学术文献库

Performance models can be very useful for understanding the behavior of applications and hence can help guide design and optimization decisions. Unfortunately, performance modeling of nontrivial computations typically requires significant expertise and human effort. Moreover, even when performed by experts, it is necessarily limited in scope, accuracy, or both. However, since models are not typically available, programmers, compilers or autotuners cannot use them easily to guide optimizations and are limited to heuristic-based methods that potentially take a lot of time to perform unnecessary transformations. We believe that streamlining model generation and making it scalable (both in terms of human effort and code size) would enable dramatic improvements in compilation techniques, as well as manual optimization and autotuning. To that end, we are building the Meliora code analysis infrastructure for machine learning-based performance model generation of arbitrary codes based on static analysis of intermediate language representations. We demonstrate good accuracy in matching known codes and show how Meliora can be used to optimize new codes though reusing optimization knowledge, either manually or in conjunction with an autotuner. When autotuning, Meliora eliminates or dramatically reduces the empirical search space, while generally achieving competitive performance.