Re: [eigen] Student contribution

[ Thread Index | Date Index | More Archives ]

Le 2013-06-14 16:04, Benoit Jacob a écrit :

2013/6/14 BRUN Gauthier <brunga@xxxxxxxxxxxxxx>

That's surprising! I've been away for a while, but I thought that the
natural thing to do for SVD on integer matrices would be to prevent
that with a static_assert, so that that would give a compilation
We could do that in the SVDBase class.

Why REALTIME? It is more usual, for performance measurements, to
measure per-process (or per-thread, in single-threaded code) CPU time,
which clock_gettime allows.
Other important pieces of information to reproduce your results include compiler version and 32/

Ok, we have to do more performance test, so we will change that settings
Concerning the details about the compiler and the OS :

 icpc version 13.1.2 (gcc version 4.7.0 compatibility)
 Linux 3.5.0-32-generic #53-Ubuntu x86_64 x86_64 x86_64 GNU/Linux

I think that actually implementing the SVD code (without resorting to JacobiSVD), and supporting complex matrices, should be the two main missing prerequisites to consider including this in Eigen.

We are going to try to fix the bug on the complex, but we cannot guaranty that it will be finished, if we do our version is still faster than JacobiSVD (x1,6) on large matrices. Concerning the last part of the algorithm which would reduce the complexity of the computation and the computation time even more, we don't plan to finished it, because we don't have enough time left.

We would like to provide you a code that could be easily re-use by someone else to finish the BDCSVD or at least to be integrated in the unsupported branch. Thus if you have any requirement concerning coding style or naming or documentation we will correct them.

PS : @Gael, we will do our best to finish the algorithm but the special SVD algorithm, which is not included in our official bitbucket, is far from working yet (we have issues with the math implied) . As a temporary fix we thought we could either redirect the computations and make them done by Jacobi for complexes or mention the limits of our algorithm in the doxygen documentation generated.

Team GL27
Jean Ceccato, Pierre Zoppitelli, Gauthier Brun, Nicolas Carre

Mail converted by MHonArc 2.6.19+