Theano and Keras setup on ubuntu with OpenCL on AMD card

ubuntu.sh

## install Catalyst proprietary
sudo ntfsfix /dev/sda2
sudo cp /etc/X11/xorg.conf /etc/X11/xorg.conf.BAK
sudo apt-get remove --purge fglrx*
sudo apt-get install linux-headers-generic
sudo apt-get install fglrx xvba-va-driver libva-glx1 libva-egl1 vainfo 
sudo amdconfig --initial

## install build essentials
sudo apt-get install cmake
sudo apt-get update && sudo apt-get install build-essential

## install AMD APP SDK
tar -xvf AMD-APP-SDK-v3.0-0.113.50-Beta-linux64.tar.bz2 
sudo ./AMD-APP-SDK-v3.0-0.113.50-Beta-linux64.sh 

## install libgpuarray
sudo apt-get install libssl0.9.8:i386
sudo apt-get install libboost-all-dev

sudo apt-get install libgtest-dev
cd /usr/src/gtest
sudo cmake .
sudo make
sudo mv libg* /usr/lib/

## download AMD acml library
mkdir ~/Ini
tar -xvf acml-6.1.0.31-gfortran64.tgz -C ~/Ini/acml
echo "###################################################################" >> ~/.bashrc
echo "export ACML_ROOT=/home/jjarutis/Ini/acml" >> ~/.bashrc
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:"/home/jjarutis/Ini/acml/gfortran64/lib" >> ~/.bashrc

## clBlas
sudo apt-get install git
git clone https://github.com/clMathLibraries/clBLAS.git
cd clBLAS/
mkdir build
cd build/
sudo apt-cache search openblas
sudo apt-get install libopenblas-base libopenblas-dev
sudo apt-get install liblapack3gf liblapack-doc liblapack-dev
cmake ../src
make
sudo make install

git clone https://github.com/Theano/libgpuarray.git
cd libgpuarray
mkdir Build
cd Build
cmake .. -DCMAKE_BUILD_TYPE=Release -DOPENCL_INCLUDE_DIRS=/opt/AMDAPPSDK-3.0-0-Beta/include
make
sudo make install
cd ..
sudo apt-get install cython
sudo apt-get install python-numpy python-scipy python-dev python-pip python-nose g++ libopenblas-dev git
python setup.py build
sudo python setup.py install

## Theano
pip install Theano
sudo pip install Theano

THEANO_FLAGS=device=opencl0:0 python test.py


## Install emacs
sudo apt-get build-dep emacs24
wget http://ftp.gnu.org/gnu/emacs/emacs-24.5.tar.gz
tar -xf emacs-24.5.tar.* && cd emacs-24.5
./configure
make
sudo make install
cd ~/
git clone --recursive https://github.com/syl20bnr/spacemacs ~/.emacs.d
emacs
echo "setxkbmap -layout us -option ctrl:nocaps" >> ~/.bashrc

Thank you for this!

Very cool! Would be really awesome to compare OpenCL and Cuda speeds :D

Hey, sorry if it sounds rude, I don't mean to offend you. But I have been told by Theano org that they have only partial support for CUDA, for very simple stuff. Not for serious coding, so I wanted to know if I can do intensive computation with this?

Any tips to get this working on Ubuntu 16.04, been trying to get it to work for a few days now, first time trying to get calculations on the GPU and giving up hope! Apparently AMD Catalyst no longer supported on, 16.04 offers AMDGPU driver. clBLAS is giving errors on 'make'. Also am using Anaconda. Is what I'm doing a lost cause? Suggestions greatly appreciated.

GPU product: Whistler [Radeon HD 6630M/6650M/6750M/7670M/7690M]

[http://www.omgubuntu.co.uk/2016/03/ubuntu-drops-amd-catalyst-fglrx-driver-16-04]

make[2]: *** No rule to make target '/opt/AMDAPPSDK-3.0/lib/x86_64/libOpenCL.so', needed by 'library/libclBLAS.so.2.10.0'. Stop.
CMakeFiles/Makefile2:223: recipe for target 'library/CMakeFiles/clBLAS.dir/all' failed
make[1]: *** [library/CMakeFiles/clBLAS.dir/all] Error 2
Makefile:149: recipe for target 'all' failed
make: *** [all] Error 2

Thanks for this gist. I actually got Theano to work on my AMD GPU using OpenCL. Sadly Lasagne is still not using the new Theano backend (gpuarray), so do not bother if you need Lasagne.

@Formulator
The Ubuntu guys suggest not upgrading if you need to use the proprietary AMD drivers. In Ubuntu 14.04 you can switch the driver with one click within the configuration.

How does the emacs section relate to the rest of the gist?

Also beware of those 2 lines :
sudo ntfsfix /dev/sda2
sudo cp /etc/X11/xorg.conf /etc/X11/xorg.conf.BAK

Hi,

Didn't expect this to get much attention, glad somebody found it useful :)
@saxenauts you are right, at least at the time of writing this script I couldn't get Theano to train a neural network on AMD GPU, do not remember exact problem now. Some basic stuff worked though. I wonder if things have changed since then.

@Formulator I wanted to try this on 16.04 as well, I'll post the outcome.

@jonsafari emacs has nothing to with this, I should have put that part somewhere else... I basically just copy pasted bash history for future reference.

@Formulator
At least for me the following experimental driver works fine. At least tests for torch-cl pass.

Can we train DNN over opencl with theano? I read theano uses CUNN for nn training. Is that still true?

@jarutis
Thanks for this, it's great to use amd graphic card with theano
But for me, I installed everything as per your instructions but now when I try to import theano I get the error "Could not initialize pygpu, support disabled".. Please help me with this, I am a newbee in deep learning and will love it if I can get theano to work with amd

Thanks for this lovely post. Everything installed just fine however CNN does not seem to work on GPU? Have you faced similar issue. Looks like Thano does not support OpenCL completely.

https://gist.github.com/daveselinger/8cba6d41eaa70b220725091390ff52c1