Batch solution of small PDEs with the OPS DSL

In this paper we discuss the challenges and optimisations opportunities when solving a large number of small, equally sized discretised PDEs on regular grids. We present an extension of the OPS (Oxford Parallel library for Structured meshes) embedded Domain Specific Language, and show how support can be added for solving multiple systems, and how OPS makes it easy to deploy a variety of transformations and optimisations. The new capabilities in OPS allow to automatically apply data structure transformations, as well as execution schedule transformations to deliver high performance on a variety of hardware platforms. We evaluate our work on an industrially representative finance simulation on Intel CPUs, as well as NVIDIA GPUs