build status

Installing PyOP2

The main testing platform for PyOP2 is Ubuntu 12.04 64-bit with Python 2.7.3. Other UNIX-like systems may or may not work. Mac OS X 10.7 and 10.9 are also known to work. Microsoft Windows may work, but is not a supported platform.

Quick start installations

Installation script for Ubuntu

For the impatient there is a script for the unattended installation of PyOP2 and its dependencies on a Ubuntu 12.04 or compatible platform. Only the sequential and OpenMP backends are covered at the moment.


This script will only work reliably on a clean Ubuntu installation and is not intended to be used by PyOP2 developers. If you intend to contribute to PyOP2 it is recommended to follow the instructions below for a manual installation.

Running with superuser privileges will install missing packages and Python dependencies will be installed system wide:

wget -O - | sudo bash


This will fail if you if you require a password for sudo. Run e.g. the following beforehand to assure your password is cached

sudo whoami

Running without superuser privileges will instruct you which packages need to be installed. Python dependencies will be installed to the user site ~/.local:

wget -O - | bash

In each case, PyOP2 will be cloned to subdirectories of the current directory.

After installation has completed and a rudimentary functionality check, the test suite is run. The script indicates whether all these steps have completed successfully and only in this case will exit with return code 0.

Only high-level progress updates are printed to screen. Most of the output is redirected to a log file pyop2_install.log. Please consult this log file in the case of errors. If you can’t figure out the cause of discover a bug in the installation script, please report it.

This completes the quick start installation. More complete instructions follow for virtual machine and native installations.

Provisioning a virtual machine

A Vagrantfile is provided for automatic provisioning of a Ubuntu 12.04 64bit virtual machine with PyOP2 preinstalled. It requires VirtualBox 4.2 and Vagrant to be installed, which are available for Linux, Mac and Windows.

Creating and launching a virtual machine is a single command: run vagrant up to automatically download the base VM image, configure it for use with VirtualBox, boot the VM and install PyOP2 and all dependencies using the above install script.

Manual Installation



You can skip over the dependencies list for now, since the instructions below tell you how to install each of these packages.

PyOP2 requires a number of tools and libraries to be available:

  • A C compiler (for example gcc or clang), make
  • A Fortran compiler (for PETSc)
  • MPI
  • Blas and Lapack
  • Git, Mercurial
  • pip and the Python headers

The following dependencies are part of the Python subsystem:

  • Cython >= 0.17
  • decorator
  • numpy >= 1.6
  • networkx

PETSc. We require very recent versions of PETSc so you will need to follow the specific instructions given below to install the right version.

Testing dependencies (optional, required to run the tests):

  • pytest >= 2.3
  • flake8 >= 2.1.0
  • gmsh
  • triangle

With the exception of the PETSc dependencies, these can be installed using the package management system of your OS, or via pip.

Installing packages with pip

To install dependencies system-wide use sudo pip install ..., to install to a user site use pip install --user .... If you don’t want PyOP2 or its dependencies interfering with your existing Python environment, consider creating a virtualenv.


In the following we will use sudo pip install .... If you want either of the other options you should change the command appropriately.


Installing to the user site does not always give packages priority over system installed packages on your sys.path.

Obtaining a build environment on Ubuntu and similar systems

On a Debian-based system (Ubuntu, Mint, etc.) install core packages:

sudo apt-get install -y build-essential python-dev git-core \
  mercurial python-pip libopenmpi-dev openmpi-bin libblas-dev \
  liblapack-dev gfortran


This may not give you recent enough versions of those packages (in particular the Cython version shipped with 12.04 is too old). You can selectively upgrade packages via pip, see below.

Install dependencies via pip:

sudo pip install "Cython>=0.17" decorator "numpy>=1.6" networkx


If your OS release is very old and you are therefore using Python 2.6 instead of 2.7, you need two additional dependencies.

