On Thu, 22 Sep 2011 02:41:05 +0100, Irwin Zaid <irwin.z...@physics.ox.ac.uk> wrote: > Hi all, > > I have a quick question about how long one should expect it to take to > regenerate kernels that have already been compiled (and therefore > presumably cached). > > Some quick tests indicate that something like 40 ms is common when using > a SourceModule object for compilation. Is this about right? It seems > quite slow to me.
I've never measured this on the assumption that nobody would be limited by that rate. Are you using #include in your code? If so, that would likely slow down the cache retrieval because it needs to see if the included files have changed. In addition, each cache retrieval does go to disk. Finally, the binary has to be accepted by the Nvidia runtime--I'm not sure how much time that takes. I do know that in the worst case this last part might also include a call to ptxas (if you're passing a mismatched binary)--but that shouldn't be happening. If you'd like to improve this time, it'd likely be good if you could measure which of these components is actually using the most time. You can just throw a few time/print statements into pycuda/compiler.py. Once you know, let the list know, and we'll consider how to fix this. HTH, Andreas
pgpnEYkHzRNl1.pgp
Description: PGP signature
_______________________________________________ PyCUDA mailing list PyCUDA@tiker.net http://lists.tiker.net/listinfo/pycuda
