On Wed, 20 Jul 2011 17:46:23 +0200, Peter17 <peter017(a)gmail.com> wrote:
2011/7/11 Andreas Kloeckner
This is rather odd--your valgrind log looks like
indeed the memory is being
lost from somewhere where numpy is being called directly from Python. Where
did you get you numpy from? Did you build it yourself? Can you try and isolate
the source of the leak by bisecting the code? Try a smaller operator? (e.g.
advection?) No operator at all? No visualization?
I tried with different versions of numpy, from Ubuntu's packages and
installed via easy_install. Then, I totally uninstalled Atlas, Lapack,
Blas and numpy, recompiled Lapack and Blas from the latest sources and
installed a fresh version of numpy from easy_install but I still had
the same problem...
After several days of analysis, I finally abandoned: I installed
Ubuntu 10.10, then Hedge and its dependencies and now, it is working
perfectly. My conclusion is that it is due to a specific version of
one (or more) of the packages of Ubuntu 11.04 but unfortunately, I
could not find which one. I had almost nothing else than the default
packages installed on my system.
Paul and I tracked down a leak in PyCUDA. If you were using GPU support,
then that might've been the source of the leak.