[Mlir-commits] [mlir] [MLIR] Add `mpi.comm` type to MPI dialect (PR #125361)

[Frank Schlimbach]

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?Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/125361 at github.com>


fschlimb wrote:

Thanks @mofeing for your exciting work. Sorry for the late reply.

I think adding a new type for the Communicator and an operation for MPI_COMM_WORLD/MPI_COMM_SELF is a good approach. A global doesn't look like the right abstraction to me.

With a dedicated Communicator type, I suggest closely following MPI and make the communicator mandatory in the IR. We might consider adding builders which do not accept a communicator (and use MPI_COMM_WORLD).

As @Moxinilian suggested, you need to add the communicator argument to all places where the modified operations are created. By default, the builders accepts arguments in the order as the inputs are defined in the .td file.

https://github.com/llvm/llvm-project/pull/125361