On Freitag 26 Februar 2010, Waldemar Kornewald wrote:
Waldemar Kornewald <wkornewald@...> writes:
> Michael Rule <mrule7404@...> writes:
> > no, I never solved this problem. It only appeared intermittently. My
> > only solution was to reboot the machine once a card got "locked up".
> > I never did find any part of my code that was obviously segfaulting
> > or causing other offense.
> The FAQ once mentioned the same exception, but claimed that it got
> fixed. What was the cause of this exception at that time? Does
> anyone know?
OK, I think the problem must be somewhere in pycuda's memory allocation or
- Are your cards having thermal issues?
- What driver version are you using? Which compiler?
- What version of PyCUDA?
- Does this still happen if you're not running a display off the card
I took all my code and put it in a custom .cu file with a main()
that loads exactly the same data as with pycuda and runs my functions with
it. With this setup, everything works even on the GPU.
Unfortunately, that doesn't say much. Can you share the offending code?
There's one thing I noticed: In my Python code I print some
information right before I launch the kernel, but after I copied all data
to the GPU. Before I see that debug information the whole screen freezes
for almost two seconds. Sometimes this leads to a crash on my current
machine, sometimes not. I also tested the same code on another machine and
there it doesn't seem to cause a crash, but I can still get the freezing.
Did anyone else notice freezes while transferring memory?
This sounds more like a poor interaction between display and compute
code to me.