Doctor of Philosophy

dissertationEmerging trends such as growing architectural diversity and increased emphasis on energy and power efficiency motivate the need for code that adapts to its execution context (input dataset and target architecture). Unfortunately, writing such code remains difficult, and is typically attempted only by a small group of motivated expert programmers who are highly knowledgeable about the relationship between software and its hardware mapping. In this dissertation, we introduce novel abstractions and techniques based on automatic performance tuning that enable both experts and nonexperts (application developers) to produce adaptive code. We present two new frameworks for adaptive programming: Nitro and Surge. Nitro enables expert programmers to specify code variants, or alternative implementations of the same computation, together with meta-information for selecting among them. It then utilizes supervised classification to select an optimal code variant at runtime based on characteristics of the execution context. Surge, on the other hand, provides a high-level nested data-parallel programming interface for application developers to specify computations. It then employs a two-level mechanism to automatically generate code variants and then tunes them using Nitro. The resulting code performs on par with or better than handcrafted reference implementations on both CPUs and GPUs. In addition to abstractions for expressing code variants, this dissertation also presents novel strategies for adaptively tuning them. First, we introduce a technique for dynamically selecting an optimal code variant at runtime based on characteristics of the input dataset. On five high-performance GPU applications, variants tuned using this strategy achieve over 93% of the performance of variants selected through exhaustive search. Next, we present a novel approach based on multitask learning to develop a code variant selection model on a target architecture from training on different source architectures. We evaluate this approach on a set of six benchmark applications and a collection of six NVIDIA GPUs from three distinct architecture generations. Finally, we implement support for combined code variant and frequency selection based on multiple objectives, including power and energy efficiency. Using this strategy, we construct a GPU sorting implementation that provides improved energy and power efficiency with less than a proportional drop in sorting throughput