On 06/24/2013 05:42 PM, Andreas Kloeckner wrote:
Søren Rasmussen<[email protected]> writes:
Sorry for the late reply - stuff came up.
Faulthandler gave me nothing. The driver version is: NVIDIA UNIX x86_64
Kernel Module 319.21 Sat May 11 23:51:00 PDT 2013
Backtrace:
#0 0x00007ffff7def181 in ?? () from /lib64/ld-linux-x86-64.so.2
#1 0x00007ffff7de9176 in ?? () from /lib64/ld-linux-x86-64.so.2
#2 0x00007ffff79ba52f in ?? () from /lib/x86_64-linux-gnu/libdl.so.2
#3 0x00007ffff79ba00f in dlclose () from /lib/x86_64-linux-gnu/libdl.so.2
#4 0x00007fffef0c1d8c in ?? () from /usr/local/cuda/lib64/libcurand.so.5.5
#5 0x00007fffef0c1e95 in ?? () from /usr/local/cuda/lib64/libcurand.so.5.5
#6 0x00007ffff68fb901 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
#7 0x00007ffff68fb985 in exit () from /lib/x86_64-linux-gnu/libc.so.6
#8 0x00007ffff68e1774 in __libc_start_main ()
from /lib/x86_64-linux-gnu/libc.so.6
#9 0x000000000041ba51 in _start ()
Huh, that looks likes it's calling into CURAND (the CUDA random number
generator) during module initialization and crashing while it's doing
that. This feels like it's some incompatibility between your system and
the binaries you're using. My first instinct would be to try and rebuild
PyCUDA from source to see if the problem goes away.
HTH,
Andreas
A normal rebuild didn't work. However, building with the
--no-cuda-enable-curand option did the trick.
-- Soren
_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda
