9 Feb
2015
9 Feb
'15
6:55 p.m.
Just a quick note about performance: since LAPACK is heavily based on BLAS level 3, and since BLAS level 3 are *usually* tuned for large matrices, bear in mind that using it repeatedly for operations on "small" objects (i.e., that fit some levels of cache) might actually slow down the computation. Eigen could also have the advantage of working around this issue, since it tends to perform better for small size problems than the various BLAS (and BLAS-like) alternatives. This (various BLAS implementations vs Eigen vs pure C) is something I started investigating with/through COFFEE, but that I eventually dropped for lack of time I don't actually think this info could be of any help, but you know, just in case -- Fabio