ETA for SVD eigenvalue

[clusty]

Hey,

I recently discovered PLASMA and am very excited about it.
Unfortunately most of my computations involve SVDs and eigenvalue problems.
Is there some estimate as to when such a thing could become available? (even beta version)

[admin]

We're working on it.
Unfortunately, all I can say right now is that it is not likely to happen this year.
Jakub

[admin]

The research is going very well and we have very high hopes for fast software.
Unfortunately, that is not going to see the light before the Supercomputing conference in November.
We don't plan to release SVD code before that.
Jakub

[wsawyer]

Has there been any progress toward an SVD implementation?

Thanks, --Will

[haidar]

Hi William,
We are working on the Eigensolver and the SVD,
and hope that they will be ready in 4-5 months.
Do you have any specific applications ?
Thanks
Azzam

[ltaief]

Dear Will,
No SVD implementation yet but we are getting closer.
There is still a stage during the bidiagonal reduction which needs to
be optimized before a PLASMA release with SVD happens.
However, if you are a fan of hybrid computing with GPU accelerators,
the next MAGMA release (SC10) will contain eigenvalue and SVD solvers.
Regards,
Hatem

[Sebastian123]

Can I rewarm this forum to get to know the status of the SVD implementation?

The current PLASMA SVD routine PLASMA_dgesvd does not support the computation of the singular vectors as for PLASMA 2.4.6. Is there any work going on in this direction?

I also implemented my own SVD code (based on Eigen's MKL wrappers) using the QR subroutines PLASMA_dgeqrf and PLASMA_dorgqr, but my tests reveal that those are not working in parallel yet. Is that true?

Best regards
Sebastian

[admin]

Sebastian123 wrote:
> Can I rewarm this forum to get to know the status of the SVD
> implementation?
>
> The current PLASMA SVD routine PLASMA_dgesvd does not support the
> computation of the singular vectors as for PLASMA 2.4.6. Is there any work
> going on in this direction?
>

The next PLASMA release (SC12 - mid November) will include routines for solving the singular value problem and computing the singular vectors, most likely also routines for computing a subset of the singular vectors in a shorter time.

> I also implemented my own SVD code (based on Eigen's MKL wrappers) using
> the QR subroutines PLASMA_dgeqrf and PLASMA_dorgqr, but my tests reveal
> that those are not working in parallel yet. Is that true?
>

PLASMA QR routines are parallel.
If you don't see the performance, there can be a couple of reasons.
If your matrix is "tall and skinny" you can switch to PLASMA's QR routines based on tree reductions.
Look at the slides: http://web.eecs.utk.edu/~kurzak/tutorials/ISC12/02_Kurzak_PLASMA.pdf
Slides 48, 49 explain the concept.
Slide 72 explains how to use that mode.
You may also have to tune the tile size.

Most likely you have a different problem, though.
If you are using PLASMA with MKL, you have to set MKL_NUM_THREADS to 1.
PLASMA will be messed up if you don't.
If you do, MKL is sequential.
It is a genuine problem of interfacing multithreaded libraries, to which we don't have a solution yet.
Best,
Jakub

[Sebastian123]

Thanks for the response.

I found my problem in QR - The Eigen MKL bindings are erroneous and my PLASMA binding code was based on that one.

I am very fine with the November release and do look forward testing it! That's indeed very great news for me!

Best regards
Sebastian