ctypes 在使用指向 class object 的指针调用 function 时给出错误

Question

I have two wrapper functions in my C++ DLL.

The function CreateObj() makes an object of some other class and returns a pointer to the object In the second function DoSomething I pass the previously created pointer as first argument and the second argument is a RGB image (1200, 1200,3) of unsigned chars.

On the python side I load the DLL and want to call the two wrapper functions to do some processing on a RGB image that I have loaded on the python side. I am getting an OSError in the very last step of calling DoSomething from python.

Any help will be greatly appreciated.

//The C++ library code

class Obj
{
    /*..../*
};

extern "C"
{
#ifdef BUILD_AS_DLL
    __declspec(dllexport)
#endif
    Obj* CreateObj();
    int DoSomething( Obj* ptr,  uint8_t*  t[1200][1200][]);  //ptr was created using CreateObj and second argument is an image array of unsigned char of shape [1200][1200][3]
#ifdef BUILD_AS_DLL
    __declspec(dllexport)
#endif
}

The Python code for calling the above DLL

from ctypes import *
import numpy as np
import numpy.ctypeslib as npct
import cv2
    
path = os.path.join(os.getcwd(), "mwe.dll") 
lib = ctypes.CDLL(path) 

create_func = lib.CreateObj
create_func.restype = c_void_p
obj = c_void_p(create_func())

dosomething_func = lib.DoSomething
ucharPtr = npct.ndpointer(dtype=np.uint8, 
                          ndim=3, 
                          shape = (1200,1200,3),
                          flags='CONTIGUOUS') 
dosomething_func.argtype = (c_void_p,
                            ucharPtr)
dosomething_func.restype = c_int

img = cv2.imread("path/to/image")  
img_ctype = byref(img.ctypes.data_as(c_void_p))
dosomething_func(byref(obj), img_ctype)   // ERROR on this line

I get error

File "mwe.py", line 24, in <module>
    res = dosomething_func(
OSError: exception: access violation reading 0xFFFFFFFFFFFFFFFF

Answer 1

Here's a minimal reproducible example of a solution demonstrating passing a small 3D numpy array into and back out of a function:

test.cpp:

Note that uint8_t* t[1200][1200][] was syntactically incorrect. The variable array dimension must be first, not last, and this is an array of pointers not an array of 8-bit data.

#include <cstdint>
#include <iostream>

using namespace std;

#define API __declspec(dllexport)

class Obj
{
    int m_x;
public:
    Obj(int x): m_x(x) {}
    int get_x() const { return m_x; }
};

extern "C" {

API Obj* CreateObj() {
    return new Obj(5);
}

API int DoSomething(Obj* ptr, uint8_t t[][3][3]) {
    for(uint8_t i=0; i < 3; ++i) {
        for(uint8_t j=0; j < 3; ++j) {
            for(uint8_t k=0; k < 3; ++k) {
                cout << static_cast<uint32_t>(t[i][j][k]) << ' ';
                t[i][j][k] *= 2;
            }
            cout << endl;
        }
        cout << endl;
    }
    return ptr->get_x();
}

}

test.py:

Assuming cv2.imread() returns an np.array of the same type and shape as the declared ndpointer type, you would just pass it to the function directly.

from ctypes import *
import numpy as np
import numpy.ctypeslib as npct

lib = CDLL('./test')

class OBJ(Structure):
    _fields_ = []

ARRAY = npct.ndpointer(dtype=np.uint8,shape=(3,3,3))

lib.CreateObj.argtypes = c_int,
lib.CreateObj.restype = POINTER(OBJ)
lib.DoSomething.argtypes = POINTER(OBJ),ARRAY
lib.DoSomething.restype = c_int

arr = np.arange(27,dtype=np.uint8).reshape((3,3,3))
obj = lib.CreateObj(5)
result = lib.DoSomething(obj,arr)
print(f'result = {result}')
print(arr)

Output:

0 1 2
3 4 5
6 7 8

9 10 11
12 13 14
15 16 17

18 19 20
21 22 23
24 25 26

result = 5
[[[ 0  2  4]
  [ 6  8 10]
  [12 14 16]]

 [[18 20 22]
  [24 26 28]
  [30 32 34]]

 [[36 38 40]
  [42 44 46]
  [48 50 52]]]