Fix #10970. Allow LU factorizing non-square sparse matrices.

[andreasnoack]

E.g.

julia> full(lufact(sprandn(5,3,0.5))[:L])
5x3 Array{Float64,2}:
  1.0        0.0        0.0      
  0.0        1.0        0.0      
  0.0        0.56509    1.0      
 -0.300521  -0.673485  -0.0597699
  0.0        0.0        0.0

[tkelman]

Cool, +1

cc @gwhowell

[gwhowell]

Looks good,
Thanks,
Gary

On Mon, Apr 27, 2015 at 2:38 AM, Tony Kelman [email protected]
wrote:

Cool, +1

cc @gwhowell https://github.com/gwhowell

—
Reply to this email directly or view it on GitHub
#11025 (comment).

[ViralBShah]

Is it realistic to backport this?

[andreasnoack]

Yes. I could cherry-pick it cleanly, so I think we should just do it by the end of the week.

[tkelman]

The tests have been moved around to different files, but otherwise the change looks pretty small - depends whether the umfpack code looks the same on release-0.3.

[andreasnoack]

@tkelman I've already tried git cherry-pick and there weren't any conflicts.

[tkelman]

Sounds good, just make sure the added tests get run under release-0.3's file organization.

At this point I wouldn't be opposed to saying all backports to release-0.3 from now on should go through pr's against release-0.3, but I might be the only one who thinks that.

[tkelman]

Wasn't sure if you were going to do the backport, so I did in a6efa59.

git cherry-pick is smarter than I thought it was, it seems to have figured out that the tests have moved around to different files.

[tkelman]

@gwhowell just your luck that we should also have 0.3.8 out momentarily, binaries are building now

[gwhowell]

Hi,
I
was able to recompile 0.3.7 with the lufact nonsquare changes. It's
probably working (returns results, have not yet checked the results).

If I make progress with the column pivoting, I'll let you know.

Another operator I'm interested in is condest. matlab and octave seem to
use the same
algorithm for that (actually estimates the L1 condition number, uses work
of
Higham and Hager). When I get more time I can submit that as a ticket (and
perhaps
give you some code that seems to work). I use condest at runtime to decide
what preconditioners to use.

So thanks,
Gary Howell

On Thu, Apr 30, 2015 at 8:10 PM, Tony Kelman [email protected]
wrote:

@gwhowell https://github.com/gwhowell just your luck that we should
also have 0.3.8 out momentarily, binaries are building now

—
Reply to this email directly or view it on GitHub
#11025 (comment).

[tkelman]

@gwhowell ref #6485