... and the winner for now is:
it solves my (non_only_diagonal, non_symmetric) test sparse system
*incredibly* faster than vnl_svd with identical results, even for
different tries with low zero_out_tol (whatever it is).
matlab gets unacceptably slow above 5,000 x 5,000 - I dont have patience
to wait, so no benchmark (ctrl^c) for matlab, vnl_lsqr solves it within
a minute, but with the following remark:
vnl_lsqr.cxx : The equations A*x = b are probably compatible. Norm(A*x
- b) is as small as seems reasonable on this machine.
which can be a key to very fast results (because my test system is too
naive). This all was only *test* for not-comming-from-real-life Ax=b
problem. I will be using it to solve real network problems within a few
next weeks - I will let you know how results from lsqr, svd and matlab
thank you for all suggestions and help,