Schedule Jan 17, 2009
Particle-Mesh Methods on Graphics Processing Units: The Application of Emulated Double-Precision to Improve Speed and Precision
Otonyo Mangete (Drexel Univ.)

The Particle-Mesh (PM) method is a standard method of simulating gravitating systems such as stars and galaxies and simulations of dark matter distributions in cosmological simulations. The Fast Fourier Transform (FFT) is an integral part of the PM method and its speed and accuracy are vital to simulations of this kind.

For large simulations, the FFT is performed in double precision. The Nvidia GPU has been used recently as a low-cost, high-performance computing device due to its speed and low power consumption. There exists a library for computing FFTs on the GPU (called CUFFT) but its speed is severely reduced when using double-precision. This limits the range of simulations that can be carried out on the GPU.

An Emulated Double-Precision FFT (EDP-FFT) is constructed for use on a Nvidia GPU to replace the slower (in double precision) CUFFT library. The EDP-FFT functions use real and complex numbers. It also has one-dimensional and multi-dimensional FFT functions.

Test simulations of a self-gravitating system of galaxies computed using the PM method is constructed along with test suites comparing the EDP-FFT to the Fastest Fourier Transform in the West (FFTW) library and the CUFFT library in double precision. The performance characteristics of all three libraries are compared in this presentation.

View poster as pdf.

Author entry (protected)