Towards Polyhedral Automatic Differentiation

Jan Hückelheim, Navjot Kukreja

TL;DR: The polyhedral model could enable AD tools to compute more efficient derivatives of arbitrary models.

Abstract: Most Automatic Differentiation (AD) tools lack a way to explicitly represent or differentiate performance-critical and hardware-dependent constructs such as parallelism, vectorisation, or tiling. Machine-learning frameworks work around this by hiding implementation details from the AD process, but lack the generality of general-purpose programming languages. Instead, this talk discusses the polyhedral model as a way for general-purpose AD tools to preserve performance tweaks through the differentiation process.

Keywords: Automatic Differentiation, Polyhedral Compiler, Parallelisation, Vectorisation