Logo image
Automatic Parallel Library Generation for General-Size Modular FFT Algorithms
Book chapter   Peer reviewed

Automatic Parallel Library Generation for General-Size Modular FFT Algorithms

Lingchuan Meng and Jeremy Johnson
Computer Algebra in Scientific Computing, pp 243-256
2013

Abstract

autotuning FFT library generation modular arithmetic parallelization
This paper presents the automatic library generation for modular FFT algorithms with arbitrary input sizes. We show how to represent the transform and its algorithms at a high abstraction level. Symbolic manipulations and code optimizations that use rewriting systems can then be systematically applied to generate a library with recursive function closure. The generated library is automatically optimized for the target computing platforms, and is intended to support modular algorithms for multivariate polynomial computations in the modpn library used by modpn. The resulting scalar and vector codes provide comparable speedup to the fixed-size code presented in [LJF10], which is an order of magnitude faster over the hand-tuned modpn library. Thread-level parallelism has also been utilized by the generated library and delivers additional speedup.

Metrics

9 Record Views
4 citations in Scopus

Details

Logo image