================== omptest.c ==================
#include <omp.h>
#include <stdlib.h>
#include <math.h>
#ifdef PYTHON_MODULE
#include <Python.h>
#endif

static unsigned short *a = NULL;
static unsigned short *b = NULL;
static unsigned short *c = NULL;
static double *r = NULL;
static int length = 0;

static void init_test(int len) {
    int i;
    if (length)
        return;
    length = len;
    a = (unsigned short*)calloc(length, sizeof(unsigned short));
    b = (unsigned short*)calloc(length, sizeof(unsigned short));
    c = (unsigned short*)calloc(length, sizeof(unsigned short));
    r = (double*)calloc(length, sizeof(double));
    for (i = 0; i < length; i++) {
        a[i] = (unsigned short)(((i + 13) * 4001) % 2017);
        b[i] = (unsigned short)(((i + 17) * 4001) % 2017);
        c[i] = (unsigned short)(((i + 19) * 4001) % 2017);
    }
}

static int do_test(int use_omp, int count) {
    int i;
    int threads = 1;
    if (use_omp) {
#ifdef PYTHON_MODULE
        Py_BEGIN_ALLOW_THREADS
#endif
        threads = omp_get_max_threads();
        omp_set_num_threads(threads);
        while (count--) {
            #pragma omp parallel for private(i)
            for (i = 0; i < length; i++)
                r[i] = (double)a[i] / pow((double)b[i] + 1.0, 2) +
                       sqrt(fabs((double)c[i]));
        }
#ifdef PYTHON_MODULE
        Py_END_ALLOW_THREADS
#endif
    }
    else {
        while (count--) {
            for (i = 0; i < length; i++)
                r[i] = (double)a[i] / pow((double)b[i] + 1.0, 2) +
                       sqrt(fabs((double)c[i]));
        }
    }
    return threads;
}

#ifdef PYTHON_MODULE
static PyObject *omptest_init(PyObject *self, PyObject *args) {
    int length;
    if (!PyArg_ParseTuple(args, "i", &length))
        return NULL;
    init_test(length);
    Py_RETURN_NONE;
}

static PyObject *omptest_do(PyObject *self, PyObject *args) {
    int use_omp, count;
    int threads;
    if (!PyArg_ParseTuple(args, "ii", &use_omp, &count))
        return NULL;
    threads = do_test(use_omp, count);
    return Py_BuildValue("i", threads);
}
static PyMethodDef OMPTestMethods[] = {
    {"init", omptest_init, METH_VARARGS, ""},
    {"do", omptest_do, METH_VARARGS, ""},
    {NULL, NULL, 0, NULL}
};

#if PY_MAJOR_VERSION >= 3
static struct PyModuleDef omptestmodule = {
    PyModuleDef_HEAD_INIT,
    "omptest",
    "",
    -1,
    OMPTestMethods
};
PyMODINIT_FUNC PyInit_omptest(void)
{
    PyEval_InitThreads();
    return PyModule_Create(&omptestmodule);
}
#else
PyMODINIT_FUNC initomptest(void)
{
    PyEval_InitThreads();
    (void) Py_InitModule("omptest", OMPTestMethods);
}
#endif
#else
#include <stdio.h>
#include <sys/time.h>
#include <time.h>
int main() {
    int use_omp, threads;
    struct timeval tv;
    double t1, t2;
    init_test(1024*1024);
    for (use_omp = 0; use_omp <= 1; use_omp++) {
        gettimeofday(&tv, NULL);
        t1 = (double)tv.tv_sec + (double)tv.tv_usec * 1e-6;
        threads = do_test(use_omp, 100);
        gettimeofday(&tv, NULL);
        t2 = (double)tv.tv_sec + (double)tv.tv_usec * 1e-6;
        printf("[C] use_omp = %d, threads = %d, elapsed %.3fs\n",
               use_omp, threads, t2 - t1);
    }
}
#endif

================== setup.py ==================
# -*- coding: utf-8 -*-

from distutils.core import setup, Extension

module1 = Extension('omptest',
                    sources=['omptest.c'],
                    extra_compile_args=['-DPYTHON_MODULE', '-O4', '-fopenmp'],
                    extra_link_args=['-lgomp', '-lm'])

setup(name='OMPTest',
      version='0.0.0',
      description='Test OpenMP withing C extension for Python',
      ext_modules=[module1])

================== pyomptest.c ==================
# -*- coding: utf-8 -*-

import time
import omptest

omptest.init(1024*1024)
for use_omp in [0, 1]:
    t1 = time.clock()
    threads = omptest.do(use_omp, 100)
    t2 = time.clock()
    print("[Py] use_omp = %d, threads = %d, elapsed %.3fs" %
          (use_omp, threads, t2 - t1))

================== Makefile ==================
all: comptest pyomptest

comptest: omptest.c
	gcc -O4 -fopenmp -lgomp -lm omptest.c -o omptest
	./omptest

pyomptest: omptest.c pyomptest.py setup.py
	python setup.py build
	python setup.py install
	python pyomptest.py

================== "sudo make" stdandard output ==================
[C] use_omp = 0, threads = 1, elapsed 1.177s
[C] use_omp = 1, threads = 4, elapsed 0.663s
[Py] use_omp = 0, threads = 1, elapsed 1.171s
[Py] use_omp = 1, threads = 4, elapsed 2.610s