NA %*% 0 == 0 (PR#4582) - R-devel

Duncan Murdoch · 2003-10-15T09:05:44Z

Peter Dalgaard BSA wrote: > Thomas Lumley writes: > > >>On Tue, 14 Oct 2003 hosking@watson.ibm.com wrote: >> >> >>>Full_Name: J . R. M. Hosking >>>Version: 1.8.0 >>>OS: Windows 2000 >>>Submission from: (NULL) (129.34.20.23) >>> >>> >>>On R 1.8.0 (and on R 1.5.1), Windows binary: >>> >>> >>>>NA %*% 0 >>> >>> [,1] >>>[1,] 0 >>> >>>This is surprising. Is it a bug? Note that >>> >> >>I think it's a bug. It doesn't seem to happen on Linux or OS X, though. R-1.8

Uwe Ligges

Wed, Oct 15, 2003 2:05 AM #

Peter Dalgaard BSA wrote:

R-1.8.0: Does not happen for me on WindowsNT4.0 either (self compiled 
binary, gcc-3.3.1).

I guess the problem is compiler-dependend, because I get that "0" in a 
self-compiled R-1.7.1 (has been compiled with gcc-3.2.?).

Is your binary self-compiled or from CRAN (Duncan Murdoch compiled the 
CRAN binary with gcc-3.3.1 as well, AFAIK)?

Uwe Ligges

Duncan Murdoch

Wed, Oct 15, 2003 6:06 AM #

On Wed, 15 Oct 2003 09:06:39 +0200 (MET DST), you wrote:

I get it in Win XP with both 1.8.0 (compiled with 3.3.1) and 1.7.1
(compiled with 3.2.x).  I'd guess one of two things:

 - it depends on the runtime library.  (Windows versions use MSVCRT,
but the exact version used depends on the system you're running on.)
This doesn't seem to explain what happens on Linux, does it? 

 - it depends on some uninitialized variable.  do_matprod is
surprisingly complicated, so I can't tell if this is what's happening.
For some reason Insight won't debug that function properly.

Duncan Murdoch

Brian Ripley

Wed, Oct 15, 2003 6:49 AM #

non-complex %*% is fundamentally done via a call to dgemm, so the answer is
going to depend on the BLAS library in use.  When Doug Bates introduced
that version of matprod, we did discuss a bit the potential problems of
cases like this.

I've cross-checked, and that _is_ what is going on with my Linux box: try
tracing around the call to dgemm in matprod in array.c.

Brian

On Wed, 15 Oct 2003, Duncan Murdoch wrote:

Brian D. Ripley,                  ripley@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 272860 (secr)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

Uwe Ligges

Wed, Oct 15, 2003 7:00 AM #

Prof Brian D Ripley wrote:

You are right!
Indeed, I used the ATLAS (Athlon) optimized Rblas.dll (that one you 
provide on CRAN) for R-1.8.0 and R's default Rblas.dll under R-1.7.1.
Using the other version of Rblas.dll changes the outcomes.

So I get "0" for default Rblas.dll, and NA for the Athlon optimzed one.

Uwe Ligges