/PyOP2

OP2 runtime library and python bindings

Primary LanguagePythonOtherNOASSERTION

Installation

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. Microsoft Windows is not supported.

Quick start

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.

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

wget -O - https://github.com/OP2/PyOP2/raw/master/install.sh | sudo bash

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

wget -O - https://github.com/OP2/PyOP2/raw/master/install.sh | bash

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.

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.

Preparing the system

OP2 and PyOP2 require a number of tools to be available:

  • gcc, make, CMake
  • bzr, Git, Mercurial
  • pip and the Python headers
  • SWIG

On a Debian-based system (Ubuntu, Mint, etc.) install them by running

sudo apt-get install -y build-essential python-dev bzr git-core mercurial \
       cmake cmake-curses-gui python-pip swig

OP2-Common

PyOP2 depends on the OP2-Common library (only sequential is needed), which is built in-place as follows:

git clone git://github.com/OP2/OP2-Common.git
cd OP2-Common/op2/c
./cmake.local -DOP2_WITH_CUDA=0 -DOP2_WITH_HDF5=0 -DOP2_WITH_MPI=0 -DOP2_WITH_OPENMP=0
cd ..
export OP2_DIR=`pwd`

For further instructions refer to the [OP2-Common README] (https://github.com/OP2/OP2-Common/blob/master/op2/c/README).

If you have already built OP2-Common, make sure OP2_DIR is exported or the PyOP2 setup will fail.

Dependencies

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

Note: In the following we will use pip install ... to mean any of the above options.

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

Common

Common dependencies:

  • Cython >= 0.17
  • decorator
  • instant >= 1.0
  • numpy >= 1.6
  • PETSc >= 3.3 with Fortran interface, C++ and OpenMP support
  • PETSc4py >= 3.3
  • PyYAML

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

sudo apt-get install cython python-decorator python-instant python-numpy python-yaml

Note: 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:

pip install Cython=>0.17 decorator instant numpy pyyaml

Additional Python 2.6 dependencies:

  • argparse
  • ordereddict

Install these via pip:

pip install argparse ordereddict

PETSc

PyOP2 uses petsc4py, the Python bindings for the PETSc linear algebra library.

We maintain a fork of petsc4py with extensions that are required by PyOP2 and requires:

  • an MPI implementation built with shared libraries
  • PETSc 3.3 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. On a Debian/Ubuntu system with PETSc 3.3 installed, this can be achieved via:

export PETSC_DIR=/usr/lib/petscdir/3.3
export PETSC_ARCH=linux-gnu-c-opt

If not, make sure all PETSc dependencies (BLAS/LAPACK, MPI and a Fortran compiler) are installed. On a Debian based system, run:

sudo apt-get install -y libopenmpi-dev openmpi-bin libblas-dev liblapack-dev gfortran

If you want OpenMP support or don't have a suitable PETSc installed on your system, build the PETSc OMP branch:

PETSC_CONFIGURE_OPTIONS="--with-fortran-interfaces=1 --with-c++-support --with-openmp" \
  pip install hg+https://bitbucket.org/ggorman/petsc-3.3-omp
unset PETSC_DIR
unset PETSC_ARCH

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

Install petsc4py via pip:

pip install hg+https://bitbucket.org/mapdes/petsc4py#egg=petsc4py

Note: When using PyOP2 with Fluidity it's crucial that both are built against the same PETSc, which must be build with Fortran support!

