Skip to content

Latest commit

 

History

History
173 lines (155 loc) · 6.72 KB

compiling-KLU.md

File metadata and controls

173 lines (155 loc) · 6.72 KB
Raw

PyBaMM developer install - The KLU sparse solver

If you wish to try a different DAE solver, PyBaMM currently offers a direct interface to the sparse KLU solver within Sundials. This solver comes as a C++ python extension module. Therefore, when installing PyBaMM from source (e.g. from the GitHub repository), the KLU sparse solver module must be compiled. Running pip install . or python setup.py install in the PyBaMM directory will result in a attempt to compile the KLU module.

Note that if CMake of pybind11 are not found (see below), the installation of PyBaMM will carry on, however skipping the compilation of the idaklu module. This allows developers that are not interested in the KLU module to install PyBaMM from source without having to install the required dependencies.

To build the KLU solver, the following dependencies are required:

The first four should be available through your favourite package manager. On Debian-based GNU/Linux distributions:

apt update
apt install python3-dev gcc gfortran cmake libopenblas-dev

pybind11

The pybind11 source directory should be located in the PyBaMM project directory at the time of compilation. Simply clone the GitHub repository, for example:

# In the PyBaMM project dir (next to setup.py)
git clone https://github.com/pybind/pybind11.git

SuiteSparse and sundials

Method 1 - Using the convenience script

The PyBaMM repository contains a script scripts/setup_KLU_module_build.py that automatically downloads, extracts, compiles and installs the two libraries.

First install the Python wget module

pip install wget

Then execute the script

# In the PyBaMM project dir (next to setup.py)
python scripts/setup_KLU_module_build.py

The above will install the required component of SuiteSparse and Sundials in your home directory under ~/.local/. Note that you can provide the option --install-dir=<install/path> to install both libraries to an alternative location. If <install/path> is not absolute, it will be interpreted as relative to the PyBaMM project directory.

Finally, reactivate your virtual environment by running

source $(VIRTUAL_ENV)/bin/activate

Alternatively, you update the LD_LIBRARY_PATH environment variable as follows

export LD_LIBRARY_PATH=$(HOME)/.local:$LD_LIBRARY_PATH

The above export statement will be ran automatically the next time you activate you python virtual environment.

If did not run the convenience script inside a python virtual environment, execute you bash config file

source ~/.bashrc

(or start a new shell).

Build files are located inside the PyBaMM project directory under KLU_module_deps/. Feel free to remove this directory once everything is installed correctly.

Method 2 - Compiling Sundials (advanced)

SuiteSparse

On most current linux distributions and macOS, a recent enough version of the suitesparse source package is available through the package manager. For instance on Fedora

yum install libsuitesparse-dev

Sundials

The PyBaMM KLU solver requires Sundials >= 4.0. Because most Linux distribution provide older versions through their respective package manager, it is recommended to build and install Sundials manually.

First, download and extract the sundials 5.0.0 source

wget https://computing.llnl.gov/projects/sundials/download/sundials-5.0.0.tar.gz .
tar -xvf sundials-5.0.0.tar.gz

Then, create a temporary build directory and navigate into it

mkdir build_sundials
cd build_sundials

You can now configure the build, by running

cmake -DLAPACK_ENABLE=ON\
      -DSUNDIALS_INDEX_SIZE=32\
      -DBUILD_ARKODE=OFF\
      -DBUILD_CVODE=OFF\
      -DBUILD_CVODES=OFF\
      -DBUILD_IDAS=OFF\
      -DBUILD_KINSOL=OFF\
      -DEXAMPLES_ENABLE:BOOL=OFF\
      -DKLU_ENABLE=ON\
      -DKLU_INCLUDE_DIR=path/to/suitesparse/headers\
      -DKLU_LIBRARY_DIR=path/to/suitesparse/libraries\
      ../sundials-5.0.0

Be careful set the two variables KLU_INCLUDE_DIR and KLU_LIBRARY_DIR to the correct installation location of the SuiteSparse libary on your system. If you installed SuiteSparse through your package manager, this is likely to be something similar to:

-DKLU_INCLUDE_DIR=/usr/include/suitesparse\
-DKLU_LIBRARY_DIR=/usr/lib/x86_64-linux-gnu\

By default, Sundials will be installed on your system under /usr/local (this varies depending on the distribution). Should you wish to install sundials in a specific location, set the following variable

-DCMAKE_INSTALL_PREFIX=install/location\

Finally, build the library:

make install

You may be asked to run this command as a super-user, depending on the installation location.

Alternative installation location

By default, it is assumed that the SuiteSparse and Sundials libraries are installed in your home directory under ~/.local. If you installed the libraries to (a) different location(s), you must set the options suitesparse-root or/and sundials-root when installing PyBaMM. Examples:

python setup.py install --suitesparse-root=path/to/suitesparse

or

pip install . --install-option="--sundials-root=path/to/sundials"

(re)Install PyBaMM to build the KLU solver

If the above dependencies are correctly installed, any of the following commands will install PyBaMM with the idaklu solver module:

pip install .
pip install -e .
python setup.py install
python setup.py develop
...

Note that it doesn't matter if pybamm is already installed. The above commands will update your exisitng installation by adding the idaklu module.

Check that the solver is correctly installed

If you install PyBaMM in editable mode using the -e pip switch or if you use the python setup.py install command, a log file will be located in the project directory (next to the setup.py file).

cat setup.log
020-03-24 11:33:50,645 - PyBaMM setup - INFO - Starting PyBaMM setup
2020-03-24 11:33:50,653 - PyBaMM setup - INFO - Not running on windows
2020-03-24 11:33:50,654 - PyBaMM setup - INFO - Could not find CMake. Skipping compilation of KLU module.
2020-03-24 11:33:50,655 - PyBaMM setup - INFO - Could not find pybind11 directory (/io/pybind11). Skipping compilation of KLU module.

If the KLU sparse solver is correctly installed, then the following command should return True.

$ python -c "import pybamm; print(pybamm.have_idaklu())"