On mercredi 13 mai 2009, Nicholas Tung wrote:
> Do you need to recompile the kernel because of application requirements or
> because there's some weird caching bug? It should be fast to load cached
> kernels. Also, there's a JIT compiler if you can make the necessary
> modifications in PTX. I haven't used it though.

  In the original version, I created one thread for each computation - this 
was done exactly as in the threading example of pycuda. Since the context (and 
therefore the compiled kernel) is thread-specific, that required a context 
creation and a kernel compilation for each kernel execution. Unless I'm wrong 
and it was possible to compile the kernel independently of the context.

  With the new version the threads are persistent, so the problems went away.

-- 
Vincent Favre-Nicolin                   

CEA/Grenoble       Institut Nanosciences & Cryogénie    -   
http://inac.cea.fr/

Université Joseph Fourier  -  http://physique-eea.ujf-grenoble.fr/

tél: (+33) 4 38 78 95 40           fax: (+33) 4 38 78 51 38


_______________________________________________
PyCuda mailing list
[email protected]
http://tiker.net/mailman/listinfo/pycuda_tiker.net

Reply via email to