I recently needed to use memory that was allocated inside a C++ library in a Python application which expects a NumPy array without performing any copies. With ctypes, this can be implemented quite easily. Let’s say we have a shared library libcpp.so where a function get_shared_memory returns the pointer to an array of doubles that is stored on the heap:


double* get_shared_memory(std::size_t num) {
    auto p = new double[num];
    ...
    return p;
}

The function can be called from Python with ctypes:


num_elem = 42

libcpp = ctypes.CDLL("libcpp.so")
libcpp.get_shared_memory.restype = ctypes.c_void_p 
pointer = libcpp.get_shared_memory(num_elem)

Afterwards, pointer can be interpreted from Python as a pointer to a double:

double_p = ctypes.cast(pointer, ctypes.POINTER(ctypes.c_double))

At this point, double_p is a object of type LP_c_double. It can already be used to set and get values of the C++ array, for instance we can do:

double_p[0] = 1.0
print(double_p[2])

But we cannot use a LP_c_double as a normal Python lists. For instance, there is no way for Python to know the length of the underlying buffer and calling len(double_p) will fail. However, we can also use pointer to create a NumPy array on top of the existing memory thanks to np.ctypeslib.as_array:

np_buff = np.ctypeslib.as_array((ctypes.c_double * num_elem).from_address(pointer))

np_buff can now be used as a normal NumPy array. When you are done using it, you need to free the allocated memory in C++ again, as you will leak memory otherwise. The pointer that is returned in get_shared_memory should therefore be stored somewhere.