On Montag 25 Januar 2010, Alexander wrote: > Precision of time.time() function on windows is about 15ms, > which makes it almost useless for profiling. This function > is used for 'time_kernel=True' launches. > > Is it possible to change it to time.clock(), > which works much better on windows?
Bizarrely, I think the situation is exactly reversed on Unix--time.clock() is awful, while time.time() is pretty good. The reasonable thing to do would seem to be the use of CUDA events that provide good accuracy on all platforms. Can you cook up a patch? Andreas
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ PyCUDA mailing list [email protected] http://tiker.net/mailman/listinfo/pycuda_tiker.net
