High performance implementation of the inverse TFT

Abstract

The inverse truncated Fourier transform (ITFT) is a key component in the fast polynomial and large integer algorithms introduced by van der Hoeven. This paper reports a high performance implementation of the ITFT which poses additional challenges compared to that of the forward transform. A general-radix variant of the ITFT algorithm is developed to allow the implementation to automatically adapt to the memory hierarchy. Then a parallel ITFT algorithm is developed that trades off small arithmetic cost for full vectorization and improved multi-threaded parallelism. The algorithms are automatically generated and tuned to produce an arbitrary-size ITFT library. The new algorithms and the implementation smooths out the staircase performance associated with power-of-two modular FFT implementations, and provide significant performance improvement over zero-padding approaches even when high-performance FFT libraries are used.