pydiso

Pydiso is a python wrapper for the pardiso solver. It is currently implemented for the Intel MKL's version of pardiso. Its goal is to expose the full functionality of pardiso to python, allowing the user to customize it as necessary for their use cases.

Installation

Installing with conda

conda install pydiso --channel conda-forge

Installing from source

The wrapper is written in cython and links MKL dynamically, through its single dynamic library (mkl_rt). It needs to find the MKL header files and libraries to compile, which the meson build backend does with pkg-config (the mkl-sdl.pc file MKL ships). Most development installations of MKL provide this. For example, conda users can install the necessary files with the mkl-devel package that is available on the default channel, conda-forge channel, the intel channel, or others, e.g.

conda install mkl-devel pkg-config

If you have installed the .pc files to a non-standard location, set PKG_CONFIG_PATH to point to that location.

After the necessary MKL files are accessible, you should be able to install by running

pip install .

in the installation directory.

Building against MKL from PyPI instead of conda

Intel also publishes MKL to PyPI: mkl (runtime libraries, a regular dependency), mkl-devel (import libs plus pkg-config files) and mkl-include (headers), the latter two listed as build requirements alongside pkgconf (a pkg-config binary from PyPI). So a plain

pip install .

in a normal (non-conda) virtual environment pulls all of them in and builds against them automatically, no PKG_CONFIG_PATH needed - meson.build locates mkl-devel's installed .pc files itself. conda-forge's pydiso package instead installs with pip install --no-deps and supplies its own conda packages, as it already does for numpy and scipy.

Linux and Windows only. Intel hasn't published MKL for macOS past 2023.2 to begin with, and that release's PyPI wheel is also missing a symlink needed at link time, so the build succeeds but the extension fails to import. conda-forge's macOS packaging doesn't have that gap - use conda there instead.

Prebuilt wheels

The prebuilt wheels on PyPI (see .github/workflows/wheels.yml) link the same way and depend on the mkl package, which supplies the MKL libraries (and its threading runtime) at import time. MKL is not bundled into them.

Licensing

pydiso's own source code is MIT licensed (see LICENSE), and that is all any pydiso distribution contains. MKL is a separate dependency, installed from its own package under Intel's license (the Intel Simplified Software License on PyPI); pydiso does not redistribute it.