changes after n-g-d 194.36.31-4

Tomasz Rybak bogomips at post.pl
Tue Sep 28 18:07:18 UTC 2010


I have checked with 256.53-1. I could check with 195.x, but
would like to avoid messing with drivers to much if it is not
necessary.

Dnia 2010-09-28, wto o godzinie 15:10 +0200, Andreas Beckmann pisze:
> On 2010-09-27 21:05, Tomasz Rybak wrote:
> > libcuda1-dev is used (Build-Depends) by python-pycuda
> > (not yet in Debian, waiting for nvidia-cuda-toolkit).
> 
> Can be fixed in python-pycuda :-) But wait until a new toolkit version
> is in NEW.
> 

OK, already made changes in SVN, will wait with building till
new version.

> > nvidia-libopencl1-dev is used (Build-Depends) by python-pyopencl,
> > which is in testing. I am not sure what will happen when you will
> > try to remove package from testing on which other package depends.
> 
> Thanks for that information. Unfortunately there is no
> 'apt-cache reverse-build-depends' command to check this easily ...

np. That's why I subscribe to NVIDIA packaging list - I trace changes
in drivers/libraries so I know whether they touch my packages.

> 
> So libcuda1-dev can be removed without problems, for the other we need a
> better solution.
> 
> OpenCL Headers are freely available from
> http://www.khronos.org/registry/cl/, so ... let's package them as a
> replacement :-)
> Please try the attached package:
> B-D: opencl-headers, nvidia-libopencl1, ...
> (after removing nvidia-libopencl1-dev you may have to create a
> libOpenCL.so -> libOpenCL.so.1 (or whatever) symlink manually as a
> temporary workaround)
> 

Yes, I had to symlink, and later ldconfig to compile
(forgot that dpkg -P lib* rebuilds library cache).

> If this works for you, I'll ITP and look for a sponsor.
> Or would we need API 1.0 instead of 1.1 for some reason?

Upstream made sure that PyOpenCL works with 1.1.
OTOH NVIDIA drivers support OpenCL 1.1 from 258.x so.
Bindings compile but when trying to use I got:
Traceback (most recent call last):
  File "make_constants.py", line 1, in <module>
    import pyopencl as cl
  File
"/usr/src/cuda/pyopencl-0.92/build/lib.linux-x86_64-2.6/pyopencl/__init__.py", line 3, in <module>
    import pyopencl._cl as _cl
ImportError: /usr/src/cuda/pyopencl-0.92/build/lib.linux-x86_64-2.6/pyopencl/_cl.so: undefined symbol: clEnqueueReadBufferRect

clEnqueueReadBufferRect is new function from OpenCL 1.1.

oclDeviceQuery from CUDA toolkit 3.1 prints:
oclDeviceQuery, Platform Name = NVIDIA CUDA, Platform Version = OpenCL
1.0 CUDA 3.1.1, SDK Revision = 6161726, NumDevs = 1

So I am afraid that we need to stick with 1.0 headers
until you package 260.x ;-)

> 
> > Should I just Build-Depend on libcuda1, nvidia-libopencl1, just like
> > nvidia-cuda-toolkit 3.1-2?
> 
> If you need some cuda headers, you'll probably need B-D:
> nvidia-cuda-dev, libcuda1, ... nvidia-cuda-dev is from
> nvidia-cuda-toolkit, but reorganization has just started, so things may
> move :-)
> 

PyCUDA dynamically compiles CUDA kernels, so it needs nvcc installed.
That's why python-pycuda both Depends and Build-Depends on
nvidia-cuda-toolkit.

In summary - change (with 1.0 headers) should be OK, but
send me the new package for testing.

-- 
Tomasz Rybak <bogomips at post.pl> GPG/PGP key ID: 2AD5 9860
Fingerprint A481 824E 7DD3 9C0E C40A  488E C654 FB33 2AD5 9860
http://member.acm.org/~tomaszrybak
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 836 bytes
Desc: This is a digitally signed message part
URL: <http://lists.alioth.debian.org/pipermail/pkg-nvidia-devel/attachments/20100928/2789ecc3/attachment.pgp>


More information about the pkg-nvidia-devel mailing list