HiLO: High Level Optimization of FFTs

Abstract

As computing platforms become more and more complex, the task of optimizing performance critical codes becomes more challenging. Recently, more attention has been focused on automating this optimization process by making aggressive assumptions about the algorithm. Motivated by these trends, this paper presents HiLO, the high-level optimizer for FFT codes. HiLO blends traditional optimization techniques into an optimization strategy tailored to the needs of FFT codes and outputs C code ready to be further optimized by the native compiler. It has already been shown that such high-level transformations are important to coax the native compiler into doing its job well. HiLO provides a more appropriate platform for researching these phenomena, suggests an optimization strategy that improves on previous approaches, and shows that even software pipelining at the C level can improve the final binary’s performance.

This work was supported in part by the National Science Foundation under grants CCR 01-21401 ITR and 03-25687 ITR/NGS; by DARPA under contract NBCH30390004; and by gifts from INTEL and IBM. This work is not necessarily representative of the positions or policies of the Army or Government.