Optimizing 3D Convolutions for Wavelet Transforms on CPUs with SSE Units and GPUs
Nanosim (LIG), France
Nanosim (LIG), 2012
@TechReport{rrlig032,
author={Videau, Brice and Marangozova-Martin, Vania and Genovese, Luigi and Deutsch, Thierry},
title={O}ptimizing 3{D} {C}onvolutions for {W}avelet {T}ransforms on {CPU}s with {SSE} {U}nits and {GPU}s},
institution={LIG},
year={2012},
number={RR-LIG-032},
address={Grenoble, France},
typepub={Research Report},
series={Les rapports de recherche du LIG},
type={Research Report},
issn={2105-0422},
pdf={http://rr.liglab.fr/research_report/RR-LIG-032_orig.pdf}
}
Nanosimulations present a big HPC challenge as they present increasing performance demands in heterogeneous execution environments. In this paper, we present our optimization methodology for BigDFT, a nanosimulation software using Density Functional Theory. We explore autotuning possibilities for BigDFT’s 3D convolutions by studying optimization techniques for several architectures. Namely, we focus on processors with vector units and on GPU acceleration. We report on the portability and the performance gains of our approach (speedup x2 on CPU, x5 on GPU) and discuss the relation between algorithmic specifics, architecture and performance.
November 27, 2012 by hgpu
