[Openmp-dev] Building and Using LLVM/Clang 7.0 with OpenMP Offloading to NVIDIA GPUs

[Joachim Protze via Openmp-dev]

This is what "module show" lists for our cuda/10 module:

prepend-path	PATH	/usr/local/sw/cuda/10.2.89/bin	:
prepend-path	PATH	/usr/local/sw/cuda/10.2.89/computeprof/bin	:
prepend-path	PATH	/usr/local/sw/cuda/10.2.89/nvvm/bin	:
prepend-path	CPATH	/usr/local/sw/cuda/10.2.89/include	:
prepend-path	LD_LIBRARY_PATH
/usr/local/sw/cuda/10.2.89/lib64:/usr/local/sw/cuda/10.2.89/lib64/lib64	:
prepend-path	LD_LIBRARY_PATH	/usr/local/sw/cuda/10.2.89/computeprof/bin	:
prepend-path	LD_LIBRARY_PATH	/usr/local/sw/cuda/10.2.89/nvvm/lib64	:
prepend-path	LD_LIBRARY_PATH	/usr/local/sw/cuda/10.2.89/extras/CUPTI/lib64	:
prepend-path	LIBRARY_PATH	/usr/local/sw/cuda/10.2.89/nvvm/libdevice	:
prepend-path	MANPATH	/usr/local/sw/cuda/10.2.89/doc/man	:


Hope this helps,

Joachim



Am 01.10.20 um 14:17 schrieb Jonas Hahnfeld via Openmp-dev:
> Am Donnerstag, den 01.10.2020, 06:54 +0900 schrieb Itaru Kitayama:
>> Jonas,
>> Could you tell me how the module file is written, removing
>> system-specific information, but
>> leaving relevant for the CUDA Toolkit installation? The path to binary
>> needs to be in PATH,
>> and libraries in LD_LIBRARY_PATH I suppose, but not sure if there
>> needs to be other than that.
> 
> Yes, that should be the relevant ones (plus maybe CPATH, but CMake
> should add the include directory itself) - as I said nothing special.
> The other variables are not actually required, but are set for user
> convenience like CUDA_ROOT.
> 
>>
>> On Wed, Sep 30, 2020 at 10:54 PM Jonas Hahnfeld <hahnjo at hahnjo.de> wrote:
>>> Am Sonntag, den 27.09.2020, 07:28 +0900 schrieb Itaru Kitayama via
>>> Openmp-dev:
>>>> Jonas,
>>>
>>> If addressing me, please include my email address to make sure it ends
>>> up in my inbox. Otherwise messages sent only to the mailing lists get
>>> filtered into folders that I may read / scroll over from time to time.
>>>
>>>> I’m re-reading your instructions and now am wondering what env vars related to CUDA Toolkit installations are required when building Clang?
>>>
>>> None that I'm aware of, at least I didn't set any in addition to those
>>> by our 'module'. IIRC Clang detects CUDA installations by looking for
>>> the first 'ptxas' in PATH. Unless that changed, make sure that clang -v
>>> prints the right installation and you should be good.
>>>
>>> Jonas
>>>
>>> _______________________________________________
>>> Openmp-dev mailing list
>>> Openmp-dev at lists.llvm.org
>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/openmp-dev


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 228 bytes
Desc: OpenPGP digital signature
URL: <http://lists.llvm.org/pipermail/openmp-dev/attachments/20201002/b1c2936d/attachment.sig>