Performance-Portable Finite-element Computations from High-level Specifications with FFC and PyOP2

Abstract—We present a tool chain for the fully automated synthesis of performance-portable finite-element solvers for mul-ticore and GPGPU platforms from high-level specifications. Our runtime code generation and just-in-time compilation pathway takes finite-element forms in the domain-specific language UFL to low-level code. Automatically generated finite-element assembly kernels are passed to PyOP2, a domain-specific language for mesh-based simulation codes, which acts as an intermediate abstraction layer for executing the numerical kernels in parallel over an unstructured mesh. Easy integration of our tool chain allows transparently adding performance portability to existing simulation codes. PyOP2 [1], [2] is a Python implementation of the unstruc-tured mesh computation framework OP2 [3], which applies numerical kernels in parallel over an unstructured mesh. Ker