[petsc-dev] No single precision support in ViennaCL?

[Matthew Knepley]

On Tue, Jan 21, 2014 at 3:15 AM, Karl Rupp <rupp at mcs.anl.gov> wrote:

> Hi,
>
> >     It looks like ViennaCL in petsc doesn't support single precision
>
>>     (which I need cause my (old) GPU doesn't support double precision).
>>     I get the following error:
>>
>>     Cannot use viennacl withOUT double precision numbers, it is not
>>     coded for this capability
>>
>>     Is there any way this can be enabled in the build?
>>
>>
>> This may be an oversight. Can you try adding
>>
>>    self.double = 0
>>
>> to the constructor in config/PETSc/packages/viennacl.py?
>>
>
> Thanks, Matt. Do we want to provide some automatic marshalling to single
> precision if not provided by the device? For 'production use' it doesn't
> really matter, because GPUs today which offer only single precision are
> generally rather low-end and won't provide any substantial benefit over
> CPUs. However, it can be handy for debugging purposes. Any opinions?
>

I think we should since it also a way to decrease the PCI penalty.

  Matt


> Best regards,
> Karli
>
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20140121/c4a09a4d/attachment.html>