9. OpenCL setup

In order to run Veles in OpenCL mode, there must be an OpenCL driver properly installed on your system.

9.1. Nvidia OpenCL Driver installation

Change to tty1 terminal (CTRL+ALT+F1) and stop X server (this will terminate all your running applications!):

sudo service lightdm stop

Copy the installation .run file from /data/veles/packages/binary and execute it or use the following shell script:

#!/bin/bash -e

if [[ -z "$1" ]]; then
  echo "You must specify the driver version, e.g., \"331.67\"" 1>&2
  exit 1

wget http://us.download.nvidia.com/XFree86/Linux-x86_64/$1/$package
chmod +x $package
./$package -a -q -l --no-opengl-files

This script requires a single command line argument which sets the driver version. Refer to Nvidia’s web site to find out what version is the latest. Anyway, “-l” argument makes the underlying Nvidia installation script to install the most recent version available, whereas “-a” accepts the license and “-q” answers “yes” on all questions. “–no-opengl-files” is strictly necessary because by default Nvidia’s installer writes OpenGL libs which conflict with integrated video OpenGL libs and Unity fails to start.

Start the X server:

sudo service lightdm start

As for CUDA installation, it should be straightforward, but pay attention to not install driver during cuda installation process.

9.2. Nvidia drivers troubleshooting

9.2.1. nouveau

In case of drivers problem check that there is no conflicts with nouveau driver:

sudo modprobe nvidia

If an error appears , e.g. “modprobe: ERROR: could not insert ‘nvidia’: No such device” check nouveau driver presence:

lsmod | grep nouveau

If nouveau is found, add it to the black list:

echo "blacklist nouveau" | sudo tee -a /etc/modprobe.d/blacklist-framebuffer.conf

Then reinstall Nvidia drivers package.

9.2.2. driver unloading

Looped nvidia-smi prevents from driver unloading:

nohup nvidia-smi -l

Table Of Contents

Previous topic

8. PyCharm setup

Next topic

10. Distributed setup

This Page