[R-SIG-Mac]Re: [Rd] timing Mac R's

[Prof Brian D Ripley]

On Fri, 8 Feb 2002, Jan de Leeuw wrote:

> Here is the same thing as before, now done a bit more systematically. The
> rather bad performance of R-patched on eigen is a mystery (to me), the
> gain
> in La.eigen when using ATLAS is probably not Altivec, but using
> threaded level 3 BLAS on a dual processor machine.

Well, it's almost entirely in the algorithm used in LAPACK.  There is no
single precision used.  DSYEVR was designed as `the holy grail' for that
problem, and it uses level 3 BLAS much better than its predecessors.
I see an 8x speed up on my PIII Linux box on your tests (and the ATLAS
there is only using one processor to keep the build portable).

My guess (and only a guess) is that you have an alignment problem. I am
pretty sure there is no change in the internal code across the versions you
used, but some compilers and chips will run code much faster with certain
alignments than others. Sparc for example needs -dalign under it's native
compiler, and without that (which used to be necessary under S-PLUS) there
could be performance changes of this nature under very small code changes.

-- 
Brian D. Ripley,                  ripley@stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272860 (secr)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595