Re: [eigen] Fwd: Re: Student contribution

[ Thread Index | Date Index | More Archives ]

2013/6/17 BRUN Gauthier <brunga@xxxxxxxxxxxxxx>


We are proud to announce that thanks to Gael's fix, our algorithm now works fine with any dynamic matrices, including complexes one.

That's great news!
It passes the same 'is_Approx' tests as Jacobi and that is not surprising since the loss of precision is caused by the deflation, actually functional but commented, and the third part of the algorithm that is yet to be implemented. We are now focusing on optimizing the existing code, cleaning some of the functions, and making the work of further developers easier. That includes sources for the next step, especially a paper to solve the Particle simulation problem and code comment to explain where to add the last part of the algorithm.

Do you have any particular thing you want us to change in the BDCSVD ?

Regarding the specifics of the current code --- it's up to Gael. At this point, I would simply recommend merging your code into unsupported/ and, rather than code changes, in my humble opinion, the most important thing that you can do to maximize the likeliness that someone else would step up to continue this work, is that you add some documentation, maybe in the form of a separate DESIGN or TODO file, carefully explaining the approach taken here, how the divide&conquer step works, and the work that remains to be done to stop using JacobiSVD and instead implement your own bidiagonal d&c SVD step.

References for "the right d&c SVD way" include:
This article:
LAPACK sources:

Thanks again for the good work made by the students, and let's hope that someone continues this work.

Best regards,

Team GL27

Jean Ceccato, Pierre Zoppitelli, Gauthier Brun, Nicolas Carre

Mail converted by MHonArc 2.6.19+