Dnia 2010-09-11, sob o godzinie 18:13 -0400, Andreas Kloeckner pisze:
> On Sat, 11 Sep 2010 15:59:50 +0200, Tomasz Rybak <[email protected]> wrote:
> > Hello everyone.
> > Recently I started looking into cooperation between PyCUDA and OpenGL.
> > It works (I am getting segfaults at the end of my program but I might
> > not be cleaning some object properly).
> 
> Shouldn't happen. Can you send tracebacks of these crashes?

OK. Will try to do simplest code and then send source and core dump.

> 
> > During reading documentation and looking at PyCUDA sources
> > I noticed that PyCUDA uses pycuda.gl.BufferObject, which uses
> > cudaGLMapBufferObject, cudaGLRegisterBufferObject, and 
> > cudaGLUnregisterBufferObject. They are described as depreciated
> > in NVIDIA CUDA reference and programming guide.
> > I would like to try to switch PyCUDA to using
> > cudaGraphicsResource (and other functions from this family)
> > if noone is already doing this.
> > 
> > Questions:
> > 1. Is is possible to do this for 0.94, or should I wait for 0.95?
> 
> I was going to release 0.94 next week (headed to Nvidia's GTC next
> weekend). I'd prefer to wait till 0.95 with anything that breaks API.

Good to hear that. Then I will try to build some Debian/Ubuntu packages
and publish them - maybe in Ubuntu Lauchpad PPA, to allow for easy
installation.

> 
> > 2. Should I try to change existing classes (BufferObject) or
> > create new ones, and existing ones will be tagged "depreciated"?
> 
> Definitely new classes, deprecating the old ones.
> (ps deprecate != depreciate :-])

Sorry.

> 
> > Two possible problems for now:
> > 1. I only have NVIDIA 9400M (NVIDIA ION) GPU with capabilities 1.1.
> 
> I'd be happy to test for you on my GTX260 if that helps.
> 
> > 2. Debian has 195.x drivers and CUDA 3.1 does not work well with
> > those drivers. There is plan to get 256.x drivers, but because of Debian
> > freeze it will not be fast. Now focus is on polishing existing driver
> > packages and on supporting legacy hardware.
> 
> Do the 195 drivers have the new GL sharing API?

Yes, simpleGL program from CUDA SDK 3.0 works without any problems.



-- 
Tomasz Rybak <[email protected]> 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

_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda

Reply via email to