A hierarchically blocked Jacobi SVD algorithm for single and multiple graphics processing units
University of Zagreb, Faculty of Mechanical Engineering and Naval Architecture, I. Lucica 5, HR-10000 Zagreb, Croatia
arXiv:1401.2720 [cs.NA], (13 Jan 2014)
@article{2014arXiv1401.2720N,
author={Novakovic}, V.},
title={"{A hierarchically blocked Jacobi SVD algorithm for single and multiple graphics processing units}"},
journal={ArXiv e-prints},
archivePrefix={"arXiv"},
eprint={1401.2720},
primaryClass={"cs.NA"},
keywords={Computer Science – Numerical Analysis, Computer Science – Distributed, Parallel, and Cluster Computing, 65Y05 (Primary) 65Y10, 65F15 (Secondary), G.1.0, G.1.3, G.4},
year={2014},
month={jan},
adsurl={http://adsabs.harvard.edu/abs/2014arXiv1401.2720N},
adsnote={Provided by the SAO/NASA Astrophysics Data System}
}
We present a hierarchically blocked one-sided Jacobi algorithm for the singular value decomposition (SVD), targeting both single and multiple graphics processing units (GPUs). The blocking structure reflects the levels of GPU’s memory hierarchy. The algorithm may outperform MAGMA’s dgesvd, while retaining high relative accuracy. To this end, we developed a family of parallel pivot strategies on GPU’s shared address space, but applicable also to inter-GPU communication. Unlike common hybrid approaches, our algorithm in a single GPU setting needs a CPU for the controlling purposes only, while utilizing GPU’s resources to the fullest extent permitted by the hardware. When required by the problem size, the algorithm, in principle, scales to an arbitrary number of GPU nodes. The scalability is demonstrated by more than twofold speedup for sufficiently large matrices on a Tesla S2050 system with four GPUs vs. a single Fermi card.
January 14, 2014 by hgpu
