[Rd] Problems with Lapack's eigen() on 64-bit

[Peter Dalgaard]

Martin Maechler <maechler at stat.math.ethz.ch> writes:

> I'm only now realizing that we have severe problems with R on our
> AMD 'Opteron' and 'Athlon64' clients running Redhat Enterprise
> with all 64-bit libraries (AFAICS).
> 
> The Lapack problem happens for R-patched and R-devel both on
> the Opteron and the Athlon64.
> 
> Here are platform details:
> 
> o  "gcc -v" and "g77 -v" both end with the line
>    gcc version 3.2.3 20030502 (Red Hat Linux 3.2.3-34)
> 
> o  I've used ...../configure --without-blas
> 
> 1) Opteron ("deb7", a dual-processor compute server):
> 
>   - uname -a :
>     Linux deb7 2.4.21-9.0.3.ELsmp #1 SMP Tue Apr 20 19:44:29 EDT 2004 x86_64 x86_64 x86_64 GNU/Linux
>   - /proc/cpuinfo  contains (among much more)
>     vendor_id       : AuthenticAMD
>     cpu family      : 15
>     model           : 5
>     model name      : AMD Opteron(tm) Processor 248
> 
> 2) Athlon64 (a simple "new generation" client - to become my desktop soon):
>   
>   - uname -a :
>     Linux setup-12 2.4.21-15.0.2.EL #1 Wed Jun 16 22:41:44 EDT
>     2004 x86_64 x86_64 x86_64 GNU/Linux 
> 
>   - /proc/cpuinfo  contains
>     vendor_id       : AuthenticAMD
>     cpu family      : 15
>     model           : 14
>     model name      : AMD Athlon(tm) 64 Processor 2800+
> 
> ----------------
> 
> Now the Lapack problem, easily seen from the base eigen()
> example:
> 
>     > eigen(cbind(1, 3:1, 1:3))
>     $values
>     [1]  5.7015621  1.0000000 -0.7015621
> 
>     $vectors
> 	      [,1]       [,2]       [,3]
>     [1,] 0.4877939 -0.7181217 -0.9576161
>     [2,] 0.6172751 -0.3893848  0.2036804
>     [3,] 0.6172751  0.5767849  0.2036804
> 
> which is simply plainly wrong and  
>     eigen(cbind(1, 3:1, 1:3), EISPACK=TRUE)
> gives the correct eigen values c(5, 1, 0)  and corresponding
> eigenvectors.
> 
> IIRC, we've already dealt with a Lapack problem, and that
> workaround (built into R-devel's Makefiles) has been to use
> -ffloat-store for the compilation of src/modules/lapack/dlamc.f
> 
> ----------
> 
> Thank you for hints / suggestions.
> 
> Others with 64-bit platforms might also try to see what
>    eigen(cbind(1, 3:1, 1:3)) 
> gives there.

SuSE Linux doesn't seem to have problems, neither with a default build
of 1.9.1 (as per Detlef's SPEC file), my ATLAS or GOTO builds of 1.9.0
or a 2.0.0-in-progress from May 5 (sorry: not SVN-ified yet, I *am* on
vacation you know...). All give

> eigen(cbind(1, 3:1, 1:3))
$values
[1]  5.000000e+00  1.000000e+00 -4.440892e-16

$vectors
          [,1]       [,2]       [,3]
[1,] 0.5773503 -0.8451543 -0.9428090
[2,] 0.5773503 -0.1690309  0.2357023
[3,] 0.5773503  0.5070926  0.2357023

The largest and smallest eigenvectors/values are easily
seen to be correct. And the middle one is basically this:

pd at linux:~/r-devel/BUILD> echo  'cbind(1, 3:1, 1:3)%*%c(-5,-1,3)' | bin/R -q --vanilla
> cbind(1, 3:1, 1:3)%*%c(-5,-1,3)
     [,1]
[1,]   -5
[2,]   -1
[3,]    3


The g77 version info says "3.3.1 (SuSE Linux)"

-- 
   O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)             FAX: (+45) 35327907