On Mon, Mar 9, 2009 at 09:39, Andreas Klöckner <lists(a)informa.tiker.net>wrote;wrote:
On Donnerstag 05 März 2009, Nicholas Tung wrote:
cc'ing mailing list
On Thu, Mar 5, 2009 at 17:40, Nicholas Tung <ntung(a)ntung.com> wrote:
> Hi Andreas,
> Is there any device emulation support at the moment? What's your
> priority for this? I often get kernel crashes that bring down my entire
> machine, which makes debugging pretty slow.
Device emulation is a feature of the runtime API. As such, PyCUDA doesn't
an easy way to provide access to it. I talked to Joe Stam from Nvidia in
January, and he seemed very interested in providing better driver-level
debugging capabilities for PyCUDA to use, but none of that has materialized
I've had cases like the one you describe, and bisection plus a 'debug
where I can just write random debug data from the kernel has always enabled
to track down what's happening.
Yes, I did fix the problem [through regular debugging] which led me to email
this [and many others before it], but I think host emulation capabilities
would be nice. I just started using Py++ to grab the argument specification
from the global functions in a CUDA file, so perhaps adding a "main"-like
method so the file could be compiled into a Python module wouldn't be out of
the question. The nVidia host emulation is also kinda imperfect -- at least
last time I used it, it had a rather non-random thread exit order out of the
__syncthreads() which led to problems.
Driver-level debugging would be great.