Matrices used by S. Williams et al for sparse matrix multiplication on GPUs.
14 matrices were used in the following paper:
S. Williams, L. Oliker, R. Vuduc, J. Shalf, K. Yelick, J. Demmel,
"Optimization of Sparse Matrix-Vector Multiplication on Emerging Multicore
Platforms", Parallel Computing Volume 35, Issue 3, March 2009, Pages
178-194. Special issue on Revolutionary Technologies for Acceleration of
Emerging Petascale Applications.
This same set of 14 matrices was also used in a subsequent technical report by
http://www.nvidia.com/object/nvidia_research_pub_001.html "Efficient Sparse
Matrix-Vector Multiplication on CUDA" Nathan Bell and Michael Garland, in,
"NVIDIA Technical Report NVR-2008-004", December 2008
file Name dim* nnz description
dense2 Dense 2K 4.0M dense matrix in sparse format
pdb1HYS Protein 36K 4.3M protein data bank 1HYS
consph FEM/Spheres 83K 6.0M FEM concentric spheres
cant FEM/Cantilever 62K 4.0M FEM cantilever
pwtk Wind Tunnel 218K 11.6M pressurized wind tunnel
rma10 FEM/Harbor 47K 2.37M 3D CFD of Charleston Harbor
qcd5_4 QCD 49K 1.90M quark propagators (QCD/LGT)
shipsec1 FEM/Ship 141K 3.98M FEM Ship section / detail
mac_econ_fwd500 Economics 207K 1.27M Macroeconomic model
mc2depi Epidemiology 526K 2.1M 2D Markov model of epidemic
cop20k_A FEM/Accelerator 121K 2.62M Accelerator cavity design
scircuit Circuit 171K 959K Motorola circuit simulation
webbase-1M webbase 1M 3.1M Web connectivity matrix
rail4284 LP 4Kx1.1M 11.3M Railways set cover,
(*) the matrix is square if only one dimension listed.
Six of the matrices are nearly identical to the matrices already in the
UF Collection. They are thus not included in the UF Collection:
pwtk Boeing/pwtk. The matrix here differs in only 2 entries.
The pattern is the same. The norm of the difference between
the two is round-off error (5e-26). Tim Davis obtained the
matrix directly from Roger Grimes, and was careful to preserve
each significant bit. The Boeing/pwtk matrix in the UF
Collection is thus the correct one.
qcd5_4 this is the binary pattern of QCD/conf5_4-* and QCD/conf6_0-*,
ids: 1598:1604 in the UF Collection. These are from B.
Medeke, and were orginally collected by R. Boisvert et al.
for the Matrix Market.
rma10 Bova/rma10. The matrix here is binary; Bova/rma10 is not.
The rma10 matrix here is binary pattern of Problem.A +
Problem.Zeros in the UF Collection. Tim Davis obtained
Bova/rma10 directly from the matrix creator, Steve Bova.
scircuit Hamm/scircuit. The pattern is the same. The 1-norm of the
difference is 5e-10, which is round-off error. This matrix was
obtained by Tim Davis from Steve Hamm (at Motorola) directly,
and thus the Hamm/scircuit matrix already in the UF Collection
is the correct version.
shipsec1 similar to DNVS/shipsec1, but different patterns. It appears
that the shipsec1 matrix here is missing many entries in
the lower right quadrant. The pattern differs by 0.9M
entries. The matrix is from C. Damhaug, collected by
rail4284 an exact submatrix of Mittelmann/rail4284. The UF matrix has
exactly 4284 more columns than the rail4284 matrix in this
collection. Let A1 be the Mittelman/rail4284 matrix, and let
A2 be the rail4284 matrix in this collection. Then A2 is
identical to A1(:,m+1:end). The submatrix A1(:,1:4284) of the
matrix in the UF collection is exactly equal to -speye(m).
Tim Davis obtained the matrix directly from Mittelmann's
test set. It's possible that multiple versions exist, or
that the identity was added as the initial basis. In any
case, the rail4284 matrix from Williams' test set is not
added to the UF Collection.
The remaining eight matrices are unique to this set of matrices, but one is of
no interest (dense2). Thus, there are seven matrices in the Williams/ group in
the UF Collection:
I presume the pdb1HYS matrix comes from this source:
Crystal structure of HIV-1 reverse transcriptase in complex with a
polypurine tract RNA:DNA.
Sarafianos, S.G., Das, K., Tantillo, C., Clark Jr., A.D., Ding,
J., Whitcomb, J.M., Boyer, P.L., Hughes, S.H., Arnold, E.
Journal: (2001) EMBO J. 20: 1449-1461
Search Related Articles in PubMed
We have determined the 3.0 A resolution structure of wild-type HIV-1
reverse transcriptase in complex with an RNA:DNA oligonucleotide whose
sequence includes a purine-rich segment from the HIV-1 genome called the
polypurine tract (PPT). The PPT is resistant to ribonuclease... [ Read
More & Search PubMed Abstracts ] We have determined the 3.0 A resolution
structure of wild-type HIV-1 reverse transcriptase in complex with an
RNA:DNA oligonucleotide whose sequence includes a purine-rich segment from
the HIV-1 genome called the polypurine tract (PPT). The PPT is resistant
to ribonuclease H (RNase H) cleavage and is used as a primer for second
DNA strand synthesis. The 'RNase H primer grip', consisting of amino
acids that interact with the DNA primer strand, may contribute to RNase H
catalysis and cleavage specificity. Cleavage specificity is also
controlled by the width of the minor groove and the trajectory of the
RNA:DNA, both of which are sequence dependent. An unusual 'unzipping' of 7
bp occurs in the adenine stretch of the PPT: an unpaired base on the
template strand takes the base pairing out of register and then, following
two offset base pairs, an unpaired base on the primer strand
re-establishes the normal register. The structural aberration extends to
the RNase H active site and may play a role in the resistance of PPT to
RNase H cleavage.