A genetic algorithm based autotuning approach for performance and energy optimization

Autotuning is an empirical optimization approach in which the configuration space of an algorithmic code is explored in a systematic manner for a variety of software and hardware parameters. The objective of such autotuning is to reduce the computational time and/or energy requirements of the generated code. We develop a genetic algorithm based autotuning strategy that can be used for optimizing performance or energy or a combination thereof. The main advantage of our approach is that the number of possible compilations and executions that are explored in the configuration space is substantially smaller than exhaustive search. We demonstrate the usefulness of our approach to the underlying small matrix multiplication routines in spectral element solvers. The latter are an important class of higher order methods that are expected to be computationally intensive portion of next generation of large scale CFD simulations. Our experimental results were conducted of a variety of platforms. On AMD Fusion, for example, the genetic algorithm is able obtain 34% improvement in performance and 37% reduction in energy consumption over existing versions of the code. Further, a very small fraction of the entire configuration space needs to be explored.