An Effective Empirical Search Method for AutomaticSoftware Tuning

Abstract Empirical software optimization and tuning is an ac-tive research topic in the high performance computing research community. It is an adaptive sys-tem to generate optimized software using empirically searched parameters. Due to the large parametersearch space, an appropriate search heuristic is an essential part of the system. This paper describes aneffective search method that can be generally applied to empirical optimization. We apply this method toATLAS (Automatically Tuned Linear Algebra Software), which is a system for empirically optimizingdense linear algebra kernels. Our experiments on four different platforms show that the new search schemecan produce parameters that can lead ATLAS to generate a library with better performance