[petsc-dev] Single configure script for SuiteSparse package

Satish Balay balay at mcs.anl.gov
Wed Feb 12 12:40:38 CST 2014


On Wed, 12 Feb 2014, Satish Balay wrote:

> On Wed, 12 Feb 2014, Jed Brown wrote:
> 
> > Satish Balay <balay at mcs.anl.gov> writes:
> > 
> > > On Wed, 12 Feb 2014, Abhyankar, Shrirang G. wrote:
> > >
> > >> Barry,
> > >>    Satish and I have attempted to create a single configure script for SuiteSparse installation. It is currently in the branch shri/update-SuiteSparse https://bitbucket.org/petsc/petsc/branch/shri%2Fupdate-SuiteSparse. This change downloads the entire SuiteSparse package and installs all the sub-packages that PETSc currently supports (UMFPACK, CHOLMOD, and KLU). We can of course have selective installation of sub-packages if need be, currently it does not. I know you prefer a 'separate' configure script for each sub-package, but we think a single configure is a better approach. This change also updates PETSc to use the latest SuiteSparse. I've tested it with both debugging and optimized modes (with complex and 64-bit ints) and there was no change needed for UMFPACK or CHOLMOD solver interfaces. Let us know what you think? petsc-devs?
> > >> 
> > >
> > > Note: Apart from--download-suitsparse=1 - --with-suitsparse=1 or
> > > --with-suitsparse works on ubuntu/linux with preisntalled suitsparse
> > > [and presumably will work with other suitsparse installs]
> > >
> > > And it avoids all the issues with spliting and maintaining our own
> > > taballs for UMFPACK, CHOLMOD, and KLU [and managing duplicate copies
> > > of dependencies like AMD, COLAMD libsuitsparse etc.] - and will avoid
> > > corner-cases like --download-umfpack --with-cholmod-dir= etc..
> > 
> > Those packages were never packaged for sane independent-install anyway.
> > 
> > However, some distributions have done the work of allowing separate
> > install of each component.  Could we have the one suitesparse configure
> > script detect which components are available?
> 
> Perhaps its possible. But I suspect it will get tricky wrt detecting
> the 'liblist' for such partial installs.
> 
> Would it not be easier for user to install all 3 packages [if they
> want to use any one pre-installed packages?] or just use
> --download-suitsparse?
> 
> I think majority of usecases will be covered by --with-suitsparse=1 or
> --download-suitsparse options.

I see fedora, macports, homebrew all have a single suitsparse package.

BTW: There is an optional dependency on metis from
suitesparse. Current code [for --download-suitsparse] doesn't use
this..

Satish

> 
> Satish
> 
> > 
> > > One caveat: we build only part of suitsparse [required by petsc] - not
> > > the whole thing.
> > >
> > > BTW: should we change PETSC_HAVE_UMFPACK PETSC_HAVE_CHOLMOD etc to
> > > PETSC_HAVE_SUITSPARSE?  [I don't think its necessary - but checking..]
> > >
> > > And then configure should print an error message if --with-umpfack
> > > option is specified [via check_for_option_changed() in
> > > config/configure.py]
> > 
> 
> 




More information about the petsc-dev mailing list