On Fri, 27 Apr 2012 22:09:47 -0400, Andrew Cron <andrew.j.c...@gmail.com> wrote: > I'm working on some multi GPU software, and I've run into a snag on > OSX. (Note that this works on Linux.) Here's some sample code: > > import multiprocessing > import pycuda.driver as drv > drv.init() ## COMMENT THIS LINE AND IT RUNS FINE > > if __name__=="__main__": > > def tar(): > print drv.Device(0).make_context() > > thd = multiprocessing.Process(target=tar) > > thd.start() > thd.join() > > This code produces the following error: > LogicError: cuDeviceGet failed: not initialized > > I want to be able to keep the first "drv.init()" so that my code will > play nice with other pycuda software on the host thread. My guess is > the source of the issue is related to how multiprocessing clones the > parent process. It probably does something weird to pycuda. > > Any suggestions? Thanks a ton in advanced for your help!!
Turns out I asked this exact question of Nvidia at some point in late '08: http://forums.nvidia.com/index.php?showtopic=83219 Judging from the replies I got, I don't think this is defined behavior in any sense, so the fact that it works on Linux is pure coincidence. I think you'll have to stick to fork-then-init, unfortunately. Andreas
pgpsJb8LbG8SM.pgp
Description: PGP signature
_______________________________________________ PyCUDA mailing list PyCUDA@tiker.net http://lists.tiker.net/listinfo/pycuda
