This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [gomp4 00/14] NVPTX: further porting
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Alexander Monakov <amonakov at ispras dot ru>
- Cc: gcc-patches at gcc dot gnu dot org, Dmitry Melnik <dm at ispras dot ru>
- Date: Thu, 29 Oct 2015 09:31:37 +0100
- Subject: Re: [gomp4 00/14] NVPTX: further porting
- Authentication-results: sourceware.org; auth=none
- References: <1445366076-16082-1-git-send-email-amonakov at ispras dot ru> <20151021085415 dot GK478 at tucnak dot redhat dot com> <alpine dot LNX dot 2 dot 20 dot 1510211210480 dot 23517 at monopod dot intra dot ispras dot ru> <20151021092431 dot GO478 at tucnak dot redhat dot com> <alpine dot LNX dot 2 dot 20 dot 1510281954360 dot 13239 at monopod dot intra dot ispras dot ru>
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
On Wed, Oct 28, 2015 at 08:19:19PM +0300, Alexander Monakov wrote:
>
>
> On Wed, 21 Oct 2015, Jakub Jelinek wrote:
>
> > On Wed, Oct 21, 2015 at 12:16:35PM +0300, Alexander Monakov wrote:
> > > > Of course that doesn't help the thread-limit-2.c testcase.
> > >
> > > Why not?
> >
> > Because the compiler can be configured for multiple offloading devices,
> > and PTX might not be the first device. So, you'd need to have a tcl
> > test whether PTX is enabled at all rather than whether it is the default
> > device.
>
> My previous response was a bit confused so let me correct myself.
>
> Checking whether PTX is enabled as an offload target is relatively easy in
> libgomp DejaGNU harness: just inspect $offload_targets_s. This helps to XFAIL
> the test that would fail at compile time, but such XFAIL'ing is, like you
> said, undesirable because it drops the test for all offload targets. I'd
> rather provide a dummy 'usleep' under #ifdef __nvptx__. WDYT?
Such ifdefs aren't really easily possible in OpenMP right now, the
preprocessing is done with the host compiler only, you'd need to arrange for
usleep being defined only in the PTX path and nowhere else.
>
> On the other hand, checking whether PTX will be the default device when
> running the compiled test seems non-trivial.
The OpenMP standard indeed does not have a function which would return you
an enum on what offloading device it is run on, it would need to be an
extension (so a header different from omp.h and perhaps gomp_* prefixed
function), or we could just use some OpenACC function for that?
What is easily possible to distinguish is offloading vs. host fallback
(omp_is_initial_device ()), and whether it has shared address space or
separate address space (so HSA vs. PTX + XeonPhi + XeonPhi emul), and
several tests already check those two properties.
> I've updated my patches locally in response to reviews to the last series,
> except where shared memory is involved. Should I resend?
Are there any dependencies in your patch series against stuff still in
gomp-4_0-branch rather than already on the trunk?
Jakub