Despite the astonishing increases in processor performance over the last forty years, delivered application performance remains a critical issue for many important problems. Compilers play a critical role in determining that performance. A modern optimizing compiler contains many transformations that attempt to increase application performance. However, the best combination of transformations is an application-specific issue [4, 5]. This paper details an experiment in which we used techniques from our work on adaptive compilation in an attempt to duplicate the application and processor specific optimization attained by the ATLAS system [6]. Our goal was to show that an adaptive compiler can achieve results similar to those achieved by ATLAS by using more computer time and less human effort.

Command-line ags provide the developer with some control over a compiler's optimization process. We selected the MIPSpro compiler on an R10000 processor and used an adaptive feedback system to nd the best blocking factors for the matrix-matrix multiply code in Linpack and ATLAS, DGEMM. (We found that the other command line ags either had little impact on DGEMM's performance or had an obvious best setting.) Our experiments show that the adaptive approach: derives good blocking factors for DGEMM at various input sizes; picks better blocking factors, in general, than the MIPSpro compiler; and achieves performance that is competitive with ATLAS (for much less human effort). By automating this kind of tuning, adaptive compilation has the potential to make the custom-tuned performance of ATLAS available for a larger class of applications.