Hi,

I just started porting some of my code to eigen2. Most of it was done with uBlas + LAPACK, and I have to say I'm very happy about the speed achieved.

Two things I would like to see in eigen2, a basic statistics module and LAPACK bindings (for missing methods). Is anyone working on any of those? I'm still getting familiar with the internals of eigen2, but am more than willing to help implementing such modules.

hi, thank you for your interest in Eigen. To answer your question, currently nobody is working on the stats module. We started a discussion a couple of months ago and we are waiting volunteers (http://eigen.tuxfamily.org/index.php?ti ... ics_module)

About LAPACK it seems your motivation is more feature wise than performance wise, so can you tell us what are the missing features ? (I thought we already covered most of lapack)

I am definitely interested in the features in the Statistics module in the to-do list. I'm still getting to understand how eigen2 works, but with a little help I'm certain I can implement those points.

As for LAPACK, my main objective was using the methods on packed matrices (I sort of asked about this on irc, my nick there is sombranonius). On most of my code, I care the most about covariance matrices (read symmetric for the purpose of this msg). For small sizes it doesn't make too much of a difference storing the whole thing or just the upper or lower triangular part, but I need to store several large such matrices.
If eigen2 eventually supports packed storage then I will be more than happy to replace all the LAPACK methods I currently use.

Support for various special matrix forms and the corresponding packed storages, is one of our biggest TODO's for Eigen 2.1. But it's going to take us a few months. I didn't think of the selfadjoint case, the packed storage looks very awry in that case, but since it saves 50% memory usage, indeed it's worth thinking...

I will then wait until 2.1 is out to replace some of the LAPACK functions I'm currently using for packed matrices. For the time being I can keep part of my uBlas code.

In the meantime I'll start looking into implementing the features in the to-do list regarding the stats module.

bjacob wrote:Support for various special matrix forms and the corresponding packed storages, is one of our biggest TODO's for Eigen 2.1. But it's going to take us a few months. I didn't think of the selfadjoint case, the packed storage looks very awry in that case, but since it saves 50% memory usage, indeed it's worth thinking...

In case this is still on the agenda for 2.1, have a look at LAPACK's new rectangular packed format, which is superior in speed to the plain old packed storage. Also, their other changes are worth having a look, such as the new SVD algorithm.

Thanks for the pointer: the rectangular packet format is indeed a simple and good idea!

For the new SVD, I can't say as they don't detail how it works, but his:

This release only include SINGLE PRECISION and DOUBLE PRECISION. In other words, we do not have complex support. Also the matrix in input needs to be with M.GE.N.

means that it's not going to be (at least not in the next LAPACK release) a viable replacement for what we'll have in Eigen 2.1.