Automatic Performance Tuning for Fast Fourier Transforms

Abstract

In this paper we discuss architecture-specific performance tuning for fast Fourier transforms (FFTs) implemented in the UHFFT library. The UHFFT library is an adaptive and portable software library for FFTs developed by the authors. We present the optimization methods used at different levels, starting with the algorithm selection used for the library code generation and ending with the actual implementation and specification of the appropriate compiler optimization options. We report on the performance results for several modern microprocessor architectures.