John C Nash wrote:
This actually goes back a very long way. Peter is right to remind us
that "optimizers" (in the sense of compilers) can corrupt algorithms
that are well-designed. Optimizing in tests is something some of us have
fought for nearly 40 years, but compiler writers don't do much
floating-point computation -- they sometimes save a few microseconds of
computer time for many days and weeks of human time designing around
such silliness. It is worse than a "bug" in that the code to work around
the optimization can be very complicated and arcane -- and may not port
across architectures. Clearly a program should execute as the code
instructs, and not as a compiler designer decides to reinterpret it.
Nevertheless, with a fairly large offset of 16.0 or 256.0 I have never
seen such a test fail -- and it can port across different precision and
different radix arithmetics. However, I must say I tend to prefer to
turn off compile optimization and try to keep my code clean i.e,
manually optimized if possible. I also save the two sides of the test to
separate variables, though I can guess that some compilers would corrupt
that pretty easily.
Yes. I might not have been quite fair to the offset method, the qbeta
issue was slightly different. (The quantity
tx = xinbta - adj;
was assumed to become equal to xintbta in the original implementation,
and didn't.
I'm pretty sure the current declaration of xinbta as "volatile" is a
leftover from a time where the compiler had put xinbta in an 80 bit
register whereas tx was 64 bit. The 1998 version in SVN also had a
static volatile xtrunc used to force tx to 64 bit, but I seem to recall
that we put that there, so presumably there was an even earlier version.)
Note that historically this issue didn't upset early codes where
optimization was manual, then became an issue to watch out for on each
implementation when we computed locally, and now could be a nasty but
hidden trap with grid and cloud computing when algorithms may be running
on a number of different systems, and possibly controlling critical
systems.
JN
Peter Dalgaard wrote:
[EMAIL PROTECTED] wrote:
This comment is orthogonal to most of the others. It seems that folk
often
want to test for equality of "real" numbers. One important one is for
convergence tests. When writing my Compact Numerical Methods book I
had to
avoid lots of logical tests, but wanted to compare two REALs. I found
that
the following approach, possibly considered a trick, is to use an offset
and compare
xnew + offset
to
xold + offset
This works on the examples given to motivate the current thread with an
offset of 10, for example.
Motivation: Small xold, xnew compare offset with itself. Large xold and
xnew are compared bitwise. Essentially we change from using a
tolerance to
using 1/tolerance.
Perfect? No. But usable? Yes. And I believe worth keeping in mind for
those annoying occasions where one needs to do a comparison but wants to
get round the issue of knowing the machine precision etc.
Hmm. Echos of some early battles with R's qbeta() in this. I don't
think it can be recommended.
The problem is that you can end up in a situation where xnew=xold-1ulp
and xnewnew is xnew+1ulp. I.e. in two iterations you're back at xold.
Even in cases where this provably cannot happen, modern optimizers may
make it happen anyway...
--
O__ ---- Peter Dalgaard Øster Farimagsgade 5, Entr.B
c/ /'_ --- Dept. of Biostatistics PO Box 2099, 1014 Cph. K
(*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918
~~~~~~~~~~ - ([EMAIL PROTECTED]) FAX: (+45) 35327907
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
