Pruning the Optimization Search Space Using Architecture-aware Cost Models⋆

Abstract. In recent years, a number of strategies have emerged for em-pirically tuning applications to different architectures. Although quite successful for certain domains, empirical tuning is yet to gain wide ac-ceptance as a viable strategy in high-performance computing. The prin-cipal bottleneck in this regard is the prohibitively large search space that needs to be explored in order to discover the best program variants for different architectures. Although there have been some efforts at using cache models in pruning the search space for kernels, the optimization search space of whole applications still remains mostly intractable. In this paper, we propose a novel search space pruning strategy. Our ap-proach is to identify architecture-dependent parameters within compiler cost models and search for the best values of those parameters. We have implemented this strategy for exploring the search space of loop fusion and tiling for whole applications. Preliminary experiments suggest that this approach of tuning for architecture-dependent model parameters is highly effective in reducing the size of the optimization search space while incurring only a small performance penalty.