[R-SIG-Mac] gfortran bug?

[Andreas Noack Jensen]

Thank you so much for looking at it. Initially my code was in the "free"
F95 format but partly because of the error and partly because of portability
concerns I chose to translate it to strict F77 even thought it is ugly and
annoying with the line width limit.

The subroutine is thought as a multivariate version of filter(..., method =
'recursive') and I use it to simulate VAR processes. dgemm is the double
precision BLAS subroutine for

C = alpha*A*B + beta*C

for general matrices A, B, C and scalars alpha and beta.

Indeed I think something is going wrong with the initial values but I cannot
figure out why. Maybe because of missing Fortran skills but I am puzzled
because the code works fine on Windows and Linux.

I think the "-mtune=core2" problem has disappeared in the latest 4.2.3
gfortran at r.research because I had the error in the beginning but after
reinstall of gfortran it was gone (right now, however,  I am on gfortran
4.2.4 aka gcc 4.2.1).

Thank you once again and hope some new ideas to isolate the problem will
appear.

Best
Andreas

Den 27/04/10 13.34 skrev "Berend Hasselman" <bhh at xs4all.nl> følgende:

> 
> On 25-04-2010, at 22:22, Andreas Noack Jensen wrote:
> 
> ...
>> On Mac I quite randomly get very weird results from the subroutine. It is
>> like the initial values explodes to fx 2e+290 even for identical calls
>> without random number generation. Most results are identical to results I
>> get from a pure R code but at some calls they explode in a non systematic
>> fashion.
> ...
> 
> I have experimented with your stuff.
> Find attached the R files needed to run my test.
> Below is what I have done together with the R output.
> As you can see the third column of the result matrix is weird.
> This result is independent from setting the seed.
> Some sort of initialisation issue?
> Since I haven't got a clue what rmfilter does and my memory needs some
> refreshing wrt dgemm,
> I can't pinpoint the problem.
> 
> Berend
> 
> 
> I am (still)  using the gfortran from gfortran-4.2.3.dmg
> ------------------------------------------------------------------
> I made the rmfilter.so with
> 
> MAKEFLAGS="FCFLAGS=-O2" R CMD SHLIB rmfilter.f90
> 
> NB. had to do it since this gfortran doesn't accept the -mtune=core2 option!!
> ----------------------------------------------------------------------
> sessionInfo()
> 
> R version 2.10.1 Patched (2010-04-07 r51689)
> x86_64-apple-darwin9.8.0
> 
> locale:
> [1] en_GB/en_GB/C/C/en_GB/en_GB
> 
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods   base
> 
> ------------------------------------------------------------------
> Mac OS X 10.6.3 Intel
> 
> ------------------------------------------------------------------
> 
> I ran this test job
> 
> 
> dyn.load('rmfilter.so')
> source("rmfilter.R")
> 
> set.seed(407)
> 
> eps <- matrix(rnorm(30), ncol = 3)
> eps
> 
> z <-  rmfilter(eps, array(diag(3)*0.5, c(3,3,2)))
> z
> 
> 
> with this as output
> 
>> 
>> dyn.load('rmfilter.so')
>> source("rmfilter.R")
>> 
>> set.seed(407)
>> 
>> eps <- matrix(rnorm(30), ncol = 3)
>> eps
>               [,1]       [,2]       [,3]
>  [1,]  0.841081288  1.7100891  0.9161884
>  [2,]  1.072881139  1.1351244 -0.4314795
>  [3,] -0.543593265 -1.0828613 -0.6333517
>  [4,]  0.441951763 -1.0539372 -1.8468529
>  [5,]  0.004920633 -0.9635668  0.6060301
>  [6,] -0.731256023  0.4951620 -1.9168646
>  [7,] -1.216400546 -1.5097126  1.1407077
>  [8,] -0.210674189 -1.1110270 -0.3563495
>  [9,] -0.378021236 -1.5624664 -2.5094364
> [10,] -0.939183015 -0.2886860  2.1596868
>> 
>> z <-  rmfilter(eps, array(diag(3)*0.5, c(3,3,2)))
>> z
>             [,1]       [,2]          [,3]
>  [1,]  0.8410813  1.7100891 1.018209e+277
>  [2,]  1.4934218  1.9901689 1.018209e+277
>  [3,]  0.6236583  0.7672677 1.018209e+277
>  [4,]  1.5004918  0.3247811 1.018209e+277
>  [5,]  1.0669957 -0.4175424 1.018209e+277
>  [6,]  0.5524877  0.4487814 1.018209e+277
>  [7,] -0.4066589 -1.4940931 1.018209e+277
>  [8,] -0.1377598 -1.6336829 1.018209e+277
>  [9,] -0.6502306 -3.1263543 1.018209e+277
> [10,] -1.3331782 -2.6687046 1.018209e+277
> attr(,"tsp")
> [1]  1 10  1
> 
> 
> 
>