Segmentation faults while running cnn_imagenet_minimal.m about matconvnet HOT 3 CLOSED

Hi, sorry for taking so long to respond...
To me it seems that the issue may be that you are using the system-wide CUDA library (/usr/local/cuda/lib64/libcudart.so.5.0) and linking it against the Matlab's libstdc++ ( /export/rel60_shadow/glue.umd.edu/software/matlab/2014a/Linux/bin/glnxa64/../../sys/os/glnxa64/libstdc++.so.6).

Not 100% sure about it, but maybe try to use the LD_PRELOAD trick to force matlab to link against the system-wide library? E.g. starting matlab with this command (and filling proper path to your system's location of the libstdc++)

LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libstdc++.so.6 matlab

from matconvnet.

Hi Karel,

My LD_PRELOAD consists of the following :
LD_PRELOAD=/usr/local/cuda/lib64/libcudart.so.5.0:/usr/local/cuda/lib64/libcublas.so.5.0:/usr/local/cuda/lib64/libcurand.so.5.0:/usr/lib/libstdc++.so.6

But I get an identical error.

from matconvnet.

Hmm, just did some googling and found these links 1, 2.

So, now it's going to be more a guess than anything, but it seems that what may help is to call gpuDevice() before calling any vl_nn mex functions. It seems that it may be caused by compiling the mex files against different version of CUDA than it is being run with?

from matconvnet.