Dnia 2012-06-01, pią o godzinie 12:59 +0200, Vedran Miletić pisze:
> 2012/5/22 Tomasz Rybak <tomasz.rybak@post.pl>:
> > Dnia 2012-05-22, wto o godzinie 14:14 +0200, Vedran Miletic pisze:
> >> Package: python-pyopencl
> >> Version: 2011.2+git20120508-1
> >> Severity: important
> >>
> >> Dear Maintainer,
> >>
> >> since 2011.2+git20120508-1, python-pyopencl depends on amd-libopencl1, which
> >> conflicts with nvidia-libopencl1. This effectively makes it unusable for
> >> people using it on NVIDIA cards.
> >>
> >> I'm sure there is a good reason for the requirement, but could it be relaxed
> >> somehow? It already requires one or the other by virtual dependancy.
> >
> > For more detailed explanation see NEWS from python-pyopencl.
> > In summary - Debian now contains opencl-headers 1.2.
> > NVIDIA's OpenCL supports only OpenCL 1.1 - which means that package was
> > being built successfully but failed to run any code, failing to find
> > OpenCL 1.2 functions.
> > OTOH AMD library provides OpenCL 1.2 and works well with NVIDIA GPUs.
> > AMD deals with managing OpenCL, and forwards all calls to NVIDIA
> > implementation - hence hard dependency on amd-libopencl1.
> >
> > I shall add another entry in NEWS file to describe this situation and
> > relieve worries of users of PyOpenCL. Any insight what would be
> > the best description from user point of view? I do not want to go
> > into details and bore with my description.
> >
> > Of course if you experience some crashes with PyOpenCL running on Debian
> > please let me know. I have tested this combination (NVIDIA+AMD OpenCL
> > libraries) on both unstable and experimental drivers and have not had
> > any troubles, but maybe I just got lucky ;-)
> >
> > Andreas - I know this is my bug, but maybe it would be good idea to add
> > some description to OpenCL-related packages? I can write some draft
> > (but no earlier than middle of next week - I am busy with fixing
> > PyCUDA) to describe situation with libOpenCL and ICD stuff?
> >
> > Best regards.
> >
> > --
> > Tomasz Rybak GPG/PGP key ID: 2AD5 9860
> > Fingerprint A481 824E 7DD3 9C0E C40A 488E C654 FB33 2AD5 9860
> > http://member.acm.org/~tomaszrybak
> >
>
> Hi Tomasz,
>
> I followed your instructions, and I'm experiencing Segmentation fault
> with pyopencl on
> http://enja.org/2011/02/22/adventures-in-pyopencl-part-1-getting-started-with-python/index.html
>
> $ python main.py
> __kernel void part1(__global float* a, __global float* b, __global float* c)
> {
> unsigned int i = get_global_id(0);
>
> c[i] = a[i] + b[i];
> }
>
> Segmentation fault
>
> Any suggestions?
>
I have done some testing and I also get segmentation fault on some
OpenCL kernels. I haven't got segfault on your sample program.
I have been running PyOpenCL unit tests and:
1. test_clmath.py runs without any segfault
2. test_array.py runs without any segfault
3. test_wrapper.py segfaults in test_get_info (in cl.Image) and test_image2d (in
cl.image_from_array).
Initially I have suspected older amd-libopencl1 version (fglrx 12-4
does not work with X.org 12), but problem remains after updating
to 12-6-beta and rebuilding PyOpenCL. I have tested on 295.53
on NVIDIA ION. It looks like in some cases (not all) NVIDIA OpenCL
segfaults when run using AMD OpenCL management library.
I have tested with ocl-icd, which it ITPd by Vincent Danjean,
but there are some problems with compilation.
I do not know yet how to solve it, but shall let you know
when there is some solution.
Regards.
--
Tomasz Rybak <bogomips@post.pl> GPG/PGP key ID: 2AD5 9860
Fingerprint A481 824E 7DD3 9C0E C40A 488E C654 FB33 2AD5 9860
http://member.acm.org/~tomaszrybak
Attachment:
signature.asc
Description: This is a digitally signed message part