Additional Python 2.6 dependencies:

  • argparse
  • ordereddict

Install these via pip:

sudo pip install argparse ordereddict


You can now skip down to installing PETSc.

Obtaining a build environment on Mac OS

We recommend using Homebrew as a package manager for the required packages on Mac OS systems. Obtaining a build environment for PyOP2 consists of the following:

  1. Install Xcode. For OS X 10.9 (Mavericks) this is possible through the App Store. For earlier versions, try (note that on OS X 10.7 (Lion) you will need to obtain Xcode 4.6 rather than Xcode 5)

  2. If you did not install Xcode 5, you will need to additionally install the Xcode command line tools through the downloads section of Xcode’s preferences

  3. Install homebrew, following the instructions at

  4. Install an MPI library (PyOP2 is tested with openmpi):

    brew install openmpi
  5. Install an up-to-date Python via homebrew:

    brew install python


    Do not follow the instructions to update pip, since they currently result in a broken pip installation (see

  6. Install numpy via homebrew:

    brew tap homebrew/python
    brew install numpy
  7. Install python dependencies via pip:

    pip install decorator
    pip install cython
    pip install mpi4py
    pip install pytest
    pip install flake8


Your system is now ready to move on to installation of PETSc and petsc4py described below.


On Mac OS we do not recommend using sudo when installing, as such when following instructions below to install with pip just remove the sudo portion of the command.


PyOP2 uses petsc4py, the Python bindings for the PETSc linear algebra library and requires:

  • an MPI implementation built with shared libraries
  • A suitable very recent PETSc master branch built with shared libraries

If you have a suitable PETSc installed on your system, PETSC_DIR and PETSC_ARCH need to be set for the petsc4py installer to find it.


There are no current OS PETSc packages which are new enough. Therefore, unless you really know you should be doing otherwise, always install PETSc using pip. The following instructions will install the firedrake branch of PETSc and petsc4py. This is a recent version of the upstream master branch which has been verified to at least build correctly. You may also use the upstream next or master branch, but be aware that these are rapidly developing and tend to break regularly.

Then install PETSc via pip

sudo PETSC_CONFIGURE_OPTIONS="--download-ctetgen --download-triangle --download-chaco" \
  pip install


If you intend to run PyOP2’s OpenMP backend, you should additionally pass the following options to the PETSc configure stage

--with-threadcomm --with-openmp --with-pthreadclasses

If you built PETSc using pip, PETSC_DIR and PETSC_ARCH should be left unset when building petsc4py.

Install petsc4py via pip:

sudo pip install git+

If you have previously installed and older version of PETSc or petsc4py, pip might tell you that the requirements are already satisfied when running above commands. In that case, use pip install -U --no-deps to upgrade (--no-deps prevents also recursively upgrading any dependencies).


If you only intend to run PyOP2 on CPUs (not GPUs) you can now skip straight to Building PyOP2, otherwise read on for additional dependencies.

CUDA backend:


  • boost-python
  • Cusp 0.3.1
  • codepy >= 2013.1
  • Jinja2
  • mako
  • pycparser >= 2.10
  • pycuda >= 2013.1

The cusp library version 0.3.1 headers need to be in your (CUDA) include path.

Note: Using the trunk version of Cusp will not work, since revision f525d61 introduces a change that break backwards compatibility with CUDA 4.x.

Install dependencies via the package manager (Debian based systems):

sudo apt-get install libboost-python-dev python-jinja2 python-mako python-pycuda

Note: The version of pycparser available in the package repositories is too old, you will need to install it via pip, see below.

