Fedora 18(64bit) and CUDA 5.0

In case you want to get CUDA 5.0 running on F18, I summarized the necessary steps here


Before you get started, make sure that you don’t have a F18 installation with the “nouveau” driver, but have selected “Install with basic video driver” on the initial installation screen !!!

After following the steps above you may need to do the following to build the examples

  1. The Makefiles don’t pay attention to LD_LIBRARY_PATH or the ldconfig. Assuming you installed to /usr/local, they use the path /usr/local/cuda/lib64 but disregard the /usr/lib64/nvidia/ directory in which the libcuda*so is located.

To fix ad a symlink in the /usr/local/cuda/lib64 directory:

  $  cd /usr/local/cuda/lib64
  $  ln -s /usr/lib64/nvidia/libcuda.so libcuda.so
  1. Run make from the /usr/local/cuda/samples directory as follows:

    cd /usr/local/cuda/samples make -k EXTRA_NVCCFLAGS="–pre-include /usr/local/include/undef_atomics_int128.h"

where /usr/local/include/undef_atomics_int128.h is the header file constructed in the previous post.

The “-k” tells make to continue even if a particular example fails to build, e.g. for examples requiring something you don’t have installed such as mpi.

Thanks for adding the comment. I am not sure why, but (1) was not necessary in my case. However, (2) is really useful as it avoids to manipulate all the Makefiles in the example directories