dolfin team mailing list archive

["Matthew Knepley" <knepley@xxxxxxxxx>]

Reason n why packing systems are broken. It would be easy to rebuild the
source package to point at Atlas (just switch the configure arg), but you
would have to update the dependencies.

   Matt

On Thu, Feb 28, 2008 at 6:19 AM, Ilmar Wilbers <ilmarw@xxxxxxxxx> wrote:
> Adding PETSC_TS_LIB did help, but not for the reasons one would think.
>  The result is the same whether reblas3-dev or atlas3-base-dev is
>  installed (hence the file dolfin.pc remains the same when running
>  configure with the two different blas libraries), but the library
>  '-lblas-3' is now added to the list of libraries. It is located in
>  /usr/lib, and set to point to the correct libblas.so (either in
>  /usr/lib/ or /usr/lib/atlas). I don't know why this is so, but it might
>  very well be set at the compiling of the debian petsc-dev package, so I
>  am not 100% sure that this is the way to proceed.
>
>  A third alternative, beside adding '-L/usr/lib/atlas' to dolfin.pc, is
>  to omit '-lblas' to be added when dolfin is compiled with petsc. As far
>  as I understand, libpetsc.so will always be linked with libblas.so anyway?
>
>  ilmar
>
>
>
>  Anders Logg wrote:
>  > On Thu, Feb 28, 2008 at 12:19:00AM +0100, Anders Logg wrote:
>  >
>  >> On Wed, Feb 27, 2008 at 11:49:46AM +0100, Ilmar Wilbers wrote:
>  >>
>  >>> Hello to all!
>  >>>
>  >>> With Ubuntu Hardy, updates of openmpi and petsc (2.3.3) are available.
>  >>> petsc will also be built against openmpi, not mpich. We are currently
>  >>> looking at building dolfin using these packages as dependencies. There
>  >>> is, however, a problem with ATLAS.
>  >>>
>  >>> When installing petsc-dev, there is a dependency on either atlas3-base,
>  >>> OR both lapack3 and refblas3. This means that if no BLAS is installed,
>  >>> atlas3-base is set to be installed as the dependency. atlas3-base is the
>  >>> faster, optimized one, and the one we really would like to use. The
>  >>> problem is that libblas.so and liblapack.so are put in the directory
>  >>> /usr/lib/atlas when using atlas3-base. If the package refblas3 is used,
>  >>> the library files are put in /usr/lib. When doing a 'make demo' in
>  >>> dolfin, when compiled with the petsc library provided by the debian
>  >>> package, this works fine when only refblas3 and lapack3 are installed,
>  >>> as /usr/lib is a default directory when the compiler (mpicxx) is looking
>  >>> for libraries. When atlas3-base also is available, libpetsc.so tries to
>  >>> use the libblas.so located in /usr/lib/atlas, and so, when running 'make
>  >>> demo', we get an error about undefined references to atlas symbols,
>  >>> unless we explicitly add '-L/usr/lib/atlas' to the linking stage of
>  >>> dolfin.pc in lib/pkgconfig.
>  >>>
>  >>> When compiling dolfin without petsc, we do not get this error, as
>  >>> /usr/lib/libblas.so is used. However, we really would like to use the
>  >>> atlas one when available, as it runs at least twice as fast (although
>  >>> building them yourself will be ten times faster, it seems).
>  >>>
>  >>> My suggestion is to make sure that the line '-L/usr/lib/atlas' is
>  >>> included in the dolfin.pc file.
>  >>>
>  >>> ilmar
>  >>>
>  >
>  > (I'm resending this because of problems with the mail-server, hope it
>  > doesn't appear twice.)
>  >
>  > The cflags/libs needed for PETSc are extracted by the DOLFIN configure
>  > script using the file petsc.conf (which is located at the top of the
>  > DOLFIN source tree). This is a makefile that includes the file
>  > ${PETSC_DIR}/bmake/common/variables which gives us all we need for
>  > PETSc.
>  >
>  > Does it help if you just add the variable
>  >
>  >      ${PETSC_TS_LIB}
>  >
>  > suggested by Matt at the end of the last line in petsc.conf?
>  >
>  > (And then rerun the configure script to create a new dolfin.pc.)
>  >
>  >
>
>
> _______________________________________________
>  DOLFIN-dev mailing list
>  DOLFIN-dev@xxxxxxxxxx
>  http://www.fenics.org/mailman/listinfo/dolfin-dev
>



-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which
their experiments lead.
-- Norbert Wiener