Pybind11: Possible to use mpi4py?

Question

Is it possible in Pybind11 to use mpi4py on the Python side and then to hand over the communicator to the C++ side?

If so, how would it work?

If not, is it possible for example with Boost? And if so, how would it be done?

I searched the web literally for hours but didn't find anything.

Answer 1

Passing an mpi4py communicator to C++ using pybind11 can be done using the mpi4py C-API. The corresponding header files can be located using the following Python code:

import mpi4py
print(mpi4py.get_include())

To convenietly pass communicators between Python and C++, a custom pybind11 type caster can be implemented. For this purpose, we start with the typical preamble.

// native.cpp
#include <pybind11/pybind11.h>
#include <mpi.h>
#include <mpi4py/mpi4py.h>

namespace py = pybind11;

In order for pybind11 to automatically convert a Python type to a C++ type, we need a distinct type that the C++ compiler can recognise. Unfortunately, the MPI standard does not specify the type for MPI_comm. Worse, in common MPI implementations MPI_comm can be defined as int or void* which the C++ compiler cannot distinguish from regular use of these types. To create a distinct type, we define a wrapper class for MPI_Comm which implicitly converts to and from MPI_Comm.

struct mpi4py_comm {
  mpi4py_comm() = default;
  mpi4py_comm(MPI_Comm value) : value(value) {}
  operator MPI_Comm () { return value; }

  MPI_Comm value;
};

The type caster is then implemented as follows:

namespace pybind11 { namespace detail {
  template <> struct type_caster<mpi4py_comm> {
    public:
      PYBIND11_TYPE_CASTER(mpi4py_comm, _("mpi4py_comm"));

      // Python -> C++
      bool load(handle src, bool) {
        PyObject *py_src = src.ptr();

        // Check that we have been passed an mpi4py communicator
        if (PyObject_TypeCheck(py_src, &PyMPIComm_Type)) {
          // Convert to regular MPI communicator
          value.value = *PyMPIComm_Get(py_src);
        } else {
          return false;
        }

        return !PyErr_Occurred();
      }

      // C++ -> Python
      static handle cast(mpi4py_comm src,
                         return_value_policy /* policy */,
                         handle /* parent */)
      {
        // Create an mpi4py handle
        return PyMPIComm_New(src.value);
      }
  };
}} // namespace pybind11::detail

Below is the code of an example module which uses the type caster. Note, that we use mpi4py_comm instead of MPI_Comm in the function definitions exposed to pybind11. However, due to the implicit conversion, we can use these variables as regular MPI_Comm variables. Especially, they can be passed to any function expecting an argument of type MPI_Comm.

// recieve a communicator and check if it equals MPI_COMM_WORLD
void print_comm(mpi4py_comm comm)
{
  if (comm == MPI_COMM_WORLD) {
    std::cout << "Received the world." << std::endl;
  } else {
    std::cout << "Received something else." << std::endl;
  }
}

mpi4py_comm get_comm()
{
  return MPI_COMM_WORLD; // Just return MPI_COMM_WORLD for demonstration
}

PYBIND11_MODULE(native, m)
{
  // import the mpi4py API
  if (import_mpi4py() < 0) {
    throw std::runtime_error("Could not load mpi4py API.");
  }

  // register the test functions
  m.def("print_comm", &print_comm, "Do something with the mpi4py communicator.");
  m.def("get_comm", &get_comm, "Return some communicator.");
}

The module can be compiled, e.g., using

mpicxx -O3 -Wall -shared -std=c++14 -fPIC \
  $(python3 -m pybind11 --includes) \
  -I$(python3 -c 'import mpi4py; print(mpi4py.get_include())') \
  native.cpp -o native$(python3-config --extension-suffix)

and tested using

import native
from mpi4py import MPI
import math

native.print_comm(MPI.COMM_WORLD)

# Create a cart communicator for testing
# (MPI_COMM_WORLD.size has to be a square number)
d = math.sqrt(MPI.COMM_WORLD.size)
cart_comm = MPI.COMM_WORLD.Create_cart([d,d], [1,1], False)
native.print_comm(cart_comm)

print(f'native.get_comm() == MPI.COMM_WORLD '
      f'-> {native.get_comm() == MPI.COMM_WORLD}')

The output should be:

Received the world.
Received something else.
native.get_comm() == MPI.COMM_WORLD -> True

Answer 2

This is indeed possible. As was pointed out in the comments by John Zwinck, MPI_COMM_WORLD will automatically point to the correct communicator, so nothing has to be passed from python to the C++ side.

Example

First we have a simple pybind11 module that does expose a single function which simple prints some MPI information (taken from one of the many online tutorials). To compile the module see here pybind11 cmake example.

#include <pybind11/pybind11.h>
#include <mpi.h>
#include <stdio.h>

void say_hi()
{
    int world_size;
    MPI_Comm_size(MPI_COMM_WORLD, &world_size);
    int world_rank;
    MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
    char processor_name[MPI_MAX_PROCESSOR_NAME];
    int name_len;
    MPI_Get_processor_name(processor_name, &name_len);
    printf("Hello world from processor %s, rank %d out of %d processors\n",
        processor_name,
        world_rank,
        world_size);
}

PYBIND11_MODULE(mpi_lib, pybind_module)
{
    constexpr auto MODULE_DESCRIPTION = "Just testing out mpi with python.";
    pybind_module.doc() = MODULE_DESCRIPTION;

    pybind_module.def("say_hi", &say_hi, "Each process is allowed to say hi");
}

Next the python side. Here I reuse the example from this post: Hiding MPI in Python and simply put in the pybind11 library. So first the python script that will call the MPI python script:

import sys
import numpy as np

from mpi4py import MPI

def parallel_fun():
    comm = MPI.COMM_SELF.Spawn(
        sys.executable,
        args = ['child.py'],
        maxprocs=4)

    N = np.array(0, dtype='i')

    comm.Reduce(None, [N, MPI.INT], op=MPI.SUM, root=MPI.ROOT)

    print(f'We got the magic number {N}')

And the child process file. Here we simply call the library function and it just works.

from mpi4py import MPI
import numpy as np

from mpi_lib import say_hi


comm = MPI.Comm.Get_parent()

N = np.array(comm.Get_rank(), dtype='i')

say_hi()

comm.Reduce([N, MPI.INT], None, op=MPI.SUM, root=0)

The end result is:

from prog import parallel_fun
parallel_fun()
# Hello world from processor arch_zero, rank 1 out of 4 processors
# Hello world from processor arch_zero, rank 2 out of 4 processors
# Hello world from processor arch_zero, rank 0 out of 4 processors
# Hello world from processor arch_zero, rank 3 out of 4 processors
# We got the magic number 6