Thanks for your help. I was using pyopencl 0.92. Now I'm using the most
recent version pyopencl-2011.1.1 ( I also updated the ATI drivers, to the
latest ones) and the cache is working great. The re-runing time is
completely reduced.

Now I'm seeing a different issue. When I run the program, I get this warning

Warning:W000:Barrier caused limited groupsize

This is shown when the GPU is computing, but not when the CPU is computing.
I was not seeing this before. I don't know if it is a pyopencl issue or a
driver issue.

There is another issue, I'm trying to calculate using an ATI4870, but the
cl_array.sum is not working. It works on my CPU, but not on the GPU. The
error I get is:
...
...
...
LogicError: clEnqueueNDRangeKernel failed: invalid work group size

Is there a way to set the group size for pyfft and for pyopencl.array.sum???

Regards

Ricardo


2011/6/23 Bogdan Opanchuk <[email protected]>

> Hello Ricardo,
>
> Currently there is a problem with nVidia OpenCL drivers (toolkit 4.0)
> — they take significant time to compile sophisticated OpenCL kernels
> (and, in fact, it was even worse in 3.* versions). Not so long ago
> Andreas added caching to PyOpenCL — check if the latest version
> improves your experience; you will still have to wait these several
> minutes, but only the first time you create the plan with certain
> parameters. But there is a caveat — the cache is kept in temporary
> folder and OS deletes it from time to time (after reboots or whatever
> the settings are). I think it is possible to keep the generated binary
> in pickle and reuse it (providing, of course, that you are using it on
> the same system after unpickling), but it requires support from
> PyOpenCL, because at the moment, AFAIK, there is no API to get the
> path to the cached file. Let Andreas correct me if I'm wrong.
>
> Best regards,
> Bogdan
>
> 2011/6/24 Ricardo Amézquita <[email protected]>:
> > I'm not sure this is the correct place to ask, but I couldn't  find a
> better
> > one.
> >
> > I'm using pyopencl, and pyfft to perform some calculations. Everything is
> > working great, but my program is taking quite a long time to generate the
> > FFT plan (a couple of minutes). Does anybody knows if it is possible to
> save
> > the FFT plans (maybe using pickle?) so they can be reused in different
> > program runs?
> >
> >
> > --
> > Ricardo Amézquita O
> >
> > Combustión Ingenieros Ltda
> > Cll 25B # 85B - 76, Bogotá / Colombia
> > Tel: 57-1-2635232
> > Fax: 57-1-4296516
> > http://www.cihologramas.com
> >
> >
> > _______________________________________________
> > PyOpenCL mailing list
> > [email protected]
> > http://lists.tiker.net/listinfo/pyopencl
> >
> >
>



-- 
Ricardo Amézquita O

Combustión Ingenieros Ltda
Cll 25B # 85B - 76, Bogotá / Colombia
Tel: 57-1-2635232
Fax: 57-1-4296516
http://www.cihologramas.com
_______________________________________________
PyOpenCL mailing list
[email protected]
http://lists.tiker.net/listinfo/pyopencl

Reply via email to