[Rd] make check of R-alpha_2006-04-08_r37675 fails: qbeta

[Prof Brian Ripley]

On Sun, 9 Apr 2006, Peter Dalgaard wrote:

> bhs2 at mevik.net (Bjørn-Helge Mevik) writes:
>
>> make check of R-alpha_2006-04-08_r37675 fails on Debian GNU/Linux 3.1 running
>> on an Intel P4 computer.
>>
>>> version
>>                _
>> platform       i686-pc-linux-gnu
>> arch           i686
>> os             linux-gnu
>> system         i686, linux-gnu
>> status         alpha
>> major          2
>> minor          3.0
>> year           2006
>> month          04
>> day            08
>> svn rev        37675
>> language       R
>> version.string Version 2.3.0 alpha (2006-04-08 r37675)
>>
>> I'm using ATLAS with the SSE2 extensions, and get the same error with ATLAS
>> 3.2.1 (packages atlas2-* 3.2.1ln-15 in Debian) and 3.6.0 (packages atlas3-*
>> 3.6.0-19 in Debian).  I've also tried without external BLAS (--without-blas),
>> but the error still occurs.
>>
>> 43 (0) $ ./configure
>> [...]
>> R is now configured for i686-pc-linux-gnu
>>
>>   Source directory:          .
>>   Installation directory:    /usr/local
>>
>>   C compiler:                gcc  -g -O2 -std=gnu99
>>   Fortran 77 compiler:       g77  -g -O2
>>
>>   C++ compiler:              g++  -g -O2
>>   Fortran 90/95 compiler:    g77 -g -O2
>>
>>   Interfaces supported:      X11, tcltk
>>   External libraries:        readline, BLAS(ATLAS)
>>   Additional capabilities:   PNG, JPEG, iconv, MBCS, NLS
>>   Options enabled:           R profiling
>>
>>   Recommended packages:      yes
>>
>> 44 (0) $ make
>> [...]
>>
>> 45 (0) $ make check
>> [...]
>> running code in 'd-p-q-r-tests.R' ... OK
>> comparing 'd-p-q-r-tests.Rout' to './d-p-q-r-tests.Rout.save' ...777,778d776
>> < Warning message:
>> < full precision was not achieved in 'qbeta'
>> make[3]: *** [d-p-q-r-tests.Rout] Error 1
>> [...]
>>
>> 46 (0) $ diff tests/d-p-q-r-tests.Rout.{save,fail}
>> 3c3
>> < Version 2.3.0 alpha (2006-04-03 r37628)
>> ---
>>> Version 2.3.0 alpha (2006-04-08 r37675)
>> 791a792,793
>>> Warning message:
>>> full precision was not achieved in 'qbeta'
>> 1073c1075
>> < Time elapsed:  2.53 0.03 2.55 0 0
>> ---
>>> Time elapsed:  4.939 0.024 4.967 0 0
>>
>> The (hopefully) interesting part of tests/d-p-q-r-tests.Rout.fail:
>>
>>> ## Check q*(p* ( log ), log) = identity
>>> All.eq(Rbeta,   qbeta    (log(Pbeta), shape1 = .8, shape2 = 2, log=TRUE))
>> [1] TRUE
>> Warning message:
>> full precision was not achieved in 'qbeta'
>>
>> Please let me know if there is more I can provide to help trace this!
>
> I suspect that this is related to recent changes in the zero-finder,
> and that it is relatively benign. (If it is the effect that I'm
> thinking of, then the warning is spurious, and it is mainly a matter
> of whether we want to try to get rid of the warning or update the
> .Rout.save file.)

I am not seeing it on my platforms (FC3 x86_64 and i686 Linux with gcc 
3.4.5 and 4.1.0, i386 Windows and Solaris, with various compilers) and the 
zero-finder changes postdate r37675 (and qbeta does not use it). I think 
several other people are testing i686 Debian, including the daiuly CRAN 
package checks ....

I suspect it is a real (and long-standing) problem of loss of accuracy 
(2.2.1 did a check but did not report the results: it seems to be a 
convergence failure looking at qbeta.c).  What compiler version is this?

ATLAS is not involved.  You'll need to trace qbeta.c.

-- 
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