Jack

Elemental's HermitianEig routines are already wrapped. I'm using them,
and this seems to work fine.

I will need about 10^5 of the 10^6 eigenvalues, so HermitianEig seems
to be the way to go.

Thanks for Elemental!

-erik


On Tue, Mar 8, 2016 at 11:23 PM, Jack Poulson <[email protected]> wrote:
> Scratch that; I misread your request as I was forwarded a link to this with
> the title "Big SVD".
>
> You should look into wrapping Elemental's HermitianEig and/or the
> appropriate routine from ELPA, unless you only want a small number of modes,
> in which case a Krylov subspace technique would likely be preferred.
>
> Jack
>
>
> On Tuesday, March 8, 2016 at 8:20:10 PM UTC-8, Jack Poulson wrote:
>>
>> There were several recent extensions to Elemental's SVD support, as
>> detailed here:
>> https://github.com/elemental/Elemental/issues/125
>>
>> In particular, FULL_SVD, THIN_SVD, COMPACT_SVD, and PRODUCT_SVD are now
>> all supported, where the latter should used if you do not need small
>> triplets as it uses a more parallelizable reduction to tridiagonal form of
>> A^H A rather than a reduction to bidiagonal form of A (as well as a more
>> parallelized MRRR Hermitian tridiagonal eigensolver instead of a QR
>> algorithm for the bidiagonal SVD).
>>
>> With that said, it's worth asking if your friend only needs the top k
>> triplets for modest k, as something like Jiahao and Andreas's TSVD
>> implementation on top of a distributed dense matrix-vector product likely be
>> the way to go (otherwise, a wrapper to SLEPc would be warranted).
>>
>> Jack
>>
>> On Tuesday, March 8, 2016 at 12:19:07 PM UTC-8, Tony Kelman wrote:
>>>
>>> May also want to look into Elemental. Last I checked Elemental uses
>>> Scalapack in a handful of places but I don't think it would be the case
>>> here.
>>>
>>>
>>> On Tuesday, March 8, 2016 at 8:19:15 AM UTC-8, Erik Schnetter wrote:
>>>>
>>>> A colleague mentioned to me that he needs to diagonalize (find
>>>> eigenvalues and eigenvectors) of a symmetric dense matrix with n=10^6
>>>> (i.e. the matrix has 10^12 entries). ScaLapack seems to be the way to
>>>> go for this.
>>>>
>>>> I'm happy to see there is
>>>> <https://github.com/JuliaParallel/ScaLAPACK.jl>. Saying its
>>>> documentation is "Spartan" is an understatement. There is a file
>>>> "test/test.jl" that seems to not be included from "runtests.jl", and
>>>> which thus might be intended as example.
>>>>
>>>> The package doesn't pass its tests. If you know more about whether
>>>> this just needs a brush-up or whether major surgery is required I'd
>>>> appreciate feedback.
>>>>
>>>> -erik
>>>>
>>>> --
>>>> Erik Schnetter <[email protected]>
>>>> http://www.perimeterinstitute.ca/personal/eschnetter/



-- 
Erik Schnetter <[email protected]>
http://www.perimeterinstitute.ca/personal/eschnetter/

Reply via email to