Hi Sebastian!

Thank you for your interest, and sorry for the delay in responding.

In a later email you said that you "managed to build the GCC source code
with gfx1153 arch" (AMD GPU).  Nice!  If you have not yet, please run
'make check-target-libgomp', which should produce meaningful tesults (not
many FAILs etc.), and then send me and Tobias privately (compressed) the
'libgomp.sum' and 'libgomp.log' files this produced.

On 2026-03-10T17:40:26-0400, Sebastian Galindo <[email protected]> 
wrote:
> I'm reading the OpenACC spec and I'm currently looking into this reopened
> issue: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93226
>
> I have a question related to reproducing the tests.

As you've seen, in this PR93226, two issues are remaining:
'libgomp.oacc-c-c++-common/acc_memcpy_device-1.c' execution test FAILs
for '-O2', but for NVIDIA GPU code offloading only -- so you won't easily
be able to debug that one with your AMD GPU system.  The other is
'libgomp.oacc-fortran/acc_memcpy_device-1.f90' execution test FAILs
generally for code offloading, for GCC built from the OG15 branch
(devel/omp/gcc-15 Git branch), for which in PR93226, Comment 6 Tobias has
posted as WIP patch.  If you could work on getting an understanding of
the underlying issue, and test Tobias' WIP patch, and/or make it work,
that would be very beneficial for your own learning, GSoC preparation,
and GCC bug fixing, of course.  Again: for that, you'll have to build
your GCCs from the devel/omp/gcc-15 Git branch.  It may also be useful to
compare the GCC trunk vs. OG15 '-fdump-tree-gimple' (I suppose).  As
Tobias wrote, it's expected that OG15 has differeces due to OpenACC
'declare' directive handling.

> - How do you test support for different devices or targets? I assume each
> target or device requires a specific hardware configuration.

For AMD GPU, yes, the GPU code has to be built for the specific AMD GPU
ISA, and therefore the code offloading GCC has to (be built with) support
(for) that.  A number of different AMD GPU ISAs get built by default
(search for "gfx" in 'gcc/config.gcc', 'TM_MULTILIB_CONFIG'.  Then you'll
need to use the correct flags when building offloading code.

For NVIDIA GPUs it's easier, as the PTX code that GCC/nvptx generates by
default runs on all NVIDIA GPUs of roughly the last decade.

> Also I was reading that I could request a machine in
> https://portal.cfarm.net/machines/list/ (to build GCC fast enough) but, How
> do you run the tests? just to know if there is some option to run the
> tests, since the beginner docs always include the build on the beginning of
> the command.

Yes, you may certainly apply for access to the Compile Farm, but indeed
you'd have the issue that for proper testing of code offloading, you'll
have to use a system with GPU attached -- which there are not many.
Currently, there are <https://portal.cfarm.net/news/56>
"New NVIDIA DGX Spark GB10 machines: cfarm107 and cfarm108", and
<https://portal.cfarm.net/news/58>
"New NVIDIA Jetson AGX Thor machine: cfarm109".

Building GCC on one system and running the tests on another also is
possible, but I would suggest to keep it simple in the beginning, so
suggest, that for now, you focus on the AMD GPU system that you already
have.  For quicker turn-around times during development, it helps to trim
down the things that GCC builds, and disable bootstrap builds.  Happy to
go into the details, once you have specific questions.

> Thanks for your time and patience!

Again, thank you for your interest -- and patience with us...  ;-\

Tobias and I are also looking into your
PR114002 "[OpenACC][OpenACC 3.3] Add 'acc_attach'/'acc_detach' routine"
questions that you sent in private email.  Will get back to you, soon.


Grüße
 Thomas


> El lun, 9 mar 2026 a las 12:18, Sebastian Galindo (<
> [email protected]>) escribió:
>
>> Thanks for your reply and your time!
>>
>> Regarding bind: I think the description is somewhat clear in the
>>> specification. For the identifier topic: In Fortran, names are lowercase
>>> and have by default a '_' suffix and, in C++, the function signature is
>>> manged into the name.
>>>
>>
>> Understood. I will go through the documentation and the specification to
>> get more involved.
>> I managed to build the GCC source code with gfx1153 arch since I'm using
>> AMD. If there is any testing or benchmarking I can run on this hardware
>> that would be useful for the project, I would be happy to help.
>>
>> Sebastian.

Reply via email to