HPC for Idealists with Deadlines: Pragmatic Abstractions for High Performance

Development of high performance code often necessitates compromise. Lowered computational cost comes at the expense of other worthy goals, including intelligibility and maintainability of programs. We discuss a "ladder" of tools ranging from low- to high-level, aiming to reduce these sacrifices in machine independence, readability, and separation of concerns, all while enabling "near-handwritten" performance to be attained. These tools include PyOpenCL, offering access to the OpenCL compute abstraction and open-source implementations thereof, including pocl. They include loopy, a polyhedrally-based code transformation tool for CPU and GPU code. And they include pytato, which captures data flow graphs of numpy-compatible array computations for transformation and processing via loopy. Moderate-to-large-scale examples illustrate successful uses of these tools.