Albrecht Gebhardt writes:
On Sat, 3 Jun 2000, Prof Brian D Ripley wrote:
On Fri, 2 Jun 2000, Albrecht Gebhardt wrote:
On Fri, 2 Jun 2000, Albrecht Gebhardt wrote:
On Fri, 2 Jun 2000, Albrecht Gebhardt wrote:
I just tried the rsync version of R-1.1.0 on one of my alphas:
It compiles without problems (gcc/g77 2.95.2, system is DU4.0E)
.........
I compiled --whithout-dxml, so I'm not using any special numeric library.
Sorry I mixed something up in my report: In this compile I used the alpha
specific DXML library, in my rpms I omit it with --without-dxml (because
it sometimes crashes). I will try once more now without dxml, may be this
changes something, don't know if it gets used at all in the svd code.
That was the reason. Now without DXML I pass the test without modifying
Eps:
Nevertheless, the tolerance was too low. Most machines (e.g. Linux,
Solaris and Windows) are giving errors of 8 or 9 times the
machine precision, and with dxml you got 10x (and this failed as
Eps was exactly 10x). eigen's help page has tolerances of 60x and 1000x,
and I've given svd's 100x.
[...]
So may be it is better do disable dxml support by default in ./configure?
Well, what does dxml support do for you? Making the errors 1x machine
precision larger than without is hardly critical. On the other hand, I am
not convinced that these optimized blas operations are worthwhile for all
but the most unusual uses of R.
I did not make any performance analysis for dxml, of course it should
be better than plain BLAS or LAPACK. But I have it disabled in my
setups for long now, because I ran frequently into troubles with
special LAPACK routines. I remember the case where I used DSYEVX
(selected symmetric eigenvalues/vectors) in one of my development
libraries and got reproducable crashes, if R was linked with dxml, it
was fine without. It even crashed when I downloaded the dsyevx source
from netlib and put it into my library, R still jumped into DSYEVX
from dxml and crashed.