Getting CUDA with cuDNN

In this guide I will show you how to install CUDA with cuDNN in your home directory without invoking admin privileges. Unless stated otherwise, we will be using this configuration for all of my guides.

Assumptions

• Ubuntu OS 16.04
• NVIDIA GPU with driver installed

What is

From NVIDIA:

CUDA® is a parallel computing platform and programming model developed by NVIDIA for general computing on graphical processing units (GPUs). With CUDA, developers are able to dramatically speed up computing applications by harnessing the power of GPUs.

The NVIDIA CUDA® Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks. cuDNN provides highly tuned implementations for standard routines such as forward and backward convolution, pooling, normalization, and activation layers.

In essence, if you are working on deep learning tools and require GPU-enabled processes, you will need CUDA along with cuDNN.

Why

The reason for a non-standard installation is two-fold:

1. You may not have admin privileges to install CUDA and cuDNN to system paths in the conventional manner
2. Even if your sysadmin have already installed CUDA or cuDNN on your machine, it may still be a good idea to install them on your own because repeating research experiments by others sometimes require specific versions of CUDA and cuDNN. Non-standard installations can thus grant you the ability to toggle between various versions under different environments easily using a virtual environment manager such as Conda

Getting CUDA

The standard installation path for cuda is the versioned path /usr/local/cuda-x.x where x.x reflects the version number. The standard installation will also create a symlink /usr/local/cuda to that versioned install path. You may check if a sysadmin had already installed cuda in your machine via the command:

cat /usr/local/cuda/version.txt
#=> CUDA Version 8.0.61

Let’s heave over to NVIDIA’s CUDA download page then select your system configurations and download the runfile (local) for the CUDA version that you want. For this guide, I shall choose version 9.0 as the example.

Let’s make the versioned cuda install directory under $HOME following NVIDIA’s convention of displaying the version in the dirname:

cd $HOME
mkdir cuda-9.0

Let’s save this under the CUDA_HOME environment variable:

export CUDA_HOME=$HOME/cuda-9.0

I downloaded the runfile named cuda_9.0.176_384.81_linux-run into my ~/Downloads directory. Let’s head over and install CUDA into the versioned install directory:

cd ~/Downloads
sh cuda_9.0.176_384.81_linux-run --silent --toolkit --toolkitpath=$CUDA_HOME

Once the installation is done, we have to ensure the following environment variables are assigned correctly:

export CUDA_HOME=$CUDA_HOME # included for sake of completeness
export PATH=$CUDA_HOME/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=$CUDA_HOME/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
# For 32-bit OS: export LD_LIBRARY_PATH=$CUDA_HOME/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

You may append them to ~/.bashrc if you’re running on your local machine, ~/.bash_profile if you’re accessing a remote machine via ssh, or better yet save the environment variables under a Conda environment!

We are done for CUDA!

Getting cuDNN

Installing cuDNN is relatively hassle-free as it’s simply about unpacking cuDNN files into our $CUDA_HOME directory.

Download the NVIDIA cuDNN here corresponding to your CUDA version. You’ll be requried to sign up for a free developer account if you have not already done so. For this guide I shall choose cuDNN version 7.0.4 as the example.

On the download page click on the link that reads “Download cuDNN v7.0.4 (Nov 13, 2017), for CUDA 9.0”. It should activate an accordion style dropdown displaying a couple of download links. We shall choose the Library for Linux because the other modes of installation requires admin privileges.

Lets download cudnn-9.0-linux-x64-v7.tgz into our ~/Downloads directory by clicking on the download link that reads “cuDNN v7.0.4 Library for Linux”

Once the download is complete, we shall head over to the downloads directory and simply unpack the cuDNN files into our CUDA directory:

cd ~/Downloads
tar -zxvf cudnn-9.0-linux-x64-v7.tgz -C $CUDA_HOME --strip-components=1

That’s it! We are done!