[R] error returned by "make check" in R-2.5.0

[Prof Brian Ripley]

On Sat, 28 Apr 2007, Eric Thompson wrote:

> Professor Ripley,
>
> Thank you for your comments.
>
> On 4/28/07, Prof Brian Ripley <ripley at stats.ox.ac.uk> wrote:
>> How big is abs(lam - lam2[i])/Meps ?
>
> Here is the result on my system:
>
>> abs(lam - lam2[i])/Meps
> [1] 60 18 17  0 12
>
> which is quite surprising to me that the maximum value exactly equals
> the cutoff, so perhaps this is not such a big problem. I changed the
> value to 61 in the reg-tests-1.R file, and then make check runs with
> no errors.
>
>> This could be one of those cases where your system (which CPU, what
>> compilers?) is unusually inaccurate, but Linux systems do not usually
>> differ much in their accuracies on the same CPU.
>
> It is somewhat disconcerting that my system is unusually inaccurate.
> I'm running Intel Xeon 3.20GHz CPUs and I've been using gcc 4.1.1 and
> ifort to compile R (these are what the compile script choose by
> default on my system). Perhaps it is a bad idea to mix the GNU and
> Intel compilers?

Ah, ifort is less accurate than gfortran quite often, because it cares 
more about speed.  It all depends on the version: on my Opteron x86_64 
system and icc/ifort 9.1 the test passes, whereas on a i686 (older Xeons) 
with icc/ifort 9.0 quite a few tests fail unless optimization is turned 
down.  (I guess those Xeons could be either architecture.)

We will raise the limit a bit, as this seems a little too strigent.

[...]

-- 
Brian D. Ripley,                  ripley at 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 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595