Install dependencies via pip:

sudo pip install codepy Jinja2 mako pycparser>=2.10

If a pycuda package is not available, it will be necessary to install it manually. Make sure nvcc is in your $PATH and in your $LIBRARY_PATH if in a non-standard location:

export CUDA_ROOT=/usr/local/cuda # change as appropriate
git clone
cd pycuda
git submodule init
git submodule update
# is in a non-standard location on Ubuntu systems
./ --no-use-shipped-boost \
python build
sudo python install
sudo cp /etc/

OpenCL backend:


  • Jinja2
  • mako
  • pycparser >= 2.10
  • pyopencl >= 2012.1

pyopencl requires the OpenCL header CL/cl.h in a standard include path. On a Debian system, install it via the package manager:

sudo apt-get install opencl-headers

If you want to use OpenCL headers and/or libraries from a non-standard location you need to configure pyopencl manually:

export OPENCL_ROOT=/usr/local/opencl # change as appropriate
git clone
cd pyopencl
git submodule init
git submodule update
./ --no-use-shipped-boost \
--cl-inc-dir=${OPENCL_ROOT}/include --cl-lib-dir=${OPENCL_ROOT}/lib
python build
sudo python install

Otherwise, install dependencies via pip:

sudo pip install Jinja2 mako pyopencl>=2012.1 pycparser>=2.10

Installing the Intel OpenCL toolkit (64bit systems only):

cd /tmp
# install alien to convert the rpm to a deb package
sudo apt-get install alien
fakeroot wget
tar xzf intel_sdk_for_ocl_applications_2012_x64.tgz
fakeroot alien *.rpm
sudo dpkg -i --force-overwrite *.deb

The --force-overwrite option is necessary in order to resolve conflicts with the opencl-headers package (if installed).

Installing the AMD OpenCL toolkit (32bit and 64bit systems):

# on a 32bit system, instead
tar xzf AMD-APP-SDK-v2.8-lnx*.tgz
# Install to /usr/local instead of /opt
sed -ie 's:/opt:/usr/local:g' default-install_lnx*.pl
sudo ./


PyOP2 allows initializing data structures using data stored in HDF5 files. To use this feature you need the optional dependency h5py.

On a Debian-based system, run:

sudo apt-get install libhdf5-mpi-dev python-h5py

Alternatively, if the HDF5 library is available, sudo pip install h5py.

Building PyOP2

Clone the PyOP2 repository:

git clone git://

PyOP2 uses Cython extension modules, which need to be built in-place when using PyOP2 from the source tree:

python build_ext --inplace

When running PyOP2 from the source tree, make sure it is on your $PYTHONPATH:


When installing PyOP2 via python install the extension modules will be built automatically and amending $PYTHONPATH is not necessary.

Setting up the environment

To make sure PyOP2 finds all its dependencies, create a file .env e.g. in your PyOP2 root directory and source it via . .env when using PyOP2. Use the template below, adjusting paths and removing definitions as necessary:

#PETSc installation, not necessary when PETSc was installed via pip
export PETSC_DIR=/path/to/petsc
export PETSC_ARCH=linux-gnu-c-opt


Alternatively, package the configuration in an environment module.

Testing your installation

PyOP2 unit tests use pytest >= 2.3. Install via package manager:

sudo apt-get install python-pytest

or pip:

sudo pip install "pytest>=2.3"

The code linting test uses flake8. Install via pip:

sudo pip install "flake8>=2.1.0"

If you install pytest and flake8 using pip --user, you should include the binary folder of your local site in your path by adding the following to ~/.bashrc or .env:

# Add pytest binaries to the path
export PATH=${PATH}:${HOME}/.local/bin

If all tests in our test suite pass, you should be good to go:

make test

This will run code linting and unit tests, attempting to run for all backends and skipping those for not available backends.


Start by verifying that PyOP2 picks up the “correct” dependencies, in particular if you have several versions of a Python package installed in different places on the system.

Run pydoc <module> to find out where a module/package is loaded from. To print the module search path, run:

python -c 'from pprint import pprint; import sys; pprint(sys.path)'

Troubleshooting test failures

Run the tests as follows, to abort after the first failed test:

Start with the unit tests with the sequential backend

py.test test/unit -vsx --tb=short --backend=sequential

With all the sequential tests passing, move on to the next backend in the same manner as required.