i just got a good news from NA-digest, it is usefull
Version 1.2 of the “librsb” library for high performance sparse matrix
computations based on the “Recursive Sparse Blocks” (RSB) format is
out and available for download.
librsb is mainly for iterative methods implementors; it features cache
efficient, multithreaded sparse matrix-vector multiplication, and it
is especially fast in the symmetric and transposed variants. Many
options and further matrix operations are available.
Both Sparse BLAS and the own interfaces are supported (C, Fortran).
Transparently usable in GNU Octave via the sparsersb package/command.
Main improvements of librsb-1.2 are:
– faster runtime autotuning of structure and threads for sparse
matrix-vector/matrix-matrix multiply and triangular solve: now uses
a parallel merge/subdivision based algorithm
– speeded up triangular solution of a matrix by multi-vector
– introduced extension BLAS property blas_rsb_autotune_next_operation
to trigger auto tuning at the next USMV/USMM call
– collection/report of time spent in librsb routines now possible (a
simple built-in profiler)
– further Encapsulated PostScript (EPS) matrix rendering options
– improved benchmark program for large parameter studies (converts and
reuses matrices rather than re-reading them from disk)
– binary performance records write/read/combine support, also with
Intel MKL CSR built-in comparison; LaTeX tables output and stats
Source code (LGPLv3 licensed), documentation, change log and extensive
benchmark results available on: http://sourceforge.net/projects/librsb
Filed under: ICT |