----------------
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
prev parent reply other threads:[~2022-02-21 18:47 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-02-21 14:24 Hafiz Abid Qadeer
2022-02-21 15:50 ` Jakub Jelinek
2022-02-21 17:02 ` Tobias Burnus
2022-02-21 17:47 ` Jakub Jelinek
2022-02-21 18:47 ` Tobias Burnus [this message]
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=a7dc28fe-9bc9-9cff-1077-7116c6a09eb3@codesourcery.com \
--to=tobias@codesourcery.com \
--cc=abidh@codesourcery.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).