pivoting, Linpack performance, tunable SPEC results

[David G. Hough on validgh]

I'm getting myopic, I guess; on my assertion that pivoting doesn't hurt
performance much, Wendy Thrash writes:

> I'd have to dispute this for compilers that do blocking, since it seems
> to be exactly the pivoting that prevents blocking of certain routines.

> Porterfield, for example, in his dissertation, classifies LINPACKD
> as "non-transformable" because of its pivoting.  To put this in perspective,
> Wolf and Lam, in their upcoming SIGPLAN PLDI paper, report about 20-25%
> speedup via blocking of a 500 x 500 LU factorization _without pivoting_
> on a single processor of an SGI 4D/380; this blocking also more than doubles
> the four-processor speed and quadruples the eight-processor speed over the
> unblocked versions of the algorithm for four and eight processors.

I guess the moral is that on the 1000x1000 Linpack benchmark (where rewriting
the source code is expected) it would be OK
to rewrite it in quad and throw away the pivoting.  Or maybe just throw
away the pivoting.