Tensorflow Preinstall: Drivers, cuDNN, CUDA

August 23, 2017
Tensorflow Environment

Preinstall

Some packages for builds & dependencies

sudo apt-get update
sudo apt-get upgrade
sudo apt-get install -y build-essential git unzip wget
    libcurl3-dev libcupti-dev python-dev python3-dev python-pip python3-pip \
    python-pydot python-numpy python3-numpy python-six python3-six \
    python-virtualenv python-wheel python3-wheel python-matplotlib \
    python-pandas python-sklearn  openjdk-8-jdk libfreetype6-dev libxft-dev \
    libncurses-dev libopenblas-dev libblas-dev liblapack-dev \
    libatlas-base-dev gfortran swig \
    linux-headers-generic linux-image-extra-virtual \
    pkg-config zip g++ zlib1g-dev libcurl3-dev
sudo pip install -U pip

Install Nvidia Drivers

  1. PPA
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt update
  1. apt-get
sudo apt-get install nvidia-375
  1. Check the result using nvidia-smi

Install Nvidia Toolkit 8.0

There’re many different ways like local/network install

# CHECK YOUR LATEST VERSION
wget https://developer.nvidia.com/compute/cuda/8.0/Prod2/local_installers/cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64-deb

sudo dpkg -i cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb
rm cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb
sudo apt-get update
sudo apt-get install -y cuda

This will install cuda into: /usr/local/cuda

cuda, cuda-8.0

redfish@lsab:~$ ll /usr/local | grep cuda
lrwxrwxrwx  1 root root    8 Aug 22 23:20 cuda -> cuda-8.0/
drwxr-xr-x 14 root root 4096 Aug 22 23:20 cuda-8.0/
  • As you can see, the CUDA Toolkit default folder is set to /usr/local/cuda-8.0
  • The /usr/local/cuda symbolic link points to the location where the CUDA Toolkit was installed.
  • This link allows projects to use the latest CUDA Toolkit without any configuration file update.

Environment Variables

  • export CUDA_HOME=/usr/local/cuda-8.0/
  • export PATH=$PATH:$CUDA_HOME/bin$
  • export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_HOME/lib

Install cuDNN

  • Not sure if r1.2 TensorFlow support cuDNN 6.0 or later, so cuDNN v5.1 is totally safe.
  • Download cuDNN Library for Linux: cudnn-8.0-linux-x64-v5.1.tgz
  • extract into /usr/local/cuda via:
sudo tar -xzvf cudnn-8.0-linux-x64-v5.1.tgz
cuda/include/cudnn.h
cuda/lib64/libcudnn.so
cuda/lib64/libcudnn.so.5
cuda/lib64/libcudnn.so.5.1.10
cuda/lib64/libcudnn_static.a

sudo cp cuda/include/cudnn.h /usr/local/cuda/include
sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

NVIDIA CUDA Profile Tools Interface

  • sudo apt-get install libcupti-dev
  • export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda/extras/CUPTI/lib64"
References
comments powered by Disqus