Re: Does Octave 4.2.1 use SPQR or CXSparse for sparse QR factorizations?

[Dmitri A. Sergatskov]

On Tue, Sep 5, 2017 at 10:27 PM, Joseph Young <address@hidden> wrote:

    Does Octave 4.2.1 use SPQR or CXSparse for sparse QR factorizations?
Looking at the code sparse-qr.cc, it appears as though the QR
factorization from CXSparse is used.  If this is the case, my next
question is why is this routine preferred over those from SPQR?  Looking
at the changelog, it looks like there used to be routines for spqr, but
they were removed on Octave 3.6.  My impression was that these routines
were just merged into the qr routine, but I'm not sure if they used SPQR
or CXSparse when they were implemented.  Part of the reason I'm asking
is that I believe that SPQR should give superior performance.  More than
that, the current sparse implementation of qr disallows permutation,
[Q R P]=qr(A).  This is pretty useful in a number of contexts such as
extracting rank information as well as various low rank solves and
factorizations.  I believe that MATLAB already leans on the SPQR
routines and allows the factorization with permutations.

    That said, I don't understand the backstory here, so I'm writing to
better understand what's currently being done and why.  Thanks for the
help!

​David Bateman ​was a key contributor to the sparse matrix code for octave.

I believe he was also the person to the CSparse -> CXSparse translator

(for the actual SuiteSparse library).

Here is a paper you may find interesting:

http://www.sce.carleton.ca/faculty/adler/publications/2006/bateman-adler-octave2006-sparse-matrix.pdf

Unfortunately David does not work on octave any more.

 

Joe

​Sincerely,

Dmitri.

--

​