[Scilab-Dev] Oh scilab!what a beaut you are you are if only you'd build without the war

[bv]

On Friday 29 October 2010 10:50:48 Sylvestre Ledru wrote:
> Le vendredi 29 octobre 2010 à 10:47 +0100, bv a écrit :
> > On Friday 29 October 2010 05:58:23 Sylvestre Ledru wrote:
> > > Le vendredi 29 octobre 2010 à 03:28 +0100, bv a écrit :
> > > > Greetings,
> > > > 
> > > > 
> > > > I   am atempting  to compile scilab-5.2.2 on  a machine with these:-
> > > > ---cpu-amd64 2 cores
> > > > ---o/s CBLFS linux (64-bit only) compiled from sources kernel-2.6.32
> > > > ---numerical progs: atlas3.9.26,lapack-3.2.2 and umfpack~co
> > > > 
> > > > 
> > > > I have  umfpack and co in /usr/local/numerics
> > > > rt [ ~ ]# ls /usr/local/numerics/lib
> > > > libamd.a  libcamd.a     libcerbla.a   libcolamd.a   libcxsparse.a
> > > > libldl.a libumfpack.a
> > > > libbtf.a  libccolamd.a  libcholmod.a  libcsparse.a  libklu.a
> > > > libspqr.a rt [ ~ ]#
> > > 
> > > Usually, Scilab is expecting some dynamic libraries to build correctly.
> > > [...]
> > 
> > ....ehm here is the home of SuiteSparse
> > http://www.cise.ufl.edu/research/sparse/SuiteSparse/
> > 
> > and as far as I know  all its components  are built using the UFconfig
> > utility AND static libs are  the default.
> > If anyone on list knows  how to  to hack UFconfig or use say cmake to
> > generate shared libs  for the SuiteSparse components
> > please let me know.
> 
> Debian & Ubuntu packages are built using these patches which enables the
> shared lib build:
> http://git.debian.org/?p=pkg-scicomp/suitesparse.git;a=tree;f=debian/patche
> s;h=0ac060b2a39ad6586040ef42f2a9bb8b733a7980;hb=fcdd9cd529c01240b8e014d329c
> 0089afd2e8cf3
> 
> 
thannks  for the link.  
A cursury glanch at the patches suggest they may well be a tad on the old side 
and  are heavily skewed towards C  (and probably i386 )  ( I am unsure in any 
of them   are patucularly useful here. )

And from  my  modest  understanding of BLAS,  it seem the best BLAS   and 
other linear algebra progs are in fortran.  I also  find   fortran shared libs 
dont do too  well on  machines built on C-shared-libraries/ldconfig.  So is it 
not best to   stick to blas-fortran-libs/static-linking especially in these 
days when 4-16 gbytes of ram are commonplace?

Also here is a link on BLAS benchmarks
http://www.ats.ucla.edu/clusters/common/software/libraries/blas_benchmark.htm

 Taken into account all the expert advice and wanting the best performance, I 
have designed my system (regarding application of BLAS) to:-
--use pure 64-bit  libraries
--use gfortran for blas
--use acml  where this is possible

could I put thse on the wish llist for  future version of your excellent 
program scilab.?