Resolution of std::chrono::high_resolution_clock doesn't correspond to measurements

Question

Let me ask my question by this test program:

#include <iostream>
#include <chrono>

using std::chrono::nanoseconds;
using std::chrono::duration_cast;

int main(int argc, char* argv[])
{
    std::cout 
      << "Resolution (nano) = " 
      << (double) std::chrono::high_resolution_clock::period::num / 
                  std::chrono::high_resolution_clock::period::den * 
                  1000 * 1000 * 1000 
      << std::endl;

    auto t1 = std::chrono::high_resolution_clock::now();
    std::cout << "How many nanoseconds does std::cout take?" << std::endl;
    auto t2 = std::chrono::high_resolution_clock::now();

    auto diff = t2-t1;
    nanoseconds ns = duration_cast<nanoseconds>(diff);

    std::cout << "std::cout takes " << ns.count() << " nanoseconds" 
              << std::endl;
    return 0;
}

Output on my machine:

Resolution (nano) = 100

How many nanoseconds does std::cout take?

std::cout takes 1000200 nanoseconds

I receive either 1000200 or 1000300 or 1000400 or 1000500 or 1000600 or 2000600 as a result (= 1 or 2 microsecond). Obviously, either the resolution of std::chrono is not 100 nano-seconds or the way I measure the time of std::cout is wrong. (Why do I never receive something between 1 and 2 microseconds, for example 1500000?)

I need a high-resolution timer in C++. The OS itself provides a high-resolution timer, because I'm able to measure things with microsecond-precision using the C# Stopwatch class on the same machine. So I would just need to correctly use the high-resolution timer that the OS has!

How do I fix my program to produce the expected results?

Answer 1

I'm going to guess you are using Visual Studio 2012. If not, disregard this answer. Visual Studio 2012 typedef's high_resolution_clock to system_clock. Sadly, this means it has crappy precision (around 1 ms). I wrote a better high-resolution clock which uses QueryPerformanceCounter for use in Visual Studio 2012...

HighResClock.h:

    struct HighResClock
    {
        typedef long long                              rep;
        typedef std::nano                              period;
        typedef std::chrono::duration<rep, period>     duration;
        typedef std::chrono::time_point<HighResClock>  time_point;
        static const bool is_steady = true;

        static time_point now();
    };

HighResClock.cpp:

namespace
{
    const long long g_Frequency = []() -> long long
    {
        LARGE_INTEGER frequency;
        QueryPerformanceFrequency(&frequency);
        return frequency.QuadPart;
    }();
}

HighResClock::time_point HighResClock::now()
{
    LARGE_INTEGER count;
    QueryPerformanceCounter(&count);
    return time_point(duration(count.QuadPart * static_cast<rep>(period::den) / g_Frequency));
}

(I left out an assert and #ifs to see if it's being compiled on Visual Studio 2012 from the above code.)

You can use this clock anywhere and in the same way as standard clocks.

Answer 2

The resolution of a clock is not necessarily the same as the smallest duration that can be represented by the data type the clock uses. In this case your implementation uses a data type which can represent a duration as small as 100 nanoseconds, but the underlying clock doesn't actually have such a resolution.

---

The low resolution of Visual Studio's high_resolution_clock has been an issue for several years. Microsoft's C++ standard library maintainer, Stephan T. Lavavej, has indicated that this has been fixed in Visual Studio 2015 via the use of QueryPerformanceCounter().

Answer 3

Maybe the implementation doesn't implement the higher resolution timer?

It seems you are using Windows (you mention C#), so if you use a timer and you are indeed using Windows, you can use QueryPerformanceFrequency and QueryPerformanceCounter.