compare normal, c++11 thread and openmp

test_omp.cpp

#include <iostream>
#include <ctime>
#include <random>
#include <vector>
#include <thread>
#include <omp.h>

using namespace std;

const static int TOTAL_SIZE = 100000;
const static int THREAD_NUM = omp_get_max_threads();
const static int RANGE = 100000;
const static int WALK_LEN = 40;

std::mt19937 rng(static_cast<int>(time(NULL)));
std::uniform_int_distribution<int> ud(0, RANGE);

inline vector<int> get_walk() {
	vector<int> walk;
	walk.reserve(WALK_LEN);
	for (int j = 0; j < WALK_LEN; ++j) {
		walk.push_back(ud(rng));
	}
	return walk;
}

inline vector<vector<int> > foo() {
	vector<vector<int> > rs(TOTAL_SIZE, vector<int>());
	for (int i = 0; i < TOTAL_SIZE; i++) {
		rs[i] = get_walk();
	}
	return rs;
}

void task(vector<vector<int> >::iterator begin, vector<vector<int> >::iterator end) {
	for (auto iter = begin; iter != end; ++iter) {
		vector<int> walk = get_walk();
		*iter = walk;
	}
}

inline vector<vector<int> > foo_thread() {
	vector<vector<int> > rs(TOTAL_SIZE, vector<int>());
	vector<thread> threads(THREAD_NUM);
	int gap = TOTAL_SIZE / THREAD_NUM;
	vector<vector<int> >::iterator iter = rs.begin();
	for (auto & t : threads) {
		t = thread(task, iter, iter + gap > rs.end() ? rs.end() : iter + gap);
		iter += gap;
	}
	for (auto & t : threads) {
		t.join();
	}
	return rs;
}

inline vector<vector<int> > foo_omp() {
	vector<vector<int> > rs(TOTAL_SIZE, vector<int>());
	# pragma omp parallel for num_threads(THREAD_NUM)
	for (int i = 0; i < TOTAL_SIZE; i++) {
		rs[i] = get_walk();
	}
	return rs;
}

int main() {
	clock_t start = clock();
	vector<vector<int> > rs = foo();
	clock_t end = clock();
	cout << ((float)(end - start)) / CLOCKS_PER_SEC * 1000 << " ms." << endl;
	cout << rs.size() << endl;

	start = clock();
	rs = foo_thread();
	end = clock();
	cout << ((float)(end - start)) / CLOCKS_PER_SEC * 1000 << " ms." << endl;
	cout << rs.size() << endl;

	start = clock();
	rs = foo_omp();
	end = clock();
	cout << ((float)(end - start)) / CLOCKS_PER_SEC * 1000 << " ms." << endl;
	cout << rs.size() << endl;

	return 0;
}

7031 ms
100000
2298 ms
100000
2381 ms
100000