Method allreduce

src/comm.cpp:559–585

template <typename TensorDataType>
void lbann_comm::allreduce(El::AbstractMatrix<TensorDataType>& m,
                           const El::mpi::Comm& c,
                           El::mpi::Op op) const
{
  if (El::mpi::Size(c) == 1 || m.Height() < 1 || m.Width() < 1) {
    return;
  }

  const int local_size = m.Height() * m.Width();
  m_bytes_sent += sizeof(DataType) * local_size;
  m_bytes_received += sizeof(DataType) * local_size * (El::mpi::Size(c) - 1);

  switch (m.GetDevice()) {
  case El::Device::CPU:
    return allreduce_impl(
      static_cast<El::Matrix<TensorDataType, El::Device::CPU>&>(m),
      c,
      op);
#ifdef LBANN_HAS_GPU
  case El::Device::GPU:
    return allreduce_impl(
      static_cast<El::Matrix<TensorDataType, El::Device::GPU>&>(m),
      c,
      op);
#endif // LBANN_HAS_GPU
  }
}

template <typename TensorDataType>
void lbann_comm::allreduce(El::AbstractDistMatrix<TensorDataType>& m,