Open carstenbauer opened 5 years ago
It's worth noting that there are also preconditioning drivers in LAPACK (with names corresponding to gejsv!
). These can make the Jacobi scheme runtime-competitive with the other methods. They have rather complicated interfaces so getting the wrapper logic right may take some effort. Would it make sense to do this work in a separate package first (question for the main LinearAlgebra developers)? That would allow for more comprehensive public testing and perhaps some nice demonstrations.
Yes, a separate package definitely sounds better. Could be in JuliaLinearAlgebra org.
In the course of #31057, which adds a
algorithm
ormethod
keyword tosvd
andsvd!
, it would be nice to have access to Jacobi methods (gesvj
in LAPACK) for SVD decomposition as well.See the discussion on discourse Calling Lapack’s Jacobi SVD - dgesvj and a draft thereout: https://gist.github.com/RalphAS/6a80a1dd4039153c8c94fc27a4938665.
Unfortunately, this is a bit above my head so I'm hoping for someone else to actually implement this. I assume this isn't really difficult if one knows what one's doing.
@andreasnoack, maybe you would be willing to take a look? I saw that you seem to have access to a
gesvdj!
(is this the same asgesvj
?) in CuArrays.jl already.