Re: [eigen] update

[Benoit Jacob <jacob@xxxxxxxxxxxxxxx>]

Hi Ben,

The plan is to include in eigen2 (possibly as a separate "eigen-sparse") a 
wrapper around the sparse classes from gmm++.

The wrapper should at least:

- provide a nicer API than gmm++ does, for instance overloaded operators.

- provide Qt-style copy-on-write, canceling the overhead of the overloaded 
operators returning by value; thus it'd still not be as good as expression 
templates but already quite good.

- allow mixing sparse types with other eigen2 types, and uniformize the 
API as much as possible.

Regarding the last point, we should at least provide 
conversion constructors; but a nicer and more sophisticated solution would 
be to let both both dense and sparse objects inherit a common curious base 
class, so as to completely unify the API.

I'd very much welcome help with the sparse stuff, as I'm already busy 
enough with the rest :)

Cheers,
Benoit

On Mon, 8 Oct 2007, Schleimer, Ben wrote:

> Hi Benoit,
>  Maybe we talked about this already but are sparse matrices/vectors going to be implemented in
> Eigen2?  I know it won't be able to use expressions but I'm hoping the basic data structures will
> be there (ala gmm++) so we can use it for adjacency graph calculations.
>
> Cheers
> Ben
>
>
> --- Benoît Jacob <jacob@xxxxxxxxxxxxxxx> wrote:
>
> > Hi List,
> >
> > Here's an update (the best part is the last one):
> >
> > 0) I committed a file that I had forgotten to "svn add", sorry about that.
> >
> > 1) Michael contributed the first meta loop unrolling one week ago. Michael,
> > your work is now in CopyHelper.h. I added you copyright on that file.
> >
> > 2) numeric traits, providing a uniform way to deal with various number types.
> >
> > 3) row-vectors are now treated on an equal footing with column vectors. The
> > Vector typedefs are still column vectors, but there now also are RowVector
> > typdefs. The row() method now returns a row vector. Operator[] accesses the
> > coefficiens of a vector in an uniform way, for both row and column vectors.
> >
> > 4) matrix conjugation, transposition, adjunction, trace; vector dot product. I
> > decided to not overload operator|. So do v.dot(w), that's all. I don't think
> > that I want to add a global dot(v,w) after all.
> >
> > 5) I extended meta unrolling to all the loops we have. I could think of some
> > cases where the compiler would have failed to unroll them, so at least now
> > that won't happen. We won't unroll every loop in the future, but these ones
> > were really important.
> >
> > 6) Big reorganization of the header files
> >
> > 7) Optimization: I reversed the order of some loops (like the inner loop of
> > matrix-matrix multiplication) and got a *huge* speedup.
> > Here's the result I get with our benchmark (g++ 4.2.1, Intel Core1 1.66GHz):
> >
> > TVMET: 6.1 seconds
> > Eigen2 with hand-unrolling of the matrix-product: 5.2 seconds
> > Eigen2 with meta-unrolling: 5.5 seconds
> > Eigen2 with reversed meta-unrolling: 3.4 seconds
> >
> > Trying to understand this speedup I ran cachegrind (with only 100000
> > repetitions) and found this difference:
> > without reversing of loops:
> > ==9840== D   refs:       8,648,008  (6,065,012 rd + 2,582,996 wr)
> > with reversing of loops:
> > ==9834== D   refs:       5,048,066  (2,765,055 rd + 2,283,011 wr)
> >
> > Anyway, Eigen2 is now almost twice faster than before -- when it already was
> > faster than TVMET and Eigen1.
> >
> > Cheers,
> > Benoit