如何将结构内的 numpy 数组传递给 ctypes 函数？

Question

I'm trying to wrap some third-party code from a library in my Python application. Essentially, the function I would like to call takes a struct as input that houses (among other things) pointers to double arrays. A simplified example would be:

myfunc.h:

typedef struct mystruct_t {
    int n;
    double *x;
} mystruct_t;

double myfunc(mystruct_t* data);

myfunc.c :

#include "myfunc.h"

double myfunc(mystruct_t* data) {
    double result = 0.;
    for(int i = 0; i < data->n; i++) {
        result += data->x[i];
    }
    return result;
}

Makefile :

CC = gcc
CFLAGS = -g -Wall -fPIC -lm -std=gnu99

all: libmyfunc.so

m.PHONY : clean

libmyfunc.so: myfunc.o
    gcc -shared -Wl,-soname,$@ -o $@ $^

%.o: %.c
    $(CC) -c $(CFLAGS) $<

clean:
    rm -vf libmyfunc.so *.o

I would like to wrap myfunc using NumPy, ctypes.Structure and numpy.ctypeslib so that I can pass NumPy arrays to myfunc as attributes of mystruct_t . So far I've been trying the following:

myfunc.py:

#!/usr/bin/env python
import numpy as np
import numpy.ctypeslib as npct
import ctypes
import os

array_1d_double = npct.ndpointer(dtype=np.double, ndim=1, flags='C_CONTIGUOUS')


class MyStruct(ctypes.Structure):
    _fields_ = [
        ('n', ctypes.c_int16),
        ('x', array_1d_double)
    ]

libmyfunc = npct.load_library('libmyfunc', os.path.dirname(__file__))
libmyfunc.myfunc.restype = ctypes.c_double
libmyfunc.myfunc.argtypes = [
    ctypes.POINTER(MyStruct)
]

x = np.array([1.0, 2.0, 3.0, 4.0])

mystruct = MyStruct()
mystruct.n = len(x)
mystruct.x = x

res = libmyfunc.myfunc(mystruct)

However, this fails with the following error message:

$ python myfunc.py
Traceback (most recent call last):
  File "./myfunc.py", line 26, in <module>
    mystruct.x = x
TypeError: cannot be converted to pointer

How do I define my function signatures correctly so that type conversion can happen? Or do I need to convert x somehow before assigning it to mystruct.x ?

Unfortunately I cannot change the signature of the method I would like to call and I would like to not have to write wrapper C code unless it is absolutely necessary. The other questions and resources that I've found on this only deal with either ctypes.Structure s or numpy.ctypeslib , but can you make them work together?

I've uploaded my reduced example as a gist so you can use it as a starting point.

Many thanks in advance!

Answer 1

You can remove array_1d_double . It is not needed.

The struct should be declared like this:

class MyStruct(ctypes.Structure):
    _fields_ = [
        ('n', ctypes.c_int),
        ('x', ctypes.POINTER(ctypes.c_double))
    ]

I've changed both types. You had c_int16 , but the type you used in the C code is int . That maps to c_int . And likewise for the array, that is double* . In ctypes that is POINTER(ctypes.c_double) .

The struct should be initialised like this:

mystruct = MyStruct()
mystruct.n = len(x)
mystruct.x = npct.as_ctypes(x)

With these changes, your code works as intended.