Re: [Boost-users] [multi-array] Performance? Difference from GIL or uBLAS?

From: joel falcou (joel.falcou_at_[hidden])
Date: 2008-08-07 08:23:00

> Also, after a rough glance through multi_array and uBLAS it looks
like
> GIL, uBLAS, and multi_array seem to repeat very similar
concepts (and
> seem to use different syntax). Isn't that kind of
bad?  I have been playing
> with GIL in the past but now I am
scratching my head about these other
> libraries.

Beware that running ''some tests'' on arbitrary large array of any kind
must take into account the fact that you may induces cache misses
especially in the above example where I don't think a quarter of the data
fit into the L1 cache. Considering cache misses may lead to a ten fold
computing time, this may render performance tests unusable. 

I
have struggled a few with optimized array and tried many way to do so.
Seems the easiest way to have :
- chained operator[ ] access
syntax
- correct performance in L1 cache
- easy way to perform
tiled computation
- comaptible with other high-performances settings
(padding & alignement for SIMD for example)

is not
easy.  Best way so far is to use a compact Numeric Recipes like
structure aka for a D1*D2*..*Dn array, you have to allocate a D1*D2*..*Dn
memory block and n-1 array of pointer of different level to each
dimension. Then returning the top level pointer (a T**.** with n *) grants
you a way to acces element i1,i2,..,in through n[] call while having no
stride computation to do manually and so, making tiling easy to write as a
multiple array loop nest.
And most of the time, a container that
can't support tiling easily is useless in most real life application as
this simple techniques can significantly increase runtime performances.

I actually wrote a thesis on the subject and still struggle to get
a decent library out of all my experiments.There is some prototype of it
but it's nowhere completed (NT2 on sourceforge for reference) in which I
use Epression tempalte to evaluates at compile time a somewhat optimal
tiling size. If needed I can reinject my experiments results here so we
can all brainstorm about this but I think the problem of "easy-to-use
and good performances" numeric data container is not easy and not
currently addressed anywhere.