high performance computing on graphics processing units: hgpu.org

How do we build maintainable, robust, and performance-portable scientific applications? This thesis argues that the answer to this software engineering question in the context of the finite element method is through the use of layers of Domain-Specific Languages (DSLs) to separate the various concerns in the engineering of such codes. Performance-portable software achieves high performance on multiple diverse hardware platforms without source code changes. We demonstrate that finite element solvers written in a low-level language are not performance-portable, and therefore code must be specialised to the target architecture by a code generation framework. A prototype compiler for finite element variational forms that generates CUDA code is presented, and is used to explore how good performance on many-core platforms in automatically-generated finite element applications can be achieved. The differing code generation requirements for multi- and many-core platforms motivates the design of an additional abstraction, called PyOP2, that enables unstructured mesh applications to be performance-portable. We present a runtime code generation framework comprised of the Unified Form Language (UFL), the FEniCS Form Compiler, and PyOP2. This toolchain separates the succinct expression of a numerical method from the selection and generation of efficient code for local assembly. This is further decoupled from the selection of data formats and algorithms for efficient parallel implementation on a specific target architecture. We establish the successful separation of these concerns by demonstrating the performance-portability of code generated from a single high-level source code written in UFL across sequential C, CUDA, MPI and OpenMP targets. The performance of the generated code exceeds the performance of comparable alternative toolchains on multi-core architectures.