1

As the CUDA programming model matures, I wonder if anyone is aware of any available research code or open-source libraries that implement sparse Cholesky factorizations on NVIDIA GPUs.

On May 2012, I 've been pointed to the following literature by V. Volkov

[1] Christen et al., 2007 General-Purpose Sparse Matrix Building Blocks using the NVIDIA CUDA Technology Platform, http://www.cs.jhu.edu/~misha/ReadingSeminar/Papers/Christen07.pdf

[2] Krawezik and Poole, 2009, Accelerating the ANSYS Direct Sparse Solver with GPUs, http://saahpc.ncsa.illinois.edu/09/papers/Krawezik_paper.pdf

[3] Yu et al., 2011, A CPU-GPU hybrid approach for the unsymmetric multifrontal method, http://www.sciencedirect.com/science/article/pii/S0167819111001293

[4] George et al., 2011, Multifrontal Factorization of Sparse SPD Matrices on GPUs, http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6012808

[5] Lucas et al., 2012, Multifrontal Sparse Matrix Factorization on Graphics Processing Units, ftp://ftp.isi.edu/isi-pubs/tr-677.pdf

Nikolaos Giotis
  • 281
  • 2
  • 23
  • 1
    implementing sparse Cholesky factorization on the GPU tends to be a fairly complicated process, which does not always exploit the maximum potential for parallel computations. haters feel free to down-vote! – Nikolaos Giotis Jan 02 '13 at 15:21
  • 1
    Have you had a chance to check out CHOLMOD (Tim Davis's Cholesky factorization package that offloads dense linear algebra calls to the GPU): http://www.cise.ufl.edu/research/sparse/cholmod/ Please note I am not personally familiar with the package, a colleague pointed it out to me. – njuffa Jan 02 '13 at 20:26
  • njuffa Tim Davis's work is great! im talking about GPU implemetations though preferabl in CUDA – Nikolaos Giotis Jan 03 '13 at 00:33
  • @njuffa I should have followed your link back then.. just realised that Tim Davis partnered with NVIDIA, so http://www.cise.ufl.edu/research/sparse/cholmod/ is a straight-forward option for a single GPU - really hope it works for one Quadro plus Tesla. – Nikolaos Giotis Mar 23 '13 at 02:42

1 Answers1

1

Tim Davis work suitesparse and nVIDIA's acceleration cusparse. AMD might support Cholesky decomposition in the near future.

Nikolaos Giotis
  • 281
  • 2
  • 23