GPUCSL: GPU-Based Library for Causal Structure Learning

Tom Braun, Ben Hurdelhey, Dominik Meier, Petr Tsayun, Christopher Hagedorn, Johannes Huegle, Rainer Schlosser

Abstract: GPUCSL is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$a$</tex> maintainable and extensible Python library for GPU-accelerated causal structure learning (CSL) based on the PC algorithm. The library supports multivariate normal and discrete distributed data, and implements multi-GPU support for multivariate normal distributed data. GPUCSL combines several stand-alone independent research implementations to allow a unified entry point into G PU -accelerated CSL. The library outperforms CPU -based implementations with an average speedup factor of 9.5 against pcalg and 19.8 against bnlearn and remains within the order of magnitude of existing G PU -accelerated CSL research. The source code is available at https://github.com/hpi-epic/gpucsl.