Nowadays, arithmetic operations for storage classes of any sizes are not necessarily compiled and executed as atomic instruction.

atomix-test.cpp

#include<iostream>
#include<atomic>
#include<thread>
#include<mutex>
#include<condition_variable>
#include<mutex>

/*
 * Refereces:
 * - std::unique_lock: https://cpprefjp.github.io/reference/mutex/unique_lock.html
 * - std:condition_variable: http://ja.cppreference.com/w/cpp/thread/condition_variable
 * - std::mutex: https://cpprefjp.github.io/reference/mutex/mutex.html
 * - std::thread: https://cpprefjp.github.io/reference/thread/thread.html
 * - random: https://cpprefjp.github.io/reference/thread/thread.html
 * - https://www.ibm.com/developerworks/jp/linux/library/l-posix3/index.html
 * - https://fossies.org/dox/glibc-2.25/atomic_8h.html
 * - http://qiita.com/termoshtt/items/c01745ea4bcc89d37edc
 * - http://torini.hateblo.jp/entry/2014/08/08/003352
 * - https://cpplover.blogspot.jp/2013/11/blog-post_4069.html
 * - An overview to polyhedral model: https://www.cs.indiana.edu/~achauhan/Teaching/B629/2010-Fall/StudentPresns/PolyhedralModelOverview.pdf
 */

constexpr std::size_t TH_MAX = 16 * 16 * 16;
constexpr std::size_t WEIGHT = 100;
//constexpr std::size_t TH_MAX = 16;
//constexpr std::size_t WEIGHT = 1;

#ifdef ATOMIC
using Mode = std::atomic<int>;
#else
using Mode = int;
#endif

struct Sync{};
struct NoSync{};

using syncQ = Sync;


// コンパイル時評価
/*
template<typename T>
void tasks_sync(auto i){ }
*/

// 評価されない
template<typename T>
void tasks_sync(decltype(TH_MAX)& i){}

template<>
void tasks_sync<Sync>(decltype(TH_MAX)& i){
	while(i < TH_MAX){
		std::this_thread::yield();
	}
}


/*
template<typename T>
void tasks_sync(decltype(TH_MAX)* i){ }

template<>
void tasks_sync<Sync>(decltype(TH_MAX)* i){
	while(*i < TH_MAX){
		std::this_thread::yield();
	}
}
*/

template<typename T>
int target(void){
	T ctr;
	ctr = 0;
	std::thread tasks[TH_MAX];

	for(auto i = decltype(TH_MAX)(0); i < TH_MAX; i++)
	{
		tasks[i] = std::thread([&]()
		{
			tasks_sync<syncQ>(i);
			//tasks_sync<syncQ>(&i);
			for(auto j = decltype(WEIGHT)(0);j++ < WEIGHT;) {
				ctr-=1;
				ctr+=2;
			}
		});
	}
	for(auto i = decltype(TH_MAX)(0); i < TH_MAX; i++){
		tasks[i].join();
	}

	std::cout << TH_MAX * WEIGHT << " ?=? "<< ctr << std::endl;
	return 0;
}

int main(void){
	target<Mode>();
	// target();
	return 0;
}