This guide is based on Ubuntu 12.04 LTS, but the same principles apply in newer versions of Ubuntu. Let’s begin.
1. Make sure you have a CUDA supported GPU
You must have a nVIDIA GPU that supports CUDA, otherwise you can’t program in CUDA code. Here’s a list with the CUDA supported GPU models.
2. Install nVIDIA proprietary drivers
Use Jockey (additional drivers) or just pick the driver you want from the NVIDIA official website.
3. Download CUDA Toolkit 5.0 for Ubuntu
I used the Ubuntu 11.10 32bit version. It’s the latest version so far, but it currently works fine. So please download.
4. Fix the libglut.so error
There will be an error when you’ll try to install the CUDA 5.0 examples. The driver is trying to find the libglut.so file and it doesn’t look for other versions, such as so.1, so.2 etc.
First confirm that you have a libglut file:
$ sudo find /usr -name libglut\*
If you do, symlink that file to libglut.so.
For 64bit:
$ sudo ln -s /usr/lib/x86_64-linux-gnu/libglut.so.3 /usr/lib/libglut.so
For 32bit:
$ sudo ln -s /usr/lib/i386-linux-gnu/libglut.so.3 /usr/lib/libglut.so
5. Install the CUDA Toolkit and Samples
Press CTRL+ALT+F1 to open a shell — yeah, we’re going to do this in old (yet powerful) command-line way, but there’s no need to be afraid of the black and white terminal with a blinking cursor. After all you know what they say, once you go black…
5.1 Shutdown the all the graphics
Ubuntu uses LightDM, so you need to stop this service:
$ sudo service lightdm stop
5.2 Run the installer
Go to (using cd) the directory where you have the CUDA installer (a file with *.run extension) and type the following:
$ sudo chmod +x *.run
$ sudo ./*.run
Accept the License and install only the CUDA 5 Toolkit and the Samples. DO NOT install the drivers because we have already done that.
6. Enable the nvcc compiler
In order to compile CUDA code you have to use the nvcc compile. In that so you have to tweak some environment variables into your ~/.bashrc file:
For 32bit:
export PATH=$PATH:/usr/local/cuda-5.0/bin export LD_LIBRARY_PATH=/usr/local/cuda-5.0/lib
For 64bit:
export PATH=$PATH:/usr/local/cuda-5.0/bin export LD_LIBRARY_PATH=/usr/local/cuda-5.0/lib64:/lib
If you want to compile a CUDA file (*.cu extension) you can use the following command:
nvcc -o file file.cu ./file
Or use the NSight Eclipse Edition.
