Why doesn't LoadLibrary work in Worker function when using multiprocessing?

Question

Python

I'm trying to optimize the program's running time by using the DLL in the mpCalcDllWorker function.
DLL returns some kind of garbage. In single-threaded mode, the DLL runs flawlessly.

Here is a simplified example.

def mpCalcDllWorker(nodes, queue, arrD):

    dll_fn = "***"
    lib = ctypes.cdll.LoadLibrary(dll_fn)
    print("lib:", lib)
    D = np.reshape( np.frombuffer(arrD), 1000000 )

    while True:
        job = queue.get()
        if job == None:
            break

        start = job[0]
        stop = job[1]
        lib.dll_add_ts.restype = ctypes.c_double
        ret = lib.dll_add_ts( nodes.ctypes.data_as(ctypes.c_void_p), ctypes.c_int(start), ctypes.c_int(stop) )
        print("start", start, "stop", stop, ret)
        print("nodes", nodes)
        D[start] = ret

        queue.task_done()
    queue.task_done()

def mpCalcWorker(nodes, queue, arrD):
    """ Worker process for the multiprocessing calculations
    """
    D = np.reshape( np.frombuffer(arrD), 1000000 )

    while True:
        job = queue.get()
        if job == None:
            break

        start = job[0]
        stop  = job[1]

        D[start] = sum(nodes[start:stop])

        queue.task_done()
    queue.task_done()

def mpCalcSums(arr, res, tp=0):
    # allocate shared array

    arrD = mp.RawArray(ctypes.c_double, 1000000)

    # setup jobs
    nCPU = mp.cpu_count()
    jobs = []
    for i in range(100000):
        jobs.append((i, i+10000))

    queue = mp.JoinableQueue()
    for job in jobs:
        queue.put(job)
    for i in range(nCPU):
        queue.put(None)

    # run workers
    workers = []
    for i in range(nCPU):
        if tp==0:
            worker = mp.Process(target=mpCalcWorker,    args=(arr, queue, arrD))
        if tp==1:
            worker = mp.Process(target=mpCalcDllWorker, args=(arr, queue, arrD))
        workers.append(worker)
        #worker.daemon=True  - не влияет на ошибку
        worker.start()
    queue.join()

    # make array from shared memory
    D = np.reshape(np.frombuffer(arrD), res.shape)
    return D

Is it possible to use functions imported from a DLL in Workers?
Please point out my mistake if possible.