How to install CUDA on Debian

debian_install_cuda.md

How to install CUDA on Debian 8 (Jessie)

This document describes how to install nvidia drivers & CUDA in one go on a fresh debian install.

Work in progress

Preparations

• Start with a fresh Debian install.

• Download the CUDA driver with the following specificatoins:

  Parameter	Type
  Operating system	Linux
  Architecture	x86_64
  Distribution	Ubuntu
  Version	14.04
  Installer type	runfile (local)

  You can download them directly:

  wget http://developer.download.nvidia.com/compute/cuda/7.5/Prod/local_installers/cuda_7.5.18_linux.run /home/john/Downloads/
  

IMPORTANT NOTE: Cuda binaries (.run file) comes bundle with graphic drivers, eg cuda_7.5.18_linux.run ships with the nvidia 352 graphic drivers. Proceed with the installation of those bundled drivers only if your card is supported. Example, I have a GTX 1080ti which is only supported by the 375.82 and up. Therefore I had to first install the 375.82 graphic drivers, then proceed with the Cuda install and answer NO when asked to install the graphic drivers; then it worked. Proceed at OPTIONAL section further down.

• Change file permission of the installation file:

  cd /home/john/Downloads/
  chmod a+x ./cuda_*.run
  

• Get necessary packages:

  sudo apt-get update
  sudo apt-get install dkms build-essential linux-headers-$(uname -r)
  

• Install 32 bit support:

  sudo dpkg --add-architecture i386
  sudo apt-get update
  sudo apt-get install lib32z1 lib32ncurses5
  

• Blacklist the nouveau driver:

  sudo mousepad /etc/modprobe.d/blacklist-nouveau.conf
  

  Copy the following line into the document:

  blacklist nouveau
  blacklist lbm-nouveau
  options nouveau modeset=0
  alias nouveau off
  alias lbm-nouveau off
  

• Disable the Kernel nouveau:

  echo options nouveau modeset=0 | sudo tee -a /etc/modprobe.d/nouveau-kms.conf
  
  sudo update-initramfs -u
  

• Now reboot the system. If the nouveau was successfully disabled, the screen resolution should be lower.

• Install the following packages:

  sudo apt-get install gcc g++ gcc-4.8 g++-4.8 gcc-4.9 g++-4.9 libxi6 libxi-dev libglu1-mesa libglu1-mesa-dev libxmu6 linux-headers-amd64 linux-source freeglut3-dev
  

Install the nvidia driver & cuda

OPTIONAL (If you have a graphic card not supported by the 352, Eg: the following will work for the 1080ti):

wget http://us.download.nvidia.com/XFree86/Linux-x86_64/375.20/NVIDIA-Linux-x86_64-375.20.run
chmod +x NVIDIA-Linux-x86_64-375.20.run
sudo ./NVIDIA-Linux-x86_64-375.20.run

• Reboot again, start system in init 3. How to boot into init 3: When in starting screen press e to edit start commands, find relevant installation (quiet), add init 3

• Log in as normal user, and activate su:

  su
  

• Navigate to downloads folder (or where ever you put the installer), run cuda installer:

  cd /home/john/Downloads
  sh ./cuda_7.5.18_linux.run
  

• Follow the installer menu. You will be asked several questions:

  Terms and condition? --> accept (If you feel like it...)
  Unsupported version, proceed? --> yes
  Install driver? --> yes (or NO if you already installed proper drivers for your card)
  Install OpenGL? --> yes
  Run nvidia-xconfig? --> 'yes' (This option is new in CUDA 8.0, not sure about the consequences.)
  Install CUDA toolkit? --> yes
  CUDA location? ---> /home/john/cuda-7.5
  Create symbolic link at /usr/local? --> yes
  Install samples? --> yes
  CUDA samples location? --> /home/john/cuda-7.5_samples

• Make sure the output of this command is fine before proceeding further.

  nvidia-smi
  

• Restart the system.

  sudo reboot
  

  This should start up successfully with GUI enabled (at native monitor resolution).

• Add path exports to your .bashrc:

  sudo gedit /home/john/.bashrc
  

  Add the following lines:

  # Manually added by john to enable CUDA:
  export LD_LIBRARY_PATH=/home/john/cuda-7.5/lib64:$LD_LIBRARY_PATH
  export LIBRARY_PATH=/home/john/cuda-7.5/lib64:$LIBRARY_PATH
  export PATH=/home/john/cuda-7.5/bin:$PATH
  

• Change permissions of cuda folders:

  sudo chown -R john /home/john/cuda-7.5
  sudo chgrp -R john /home/john/cuda-7.5
  sudo chown -R john /home/john/cuda-7.5_samples
  sudo chgrp -R john /home/john/cuda-7.5_samples
  

Check the result

Check whether the samples that do not require graphics ouput are working

• Navigate to cuda samples location (as specified during installation):

  cd /home/john/cuda-7.5_samples/0_Simple/simplePrintf
  

• Generate executable:

  make
  

• An executable should have been created. Run it:

  ./simplePrintf
  

For samples that do require graphics ouput, the library paths need to be fixed.

• Naviagte to a sample that does require graphics output, e.g.:

  cd /home/john/cuda-7.5_samples/5_Simulations/nbody
  

• Locate the libary files by running:

  locate libGL.so
  locate libGLU.so
  locate libX11.so
  

• Check which folder they are located in (e.g. /usr/lib/x86_64-linux-gnu).

• In the samples folder, open findgllib.mk:

  gedit /home/john/cuda-7.5_samples/5_Simulations/nbody/findgllib.mk
  

• Search for the following if-condition:

  ifeq ("$(TARGET_OS)","linux")
      # $(info) >> findgllib.mk -> LINUX path <<<)
      # Each set of Linux Distros have different paths for where to find their OpenGL libraries reside
      UBUNTU_PKG_NAME = "nvidia-352"
      UBUNTU = $(shell echo $(DISTRO) | grep -i ubuntu >/dev/null 2>&1; echo $$?)
      ...
  

  and replace it with the following code:

  ifeq ("$(TARGET_OS)","linux")
      GLPATH    ?= /usr/lib/x86_64-linux-gnu
      GLLINK    ?= -L/usr/lib/x86_64-linux-gnu
      DFLT_PATH ?= /usr/lib64
  

• Generate executable:

  make
  

• An executable should have been created. Run it:

  ./nbody
  

References:

http://www.allaboutlinux.eu/remove-nouveau-and-install-nvidia-driver-in-debian-8/2/
https://andrewbolster.info/2016/04/fixcuda-on-debian-jessie
http://unix.stackexchange.com/questions/218163/how-to-install-cuda-toolkit-7-x-or-8-on-debian-8-jessie-or-9-stretch