Questions on Intel MPI are more likely to find an answer on the HPC forum. I don't find any indication of support for MPI_LONG_DOUBLE. For one thing, Intel C++ can't support long double consistently across platforms, given the totally different treatment by the run-time libraries. Of course, if you are compiling in the default mode where long double is treated as double, then -DMPI_LONG_DOUBLE=MPI_DOUBLE should run. Intel MPI doesn't exhibit the opportunistic order of MPI_Allreduce operations resulting in inconsistent results of certain other MPI implementations, so it's possible you may be able to take advantage of the much better performance of MPI_DOUBLE.