CUDA backend:

Dependencies:

  • boost-python
  • Cusp 0.3.1
  • codepy >= 2013.1
  • Jinja2
  • mako
  • pycparser >= 2.09.1 (revision a460398 or newer)
  • pycuda revision a6c9b40 or newer

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:

pip install codepy Jinja2 mako git+https://github.com/eliben/pycparser.git#egg=pycparser-2.09.1

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

export CUDA_ROOT=/usr/local/cuda # change as appropriate
cd /tmp
git clone http://git.tiker.net/trees/pycuda.git
cd pycuda
git submodule init
git submodule update
# libcuda.so is in a non-standard location on Ubuntu systems
./configure.py --no-use-shipped-boost \
  --cudadrv-lib-dir='/usr/lib/nvidia-current,${CUDA_ROOT}/lib,${CUDA_ROOT}/lib64'
python setup.py build
sudo python setup.py install
sudo cp siteconf.py /etc/aksetup-defaults.py

OpenCL backend:

Dependencies:

  • Jinja2
  • mako
  • pycparser >= 2.09.1 (revision a460398 or newer)
  • pyopencl >= 2012.1

Install via pip:

pip install Jinja2 mako pyopencl>=2012.1 git+https://github.com/eliben/pycparser.git#egg=pycparser-2.09.1

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 http://registrationcenter.intel.com/irc_nas/2563/intel_sdk_for_ocl_applications_2012_x64.tgz
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):

wget http://developer.amd.com/wordpress/media/2012/11/AMD-APP-SDK-v2.8-lnx64.tgz
# on a 32bit system, instead
# wget http://developer.amd.com/wordpress/media/2012/11/AMD-APP-SDK-v2.8-lnx32.tgz
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 ./Install-AMD-APP.sh

HDF5

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, pip install h5py.

Building PyOP2

Clone the PyOP2 repository:

git clone git://github.com/OP2/PyOP2.git

If not set, OP2_DIR should be set to the location of the 'op2' folder within the OP2-Common build. PyOP2 uses Cython extension modules, which need to be built in-place when using PyOP2 from the source tree:

python setup.py build_ext --inplace

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

export PYTHONPATH=/path/to/PyOP2:$PYTHONPATH

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

FFC Interface

Solving UFL finite element equations requires a fork of FFC and dependencies:

Install via the package manager

On a supported platform, get all the dependencies for FFC by installing the FEniCS toolchain from packages:

sudo apt-get install fenics

Our FFC fork is required, and must be added to your $PYTHONPATH:

git clone -b pyop2 https://bitbucket.org/mapdes/ffc.git $FFC_DIR
export PYTHONPATH=$FFC_DIR:$PYTHONPATH

This branch of FFC also requires the latest version of UFL, also added to $PYTHONPATH:

git clone https://bitbucket.org/fenics-project/ufl.git $UFL_DIR
export PYTHONPATH=$UFL_DIR:$PYTHONPATH

Install via pip

Alternatively, install FFC and all dependencies via pip:

pip install \
  git+https://bitbucket.org/mapdes/ffc.git@pyop2#egg=ffc
  bzr+http://bazaar.launchpad.net/~florian-rathgeber/ufc/python-setup#egg=ufc_utils
  git+https://bitbucket.org/fenics-project/ufl.git#egg=ufl
  git+https://bitbucket.org/fenics-project/fiat.git#egg=fiat
  hg+https://bitbucket.org/khinsen/scientificpython

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:

# Root directory of your OP2 installation, always needed
export OP2_DIR=/path/to/OP2-Common/op2
# If you have installed the OP2 library define e.g.
export OP2_PREFIX=/usr/local

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

# Add UFL and FFC to PYTHONPATH if in non-standard location
export UFL_DIR=/path/to/ufl
export FFC_DIR=/path/to/ffc
export PYTHONPATH=$UFL_DIR:$FFC_DIR:$PYTHONPATH
# Add any other Python module in non-standard locations

# Add PyOP2 to PYTHONPATH
export PYTHONPATH=/path/to/PyOP2:$PYTHONPATH

Alternatively, package the configuration in an environment module.

Testing your installation

PyOP2 unit tests use pytest. Install via package manager

sudo apt-get install python-pytest

or pip

pip install pytest

If you install pytest using pip --user, you should include the pip binary folder in you path by adding the following to .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 both unit and regression tests, the latter require UFL and FFC.

This will attempt to run tests for all backends and skip those for not available backends. If the FFC fork is not found, tests for the FFC interface are xfailed.

Troubleshooting

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)'