How can we cast a `ctypes.POINTER(ctypes.c_float)` to `int`?

Question

I think this is a simple task, but I could not find a solution on the web to this. I have a external C++ library, which I'm using in my Python code, returning a ctypes.POINTER(ctypes.c_float) to me. I want to pass an array of these pointers to a jax.vmap function. The problem is that jax does not accept the ctypes.POINTER(ctypes.c_float) type. So, can I somehow cast this pointer to an ordinary int. Technically, this is clearly possible. But how do I do this in Python?

Here is an example:

lib = ctypes.cdll.LoadLibrary(lib_path)
lib.foo.argtypes = None
lib.foo.restype = ctypes.POINTER(ctypes.c_float)

bar = jax.vmap(lambda : dummy lib.foo())(jax.numpy.empty(16))

x = jax.numpy.empty(16, 256, 256, 1)
y = jax.vmap(lib.bar, in_axes = (0, 1))(x, bar)

So, I want to invoke lib.foo 16-times so that I have an array bar containing all the pointers. Then I want to invoke another library function lib.bar which expects bar together with another (batched) parameter x.

The problem is that jax claims that ctypes.POINTER(ctypes.c_float) is not a valid jax type. This is why I think the solution is to cast the pointers to ints and store those ints in bar instead.

Answer 1

Listing:

• [SO]: C function called from Python via ctypes returns incorrect value (@CristiFati's answer) - a common pitfall when working with CTypes (calling functions)

• [Python.Docs]: ctypes - A foreign function library for Python

Here's a piece of code exemplifying how to handle pointers and their addresses. The trick is to use ctypes.addressof (documented in the 2^nd URL).

code00.py:

#!/usr/bin/env python

import ctypes as cts
import sys


CType = cts.c_float
CTypePtr = cts.POINTER(CType)


def ctype_pointer(seq):  # Helper
    CTypeArr = (CType * len(seq))
    ctype_arr = CTypeArr(*seq)
    return cts.cast(ctype_arr, CTypePtr)


def pointer_elements(addr, count):  # Helper
    return tuple(CType.from_address(addr + i * cts.sizeof(CType)).value for i in range(count))


def main(*argv):
    seq = (2.718182, -3.141593, 1.618034, -0.618034, 0)
    ptr = ctype_pointer(seq)
    print(f"Pointer: {ptr}")
    print(f"\nPointer elements: {tuple(ptr[i] for i in range(len(seq)))}")  # Check if pointer has correct data
    ptr_addr = cts.addressof(ptr.contents)  # @TODO - cfati: Straightforward
    print(f"\nAddress: {ptr_addr} (0x{ptr_addr:016X})\nElements from address: {pointer_elements(ptr_addr, len(seq))}")
    ptr_addr0 = cts.cast(ptr, cts.c_void_p).value  # @TODO - cfati: Alternative
    print(f"\nAddresses match: {ptr_addr == ptr_addr0}")


if __name__ == "__main__":
    print(
        "Python {:s} {:03d}bit on {:s}\n".format(
            " ".join(elem.strip() for elem in sys.version.split("\n")),
            64 if sys.maxsize > 0x100000000 else 32,
            sys.platform,
        )
    )
    rc = main(*sys.argv[1:])
    print("\nDone.\n")
    sys.exit(rc)

Notes:

• Although it adds a bit of complexity, I introduced the CType "layer" to show that it should work with any type, not just float (as long as the values in the sequence are of that type)

• The only truly relevant lines are those marked with @TODO

Output:

(py_pc064_03.08_test0_lancer) [cfati@cfati-5510-0:/mnt/e/Work/Dev/StackExchange/StackOverflow/q078366208]> python ./code00.py 
Python 3.8.19 (default, Apr  6 2024, 17:58:10) [GCC 11.4.0] 064bit on linux

Pointer: <__main__.LP_c_float object at 0x7203e97e7d40>

Pointer elements: (2.71818208694458, -3.1415929794311523, 1.6180340051651, -0.6180340051651001, 0.0)

Address: 125361127594576 (0x00007203E97A9A50)
Elements from address: (2.71818208694458, -3.1415929794311523, 1.6180340051651, -0.6180340051651001, 0.0)

Addresses match: True

Done.