Re: [this is the full note] the game's afoot, Watson

[Gerard Jungman <jungman at lanl dot gov>]

Hi. Thanks for your message.

Edward Jason Riedy wrote:
> 
> Also be sure to look at MTL (http://www.lsc.nd.edu/research/mtl/).
> The MTL is a great system for linear algebra libraries, although
> it's license is too restrictive.  Dr. Lumsdaine is at LBL for a
> few more months and will probably have an impact on a future C++
> linear algebra semi-standard.

Yes, I have looked at it. I didn't mention it because of the
licensing. Kinda puts a crimp in our number-one code design
strategy: thievery. But I will study it some more in any case.


> Blitz++ is more a matrix package than a linear algebra package.
> They both provide a high-performance matrix system, but Blitz++
> focuses more on local operations within matrices and writing
> operations like you would on paper. The MTL has more of a global-
> operation focus (e.g. factorizations) and looks more `traditional.'

Yes. I think the key to progress here must be to achieve the
highest possible degree of orthogonality and genericity; otherwise
everybody just keeps building their own private fortresses over
and over. I like to pick on Blitz++ and TNT as examples of
container/algorithm orthogonality. Maybe MTL succeeds at achieving
these goals; it certainly looks good.


> They are both good, but not necessarily for the same uses.  The
> MTL works well when you want a LAPACK that works well with C++
> features.  Blitz++ works well when you want a fast matrix without
> the baggage, the type of thing you might want in an element-based,
> sparse, iterative solver.

I think ideally you want all of this in the design, or at least you
want to create no barriers to having all of this. We have to avoid the
traditional fortress problem for high-level abstractions.


> I don't believe TNT is still active,
> but I could be wrong.

Hmmm. I don't know either.


> [As a side note, the new version of a super-BLAS standard (BLAST)
> is being finalized.  The draft is at
> http://www.netlib.org/cgi-bin/checkout/blast/blast.pl .  To say
> some of the new chapters are controversial is a bit of an
> understatement.  Just something to keep an eye on.]

Good. I will look at this.


> Jason, in the process of writing a (hopefully free) C++ framework
> for sparse direct methods, yet another goal that's lead to failure
> too many times...

Good luck. And hurry up, so we can steal it.


-- 
G. Jungman