On Fri, Nov 27, 2015 at 3:57 PM, John Peterson <jwpeter...@gmail.com> wrote:

>
>
> On Tue, Nov 24, 2015 at 12:32 AM, David Knezevic <
> david.kneze...@akselos.com> wrote:
>
>> A follow-up comment on this: We added lapack-based SVD in DenseMatrix
>> before we included Eigen in contrib, and I don't think it was ever used
>> very extensively (and it seems to be buggy, as shown by Jonas's example).
>>
>
> It should be fixed as of https://github.com/libMesh/libmesh/pull/734
>

Great, thanks.



>
> I find the lapack API to be quite opaque,
>>
>
> Uhh, yeah, that's an understatement.
>
>
>> so the less we rely on it the better IMO. So now that we include Eigen, I
>> would be in favor of deprecating the lapack-based SVD and EVD in
>> DenseMatrix and suggesting that users use Eigen instead, as Jonas did in
>> his example. I'm not sure that we need a wrapper for Eigen's SVD or EVD in
>> DenseMatrix --- it's just as easy (and more general) for users to call
>> Eigen themselves however they need to.
>>
>
> Deprecating sounds fine to me, I don't think MOOSE uses these routines at
> all, it has a ColumnMajorMatrix class that implements its own Lapack calls
> for eigenvalue computations.
>


OK. I vote for deprecating, rather than having these SVD and EVD methods in
DenseMatrix which are apparently unused. Not a big deal either way though...




> A question for people that know more about Eigen than me: Could
> DenseMatrix<T> just be changed to an Eigen::MatrixXf or something similar?
> There might be some trickiness with interfacing with DenseSubMatrix, but it
> would be nice to get all the fast numerics stuff "for free"...
>

It seems to me that this would be possible, but it might be quite an
extensive change? Whenever I need to use Eigen, I just make an Eigen "copy"
of the DenseMatrix and use the copy. A bit ugly, but very simple.

David
------------------------------------------------------------------------------
_______________________________________________
Libmesh-users mailing list
Libmesh-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libmesh-users

Reply via email to