Bempp

Boundary element method Python package

Installing Bempp

The page contains information about how to install the latest version of Bempp (Bempp-cl 0.3.2). You can find information about installing the previous version (Bempp 3.3.4) here.

Running Bempp with Docker

If you want to use Bempp without installing it directly on your computer, you can run a Jupyter notebook lab containing Bempp using Docker. To do this, run the following commands:

docker pull bempp/cl-notebook:latest
docker run -p 8888:8888 bempp/cl-notebook

Instructions for opening your Jupyter lab in a web browser will then be output to your terminal.

In order to share a directory with the docker image, replace the second command with:

docker run -v /path/to/folder:/root/shared -p 8888:8888 bempp/cl-notebook

The folder /path/to/folder on your computer will then be visible as shared in the root directory of the Jupyter lab.

Running Numba-only Docker image

On some systems, running OpenCL inside a Docker image can cause issues with drivers. Typically, this will cause errors to be thrown when trying to import Bempp. If you experience this, you can run the Numba-only Bempp Docker image using the following commands:

docker pull bempp/cl-notebook-numba:latest
docker run -p 8888:8888 bempp/cl-notebook-numba

A folder can be shared with the image in the same way as above.

Quick install

The latest release of Bempp can be installed using pip:

pip install bempp-cl

Alternatively, the latest development version of Bempp can be installed using pip:

pip install numpy scipy numba meshio>=4.0.16 gmsh==4.6.0
pip install git+https://github.com/bempp/bempp-cl.git

Alternatively, you can install Bempp using conda. First, if not already done, enable the conda-forge channel:

conda config --add channels conda-forge
conda config --set channel_priority strict

You can then install Bempp with:

conda install bempp-cl

We noticed some recent problems with Bempp and the pocl driver available in conda, which can lead to segfaults. We are in the process of updating the conda environment to solve these problems. We strongly recommend use of the Docker image instead. Alternatively, you can attempt the following fix within the conda environment.

conda uninstall pocl
conda installl intel-opencl-rt

Installing from Source

You can find the source code of bempp-cl on GitHub.

Requirements

Bempp can be installed on any Windows, Mac or Linux system with Python 3.7+. The following dependencies are required:

Please note that this installs only a slow base environment. To profit from fast compute kernels via OpenCL we strongly recommend to use the Docker image. Alternatively, read the section on OpenCL further below.

These can be installed with:

pip install numpy scipy numba meshio>=4.0.16

Some features will not work without the following dependencies:

These can be installed with:

pip install plotly pyopencl gmsh==4.6.0

git clone https://github.com/exafmm/exafmm-t.git
cd exafmm-t
./configure 
make
make install
python setup.py install

Installing Bempp

Once you have installed all the requirements, you can install Bempp with the following commands:

git clone https://github.com/bempp/bempp-cl
cd bempp-cl
python setup.py install

OpenCL

Bempp-cl uses OpenCL kernels to launch computations. OpenCL is a standard for accessing various types of compute devices, including CPUs and GPUs. Bempp-cl uses OpenCL to compile C99 kernels during runtime for the underlying compute device. This requires a runtime environment to be installed. Runtime environments are available from Apple (CPU, GPU), Nvidia (GPU), AMD (GPU), Intel (CPU/GPU) and through the open-source Pocl Project (CPU). In the following a few important remarks about OpenCL runtime environments.

Using other compute devices

Any compute device that provides a valid ICD can in principle be used with Bempp. However, if Bempp is installed within a conda environment these may not be seen by the installation. The reason is that conda expects files in different locations than may be provided by the system. The following two solutions have been tested on Linux based systems.

Set the ICD location via environment variable

Typically, GPU vendors install ICDs in /etc/OpenCL/vendors. You can set this as default search path for ICD files with the command ::

export OPENCL_VENDOR_PATH=/etc/OpenCL/vendors

The drawback is that as long as this variable is set the Pocl driver provided through conda is not visible any more.

Copy ICD files into the correct location

An alternative to setting an environment variable is to copy the system ICD files into a location that conda knows.

First, find out where your conda environment is installed, using e.g. the command which python when the environment is active. The output of the command may be something like

/home/user/miniconda3/envs/bempp/bin/python

But this depends on whether Anaconda Python or Miniconda is installed and where the base installation resides. The directory in which conda installs ICDs is

/home/user/miniconda3/envs/bempp/etc/OpenCL/vendors

This directory already contains a file called pocl.icd which has been installed through the pocl conda package. To use other ICD drivers copy them over from the system directory /etc/OpenCL/vendors into the above directory of your conda environment.