from typing import Callable

def pack_lut_4(f: Callable[..., int], order: tuple[int, int, int, int] = (0, 1, 2, 3)) -> int:
    lut = 0
    for idx in range(16):
        bits = [(idx >> i) & 1 for i in range(4)]  # b0, b1, b2, b3
        args = [0, 0, 0, 0]
        for bit_pos, arg_pos in enumerate(order):
            args[arg_pos] = bits[bit_pos]
        lut |= (f(*args) & 1) << idx
    return lut


# F31 coordinates
def F31_0(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1) ^ (x0 & y0 & y1) ^ (x0 & y0) ^ (x0 & y1) ^ (x1 & y0 & y1) ^ (y0 & y1) ^ y0 ^ y1

def F31_1(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1) ^ (x0 & y0 & y1) ^ (x0 & y1) ^ x0 ^ (x1 & y0) ^ (x1 & y1) ^ (y0 & y1) ^ y0 ^ y1 ^ 1

def F31_2(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & y0 & y1) ^ (x0 & y0) ^ x0 ^ (x1 & y1) ^ y0

def F31_3(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1 & y1) ^ (x0 & x1) ^ (x0 & y0 & y1) ^ (x0 & y1) ^ (x1 & y0 & y1) ^ x1 ^ y1

def F31_4(x0, x1, y0, y1):
    return (x0 & x1 & y1) ^ (x0 & x1) ^ (x0 & y0) ^ (x0 & y1) ^ (x1 & y0 & y1) ^ (x1 & y0) ^ x1 ^ (y0 & y1) ^ y0 ^ y1 ^ 1

def F31_5(x0, x1, y0, y1):
    return (x0 & x1 & y1) ^ (x0 & y0) ^ x0 ^ (x1 & y0 & y1) ^ (x1 & y1) ^ x1 ^ y0 ^ y1 ^ 1

def F31_6(x0, x1, y0, y1):
    return (x0 & x1 & y1) ^ (x0 & y0) ^ (x0 & y1) ^ (x1 & y0 & y1) ^ (x1 & y1) ^ x1 ^ (y0 & y1) ^ 1

def F31_7(x0, x1, y0, y1):
    return (x0 & x1 & y1) ^ (x0 & y0 & y1) ^ (x0 & y0) ^ (x0 & y1) ^ (x1 & y0 & y1) ^ x1 ^ y1 ^ 1


# F32 coordinates
def F32_0(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1) ^ (x0 & y0 & y1) ^ (x0 & y0) ^ (x0 & y1) ^ (x1 & y0 & y1) ^ y0 ^ y1 ^ 1

def F32_1(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1) ^ (x0 & y0 & y1) ^ (x0 & y1) ^ x0 ^ (x1 & y1) ^ x1 ^ y0 ^ 1

def F32_2(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1 & y1) ^ (x0 & y0 & y1) ^ (x0 & y0) ^ x0 ^ (x1 & y0 & y1) ^ y0

def F32_3(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1 & y1) ^ (x0 & y0 & y1) ^ (x1 & y0 & y1) ^ (x1 & y0) ^ x1 ^ (y0 & y1) ^ y1

def F32_4(x0, x1, y0, y1):
    return (x0 & x1 & y0) ^ (x0 & x1 & y1) ^ (x0 & x1) ^ (x0 & y0 & y1) ^ (x0 & y0) ^ x0 ^ (x1 & y0 & y1) ^ x1 ^ (y0 & y1) ^ y1

def F32_5(x0, x1, y0, y1):
    return (x0 & x1 & y1) ^ (x0 & y0) ^ (x1 & y0 & y1) ^ (x1 & y1) ^ x1 ^ y1

def F32_6(x0, x1, y0, y1):
    return (x0 & x1 & y1) ^ (x0 & x1) ^ (x0 & y0) ^ (x0 & y1) ^ (x1 & y0 & y1) ^ (x1 & y0) ^ x1 ^ (y0 & y1) ^ y0 ^ y1

def F32_7(x0, x1, y0, y1):
    return (x0 & x1 & y1) ^ (x0 & y0 & y1) ^ (x0 & y0) ^ (x1 & y0 & y1) ^ x1 ^ y0 ^ y1 ^ 1

def main():
    F31_orders = [
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (3, 0, 1, 2),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
    ]

    F32_orders = [
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (3, 0, 1, 2),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
        (0, 1, 2, 3),
    ]

    F31_coords = [F31_0, F31_1, F31_2, F31_3, F31_4, F31_5, F31_6, F31_7]
    F32_coords = [F32_0, F32_1, F32_2, F32_3, F32_4, F32_5, F32_6, F32_7]

    lut_B = [pack_lut_4(f, o) for f, o in zip(F31_coords, F31_orders)]
    lut_A = [pack_lut_4(f, o) for f, o in zip(F32_coords, F32_orders)]

    print("TEA3_LUT_A = [", end="")
    for x in lut_A:
        print(f"0x{x:04X}, ", end="")
    print("]")

    print("TEA3_LUT_B = [", end="")
    for x in lut_B:
        print(f"0x{x:04X}, ", end="")
    print("]")