A PORTABLE SHORT VECTOR VERSION OF FFTW
Franz Franchetti · 2006
Abstract. This paper presents a portable short vector extension for the popular FFT library Fftw. Fftw is a freely available portable FFT software-library that achieves top performance across a large number of platforms. The newly developed extension enables the utilization of short vector extensions like Intel’s SSE and SSE 2 as well as Motorola’s AltiVec for any problem sizes. The method is independent of the machine’s vector length. Experimental results show good speed-up for arbitrary problem sizes and excellent speedup for problem sizes of powers of two and multiples of 16. The newly developed short vector version of Fftw compares favorably with the hand-tuned vendor library Intel MKL on IA-32 compatible machines (Intel Pentium III and 4, and AMD Athlon XP). 1.