atomic load/store

Hi

In the following code running the task5 routine from 2 thread simultaneously on a multiprocessor machine takes approx twice as long as running it from a single thread. The atomic loads/stores seem to influence each other (at least on x86/64) even though there is no explicit synchronization.

std::atomic<size_t> sdata;

#pragma optimize( "", off )
void task5()
{
	size_t acc;
	for (int i = 0; i < 1000000000; ++i)
	{
		sdata.store(i, std::memory_order_relaxed);
		acc += sdata.load(std::memory_order_relaxed);
	}
	std::cout << acc << std::endl;
}
#pragma optimize( "", on ) 

int main()
{	
	
	StopWatch sw;

	sw.Start();

	std::thread trd1 = std::thread(task5);
	std::thread trd2 = std::thread(task5); // comment out to test with a single thread...
	trd1.join();
	trd2.join(); // ... this one as well

	sw.Stop();

	std::cout << "DONE " << sw.Get_ElapsedMilliseconds() << std::endl;

return 0;
}

Why are these threads blocking each other?