Manual setup¶

Alternatively, you can install manually (refer to the dockerfile for details). Install requirements:

Python == 2.7
CUDA == 8.0
openMPI >= 2.0 (make sure that the –with-cuda flag is set during configure time)
openBLAS >= 1.12

Python¶

apt-get install python-dev python-pip

CUDA¶

If you have CUDA-enabled GPU devices, you have probably already installed nvidia drivers and CUDA. If not, you can install them from your package repository, like this:

sudo apt-get install nvidia-cuda-toolkit
reboot # nvidia driver requires reboot

This should also take care of installing current nvidia driver. If you want the latest version of CUDA or specific instructions for different platform, follow the Official CUDA installation guide.

openMPI¶

wget https://www.open-mpi.org/software/ompi/v2.0/downloads/openmpi-2.0.0.tar.gz
tar xzf openmpi-2.0.0.tar.gz
cd openmpi-2.0.0
./configure --with-cuda=/usr/local/cuda --prefix=/openmpi
make -j$(nproc) && make install

openBLAS¶

git clone https://github.com/xianyi/OpenBLAS
cd OpenBLAS && make -j$(nproc) FC=gfortran && make PREFIX=/OpenBLAS install
echo '/OpenBLAS/lib' > /etc/ld.so.conf.d/openblas.conf && ldconfig

NumPy¶

Compile numpy with openblas support

pip install cython
git clone https://github.com/numpy/numpy
printf "[openblas]\n" \
"libraries = openblas\n" \
"library_dirs = /OpenBLAS/lib\n" \
"include_dirs = /OpenBLAS/include\n" \
"runtime_library_dirs = /OpenBLAS/lib\n" >> numpy/site.cfg

cd numpy && python setup.py config && \
    python setup.py build -j $(nproc) && python setup.py install

Other system dependencies¶

Install libraries required for cuda-cffi module and zmq module.

apt-get install libffi-dev libzmq3-dev

Install crow and python dependencies¶

Set up environment variables, so it can find CUDA, OpenBLAS and openMPI libraries. To make changes permanent after reboot, you need to append these lines to /etc/profile or ~/.bashrc (depending on your configuration).

export PATH="/OpenBLAS/bin:/openmpi/bin:$PATH"
export LD_LIBRARY_PATH="/OpenBLAS/lib:/openmpi/lib:$LD_LIBRARY_PATH"
export CUDA_ROOT="/usr/local/cuda"

First download crow source from github.

git clone https://github.com/acopar/crow
cd crow

Install python requirements for CPU and GPU systems.

pip install -r requirements.txt

GPU module dependencies¶

Install python modules for GPU environment. Skip this step on computer without GPU devices.

pip install -r requirements-gpu.txt

You may need to get more recent version of scikit-cuda.

git clone https://github.com/lebedov/scikit-cuda
cd scikit-cuda && python setup.py install

Install crow module¶

Finally, you can install the package:

python setup.py install

You can also use pip to install current package.

pip install -e .

Test the configuration¶

You can check if the installed modules work correctly by importing them. Also check if your CUDA_ROOT and PATH environment variables are set correctly. If you use different distribution, path to CUDA can be different.

echo $CUDA_ROOT
ls $CUDA_ROOT

python -c 'import pycuda' # CUDA for python library
python -c 'from cuda_cffi import cusparse' # for sparse GPU operations
python -c 'import numpy; numpy.__config__.show()' # check blas info