Automating Compiler-Directed Autotuning for Phased Performance Behavior

We describe an integration of the CHiLL compiler with OpenTuner to reduce the programmer burden in using autotuning. We use as a case study optimizing the smooth operator and its associated stencil computations in the context of Geometric Multigrid (GMG), a hierarchical linear solver that operates in multiple grid resolutions (levels). Smooth is the most performance-critical operation that runs multiple times at each grid level and effectively performs a relaxation of the approximated solution at a given grid resolution. This computation poses a particular challenge for autotuning, as the desired optimization strategy varies at different grid resolutions within the same application execution. Even though the compiler provides a number of standard and domain-specific optimizations for stencil computations, it is challenging for a programmer to decide which optimizations to perform and implement all the steps of the autotuning search. In this paper, we make the following contributions to simplify this process and make it possible to configure the application for its different phases: (1) we provide an interface (called a superscript) to concisely describe a search space and automatically generate CHiLL transformation recipes; and, (2) we use OpenTuner tailored to CHiLL transformation recipes to employ sophisticated heuristic algorithms that manage the computational complexity of search. We demonstrate performance that far exceeds that of fixed optimization strategies, while only sampling a tiny subset of the autotuning search space.