Installing Nvidia’s CUDA 9.1 on Fedora 27

 

If you’re like me, you to want to make sure that you’re using an up-to-date version of your operating system. This is not a problem for the vast majority of software, but when it comes to CUDA, it seems that NVidia is always a bit behind the times and it seems that only end-of-life versions of Fedora are ever officially supported. This doesn’t prevent you from installing CUDA on non-supported versions of Fedora, it just means you have to jump through a few more hoops.

Below you’ll find how I installed CUDA 9.1 on my Lenovo Ideapad Y700 laptop with NVidia Optimus technology. In the near future I’ll update this post to include desktop instructions, though it seems like there shouldn’t be any major issues there.

Update: A few days ago, I finally updated my desktop to Fedora 27 and installed CUDA 9.1 there. The install went quite smoothly, and I was able to build all the samples by simple going into the main sample directory and typing ‘make’. On my laptop, this worked to a degree, but since I had to use bumblebee, it couldn’t locate an NVidia library for one of the samples. On my desktop, this was not an issue. Below, I’ve updated the text to be a bit more general.

1. The first thing you need to do is install NVidia’s proprietary driver.
   1. For an Optimus laptop this means making use of Bumblebee. However, it turns out that CUDA 9.1 requires the 387.26 driver or later, which is a short-lived branch driver and not in the bumblebee managed repository. This means you’ll have to install the unmanaged repository and then manually download the driver yourself from here. Once you’ve downloaded the driver, run the following commands to enable the bumblebee repos

       sudo dnf -y --nogpgcheck install http://install.linux.ncsu.edu/pub/yum/itecs/public/bumblebee/fedora$(rpm -E %fedora)/noarch/bumblebee-release-1.2-1.noarch.rpm
      sudo dnf -y --nogpgcheck install http://install.linux.ncsu.edu/pub/yum/itecs/public/bumblebee-nonfree-unmanaged/fedora$(rpm -E %fedora)/noarch/bumblebee-nonfree-unmanaged-release-1.2-1.noarch.rpm

      Then, copy the downloaded driver to the appropriate directory as mentioned on the bumblebee page

      sudo cp Downloads/NVIDIA-Linux-x86_64-387.34.run /etc/sysconfig/nvidia/

      Finally, install bumblebee

      sudo dnf install bumblebee-nvidia bbswitch-dkms VirtualGL.x86_64 VirtualGL.i686 primus.x86_64 primus.i686 kernel-devel

      Once finished, reboot your system for the changes to take effect.

   2. On a desktop you can do the install in a few different ways. The best options would be to make use of either the RPMFusion or  negativo17 repositories. This way, the driver will stay up-to-date automatically and everything is managed by the package manager. The negativo17 site has instructions for installation, so simply follow as it says. They also have CUDA in their repository, but I have no experience installing it from there. For RPMFusion, once you’ve enabled the repos, the installation of the driver should, most likely just be

      sudo dnf install akmod-nvidia

      If you’re system is quirky like mine, the above repos may not work. In that case you can follow the instructions from “If Not True Then False” to install the driver downloaded directly from NVidia. As mentioned above, make sure you download version 387.26 or later, otherwise you’ll see an error about insufficient driver version when trying to run CUDA code.No matter the method you choose, make sure to reboot after installing and to verify that you are running the NVidia driver you can use the following commands:

      nvidia-installer --version

      and/or

      glxinfo | grep OpenGL

      The glxinfo output should give the vendor string as NVIDIA Corporation.

2. Next, install some useful libraries

   sudo dnf install libGLU-devel libXi-devel libXmu-devel glut glut-devel
   sudo dnf groupinstall 'C Development Tools and Libraries'

3. Go to the CUDA Downloads page and download the .run file for Fedora 25.
4. Once downloaded change directories to Downloads and run

   sudo sh cuda_9.1.85_387.26_linux.run --override

5. Press and hold enter to scroll through the EULA, then type accept.
6. You will be informed that you are installing on an unsupported configuration, just type ‘y’ to install anyway.
7. You will be asked to install the driver, make sure to answer ‘n’ for no to this question.
8. Answer yes to installing the CUDA 9.1 toolkit. Then answer the remaining questions at your discretion, though I always choose to install the symbolic link and you’ll want to install the samples so that you can test the installation.
9. Once done, navigate to /usr/local/cuda-9.1/include/crt and open host_config.h in your favorite text editor. CUDA will complain if your gcc version is above 6, and Fedora 27 uses 7.2 as of this writing. Change line 119 of host_config.h to something like

   #if __GNUC__ > 10

   to prevent that from being a problem.

10. One last thing to do in order for CUDA to work with gcc 7.2/glibc 2.26 which some may not be comfortable with. There are problems with the floatn.h that cause compilation of CUDA code to fail. To fix this you can edit /usr/include/bits/floatn.h. First make a back-up copy of the file somewhere safe in case you want to undo the changes. Then insert the following just below the #endif on line 35:

    #if CUDART_VERSION
    #undef __HAVE_FLOAT128
    # define __HAVE_FLOAT128 0
    #endif

    For the first time since the install, glibc updated. This overwrote the change I made to this header file and I had to add these lines back. I’m planning on looking into what in CUDA is conflicting with the float128 stuff in glibc to see if there is another work-around that only modifies CUDA specific files.

11. Now, we need to modify our .bash_profile to include the CUDA library locations. Here’s what mine looks like:

     # .bash_profile# Get the aliases and functions
    if [ -f ~/.bashrc ]; then
    . ~/.bashrc
    fi# User specific environment and startup programs
    
    PATH=$PATH:$HOME/.local/bin:$HOME/bin:/usr/local/cuda-9.1/bin
    LIBRARY_PATH=$LIBRARY_PATH:$HOME/lib
    LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-9.1/lib64:$HOME/lib
    CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:$HOME/NVIDIA_CUDA-9.1_Samples/common/inc:$HOME/include
    
    export PATH
    export LIBRARY_PATH
    export LD_LIBRARY_PATH
    export CPLUS_INCLUDE_PATH

    You may not need to have the $HOME/bin, $HOME/lib or $HOME/include in your profile. I include these as I use some custom libraries across a lot of the codes that I write, so having them in a centralized location makes using them in different projects much easier as I can simply link to them. Putting the executables in a the home directory bin folder also makes it so that I can run my codes from any directory without having to point to their location. Essentially, I do all that because I’m lazy.

12. After that, navigate to the directory where you installed the samples. You can try simply typing make. If you are using bumblebee, some samples may not build because certain libraries cannot be found. I tried fixing this on my laptop by including the directories where those libraries exist in my .bash_profile, which caused the GUI desktop environment to fail to load, likely do to conflicts with the libraries needed for the Intel driver. The fix here would be to edit the specific make file for the affected codes to add the appropriate directory for the linker, however, I have yet to test this.

This is a big step forward from CUDA 8 in terms of what needs to be done so that things work. The changes to floatn.h are definitely an ugly hack, and something that I’d rather not have done, but it seems to be okay so far.

Post in the comments below if you were successful with the installation or not, and let us all know of any problems you run into. I’ll do my best to respond and help when I can.