[Mlir-commits] [mlir] [mlir][mpi] Mandatory Communicator (PR #133280)

[Jeff Hammond]

Sergio =?utf-8?q?Sánchez_Ramírez?=,
Sergio =?utf-8?q?Sánchez_Ramírez?=,
Sergio =?utf-8?q?Sánchez_Ramírez?=,
Sergio =?utf-8?q?Sánchez_Ramírez?=,
Sergio =?utf-8?q?Sánchez_Ramírez?=,"Schlimbach, Frank"
 <frank.schlimbach at intel.com>,"Schlimbach, Frank" <frank.schlimbach at intel.com>
 =?utf-8?q?,?=Frank Schlimbach <frank.schlimbach at intel.com>,Frank
 Schlimbach <frank.schlimbach at intel.com>,"Schlimbach, Frank"
 <frank.schlimbach at intel.com>,"Schlimbach, Frank" <frank.schlimbach at intel.com>
 =?utf-8?q?,?="Schlimbach, Frank" <frank.schlimbach at intel.com>
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/133280 at github.com>


jeffhammond wrote:

> The main problem is that as of right now, MPI doesn't have a stable ABI. MPI 5 specification will fix this but it's yet unfinished, even less are there any implementations (mpitrampoline being the exception .

It is finished pending the final confirmation vote in June. All of the C features are fixed. Only Fortran related things are pending.

What aspect of the ABI are you dependent on? I'll post the exact details for that from the release candidate spec.

MPICH has an implementation already. The MPI Forum publishes a stubs repo that implements the ABI, which is regularly tested by mpi4py. I will update Mukautuva to support the ABI soon.

Neither wi4mpi nor MPItrampoline support the MPI-5 ABI, although both can if they want. 

https://github.com/llvm/llvm-project/pull/133280