如何通过cython将MPI通信器从python传递到C? [英] How to pass an MPI communicator from python to C via cython?

问题描述

我正在尝试通过cython封装一个以MPI_Comm通信器句柄为参数的C函数.结果，我希望能够从python调用该函数，并将其传递给mpi4py.MPI.Comm对象.我想知道的是，如何进行从mpi4py.MPI.Comm到MPI_Comm的转换.

I am trying to wrap a C function taking an MPI_Comm communicator handle as a parameter via cython. As a result, I want to be able to call the function from python, passing it an mpi4py.MPI.Comm object. What I am wondering is, how to make the conversion from mpi4py.MPI.Comm to MPI_Comm.

为了演示，我使用一个简单的"Hello World！"类型的函数:

To demonstrate, I use a simple "Hello World!"-type function:

helloworld.h:

#ifndef HELLOWORLD
#define HELLOWORLD
#include <mpi.h>

void sayhello(MPI_Comm comm);

#endif

helloworld.c:

#include <stdio.h>
#include "helloworld.h"

void sayhello(MPI_Comm comm){
    int size, rank;
    MPI_Comm_size(comm, &size);
    MPI_Comm_rank(comm, &rank);
    printf("Hello, World! "
           "I am process %d of %d.\n",
           rank, size);
}

我现在想像这样从python调用此函数:

I now want to call this function from python like this:

from_python.py:

import mpi4py
import helloworld_wrap

helloworld_wrap.py_sayhello(mpi4py.MPI.COMM_WORLD)

表示mpirun -np 4 python2 from_python.py应该给出类似的内容:

Meaning that mpirun -np 4 python2 from_python.py is supposed to give something like:

你好，世界！我正在处理0之4.
你好，世界！我是第1步，共4步.
你好，世界！我正在执行第2步，共4步.
你好，世界！我正在处理3之4.

Hello, World! I am process 0 of 4.
Hello, World! I am process 1 of 4.
Hello, World! I am process 2 of 4.
Hello, World! I am process 3 of 4.

但是，如果我尝试通过cython来实现这一目标，

But if I try to achieve this via cython like this:

helloworld_wrap.pyx:

cimport mpi4py.MPI as MPI
cimport mpi4py.libmpi as libmpi

cdef extern from "helloworld.h":
   void sayhello(libmpi.MPI_Comm comm)

def py_sayhello(MPI.Comm comm):
    sayhello(comm)

和:

setup.py:

import os
from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext

mpi_compile_args = os.popen("mpicc --showme:compile").read().strip().split(' ')
mpi_link_args    = os.popen("mpicc --showme:link").read().strip().split(' ')

ext_modules=[
    Extension("helloworld_wrap",
              sources            = ["helloworld_wrap.pyx", "helloworld.c"],
              language           = 'c',
              extra_compile_args = mpi_compile_args,
              extra_link_args    = mpi_link_args,
          )
]

setup(
  name = "helloworld_wrap",
  cmdclass = {"build_ext": build_ext},
  ext_modules = ext_modules
)

我收到以下错误消息:

helloworld_wrap.pyx:8:13:无法将Python对象转换为"MPI_Comm"

helloworld_wrap.pyx:8:13: Cannot convert Python object to 'MPI_Comm'

表示不能将mpi4py.MPI.Comm转换为MPI_Comm.那么如何将mpi4py.MPI.Comm转换为MPI_Comm以便使包装器正常工作?

indicating that mpi4py.MPI.Comm can't be converted into an MPI_Comm. So how can I convert the mpi4py.MPI.Comm into an MPI_Comm in order to get my wrapper to work?

推荐答案

转换非常简单，因为mpi4py.MPI.Comm对象内部将MPI_Comm句柄存储为成员ob_mpi ¹ .因此，如果将helloworld_wrap.pyx的最后一行更改为通过comm.ob_mpi而不是comm，则模块将按预期进行编译和工作:

The conversion is rather simple, as a mpi4py.MPI.Comm-object internally stores an MPI_Comm handle as member ob_mpi¹. Therefore, if one changes the last line of helloworld_wrap.pyx to pass comm.ob_mpi instead of comm, the module compiles and works as intended:

helloworld_wrap.pyx:

cimport mpi4py.MPI as MPI
cimport mpi4py.libmpi as libmpi

cdef extern from "helloworld.h":
   void sayhello(libmpi.MPI_Comm comm)

def py_sayhello(MPI.Comm comm):
    sayhello(comm.ob_mpi)

令人惊讶的是，我没有找到任何相关文档，只是在研究

Surprisingly, I did not find any documentation for this but only realized this when studying the sources of mpi4py.MPI.Comm. I am not sure, if this is the intended way to handle this, but I could not get it to work otherwise.

¹ 实际上，如果mpi4py.MPI中的大多数对象(如果不是在C中为相应的MPI句柄建模的所有对象)，都将相应的句柄作为成员ob_mpi持有.

¹ In fact, most, if not all objects in mpi4py.MPI that model a corresponding MPI handle in C, hold a corresponding handle as member ob_mpi.

这篇关于如何通过cython将MPI通信器从python传递到C?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！