RE: [boost] Re: [uBlas] Any Performance Resulst Data with uBlas?

From: Hurd, Matthew (hurdm_at_[hidden])
Date: 2003-08-26 17:50:44

> From: David Abrahams [mailto:dave_at_[hidden]] 
> Subject: [boost] Re: [uBlas] Any Performance Resulst Data with uBlas?
> 
> 
> "Patrick Kowalzick" <Patrick.Kowalzick_at_[hidden]> writes:
> 
> > Hi Matthew,
> >
> > I like these statistics. Nice graphs.
> >
> > Did you test the block_prod compared to normal prod as well ? This 
> > would be interessting compared to ATLAS as well.

An early version of block prod is there in my pdf, it is the one in the
centre at the 1000 point, on the right.  The BTL/opencascade link shows
Atlas, but the current version of atlas performs quite a bit better.  

The uBLAS performance may be a little dated as the guys have done lots of
work.  Especially w.r.t. SSE/SSE2/vectorisation benefits.  Atlas wrapping
has been improved AFAIK, as ATLAS has since the results on the BTL page.
Also the Intel Math Kernel Library results I had were called from a uBLAS
matrix, calling DGEMM directly, so it still benefits, I feel, from nicety of
uBLAS.
 
> It's very hard to read, IMO.  It's not clear what the lines 
> are referring to and some of them look too similar to be 
> distinguished.

For sure.  That is nearly the point for some though.  GCC and VC benefit a
little from the assign syntax, not not a huge amount as the lines are close
together and are hard to distinguish.  It is interesting to note the great
differences you can get with Intel C++ depending on your optimization
settings, which I find a battle to get right.

But yeah, the graph aint pretty and it isn't too meaningful without more
context.

Matt.

> >> > Now, is there any performance result data about uBlas ?
> >> >
> >> This is one of the better sources of results...
> >>
> >> http://www.opencascade.org/upload/87/index.html
> >>
> >> There is a small chart I did a while back at 
> >> http://groups.yahoo.com/group/Boost-Users/files/
> >> Called uBLA performance.
> >>