[Boost-users] [mpi] could all_reduce handle MPI_IN_PLACE ?

Subject: [Boost-users] [mpi] could all_reduce handle MPI_IN_PLACE ?
From: Alain Miniussi (Alain.Miniussi_at_[hidden])
Date: 2009-06-05 18:58:32

Hi,

There doesn't seems to be an elgant way to do an in place all_reduce
on array right now (Although I only stumbled on the issue after
I moved to openmpi 1.3.2).
That is, if all_reduce_impl is called with (in_values == out_values)
MPI is allowed to compain with a MPI_ERR_BUF. Of course, one
can call all_reduce with static_cast<T*>(MPI_IN_PLACE) but I do not
find that solution very elegant.

Would it be a problem to modify all_reduce_impl into:

  template<typename T, typename Op>
   void
   all_reduce_impl(const communicator& comm, const T* in_values, int n,
                   T* out_values, Op /*op*/, mpl::true_ /*is_mpi_op*/,
                   mpl::true_ /*is_mpi_datatype*/)
   {
     if (in_values == out_values) {
        in_values = static_cast<T*>(MPI_IN_PLACE);
     }
     BOOST_MPI_CHECK_RESULT(MPI_Allreduce,
                            (const_cast<T*>(in_values), out_values, n,
                             boost::mpi::get_mpi_datatype<T>(*in_values),
                             (is_mpi_op<Op, T>::op()), comm));
   }

It worked for me, and if there is an agreement I can propose a patch.

Regards,

-- 
Alain Miniussi                                       
http://www.oca.eu/alainm
Equipe CRIMSON                                 https://crimson.oca.eu
[phone]