A duration has a representation and a tick period (precision).

template <class Rep, class Period = ratio<1> > class duration;

The representation is simply any arithmetic type, or an emulation of such a type. The representation stores a count of ticks. This count is the only data member stored in a duration. If the representation is floating-point, it can store fractions of a tick to the precision of the representation. The tick period is represented by a ratio and is encoded into the duration's type, instead of stored. The tick period only has an impact on the behavior of the duration when a conversion between different durations is attempted. The tick period is completely ignored when simply doing arithmetic among like durations.

Example:

typedef boost::chrono::duration<long, boost::ratio<60> > minutes;
minutes m1(3);                 // m1 stores 3
minutes m2(2);                 // m2 stores 2
minutes m3 = m1 + m2;          // m3 stores 5

typedef boost::chrono::duration<long long, boost::micro> microseconds;
microseconds us1(3);           // us1 stores 3
microseconds us2(2);           // us2 stores 2
microseconds us3 = us1 + us2;  // us3 stores 5

microseconds us4 = m3 + us3;   // us4 stores 300000005

In the final line of code above, there is an implicit conversion from minutes to microseconds, resulting in a relatively large number of microseconds.

If you need to access the tick count within a duration, there is a member count() which simply returns the stored tick count.

long long tc = us4.count();    // tc is 300000005

These duration's have very simple, very predictable, and very observable behavior. After all, this is really nothing but the time-tested interface of Jeff's boost time duration library (unified with templates instead of inheritance).

minutes m4 = m3 + us3;

It won't compile! The rationale is that implicit truncation error should not be allowed to happen. If this were to compile, then m4 would hold 5, the same value as m3. The value associated with us3 has been effectively ignored. This is similar to the problem of assigning a double to an int: the fractional part gets silently discarded.

There is a duration_cast facility to explicitly ask for this behavior:

minutes m4 = boost::chrono::duration_cast<minutes>(m3 + us3);  // m4.count() == 5

In general, one can perform duration arithmetic at will. If duration_cast isn't used, and it compiles, the arithmetic is exact. If one wants to override this exact arithmetic behavior, duration_cast can be used to explicitly specify that desire. The duration_cast has the same efficiency as the implicit conversion, and will even be exact as often as it can.

You can use duration_cast<> to convert the duration into whatever units you desire. This facility will round down (truncate) if an exact conversion is not possible. For example:

boost::chrono::nanoseconds start;
boost::chrono::nanoseconds end;
typedef boost::chrono::milliseconds ms;
ms d = boost::chrono::duration_cast<ms>(end - start);

// d now holds the number of milliseconds from start to end.

std::cout << ms.count() << "ms\n";

We can convert to nanoseconds, or some integral-based duration which nanoseconds will always exactly convert to, then duration_cast<> is unnecessary:

typedef boost::chrono::nanoseconds ns;
ns d = end - start;
std::cout << ns.count() << "ns\n";

If you need seconds with a floating-point representation you can also eliminate the duration_cast<>:

typedef boost::chrono::duration<double> sec;  // seconds, stored with a double
sec d = end - start;
std::cout << sec.count() << "s\n";

If you're not sure if you need duration_cast<> or not, feel free to try it without. If the conversion is exact, or if the destination has a floating-point representation, it will compile: else it will not compile.

If you need to use duration_cast<>, but want to round up, instead of down when the conversion is inexact, here is a handy little helper function to do so. Writing it is actually a good starter project for understanding Boost.Chrono:

template <class ToDuration, class Rep, class Period>
ToDuration
round_up(boost::chrono::duration<Rep, Period> d)
{
    // first round down
    ToDuration result = boost::chrono::duration_cast<ToDuration>(d);
    if (result < d)  // comparisons are *always* exact
       ++result;     // increment by one tick period
    return result;
}

typedef boost::chrono::milliseconds ms;
ms d = round_up<ms>(end - start);
// d now holds the number of milliseconds from start to end, rounded up.
std::cout << ms.count() << "ms\n";

Boost.Chrono provides few simple rounding utility functions for working with durations.

// round down
template <class To, class Rep, class Period>
To
floor(const duration<Rep, Period>& d)
{
    return duration_cast<To>(d);
}

// round to nearest, to even on tie
template <class To, class Rep, class Period>
To
round(const duration<Rep, Period>& d)
{
    To t0 = duration_cast<To>(d);
    To t1 = t0;
    ++t1;
    BOOST_AUTO(diff0, d - t0);
    BOOST_AUTO(diff1, t1 - d);
    if (diff0 == diff1)
    {
        if (t0.count() & 1)
            return t1;
        return t0;
    }
    else if (diff0 < diff1)
        return t0;
    return t1;
}
// round up
template <class To, class Rep, class Period>
To
ceil(const duration<Rep, Period>& d)
{
    To t = duration_cast<To>(d);
    if (t < d)
        ++t;
    return t;
}

The beauty of the chrono library is the ease and accuracy with which such conversions can be made. For example to convert from milliseconds (1/1000 of a second), to 1/30 of a second, one must multiply the milliseconds by 0.03. It is common knowledge that you can't exactly represent 0.03 in a computer. Nevertheless round will exactly (with no round off error) detect a tie and round to even when this happens. The differences diff0 and diff1 are not approximate, but exact differences, even when d has the units of millisecond and To is 1/30 of a second. The unit of diff0 and diff1 is 1/3000 of a second which both millisecond and 1/30 of a second exactly convert to (with no truncation error).

Similarly, the comparison t < d in ceil is exact, even when there is no exact conversion between t and d. Example use of rounding functions

#include <iostream>
#include <boost/chrono/chrono_io.hpp>
#include <boost/chrono/floor.hpp>
#include <boost/chrono/round.hpp>
#include <boost/chrono/ceil.hpp>

int main()
{
    using namespace boost::chrono;
    milliseconds ms(2500);
    std::cout << floor<seconds>(ms) << '\n';
    std::cout << round<seconds>(ms) << '\n';
    std::cout << ceil<seconds>(ms) << '\n';
    ms = milliseconds(2516);
    typedef duration<long, boost::ratio<1, 30> > frame_rate;
    std::cout << floor<frame_rate>(ms) << '\n';
    std::cout << round<frame_rate>(ms) << '\n';
    std::cout << ceil<frame_rate>(ms) << '\n';

    return 0;
}

The output of this program should be

2 seconds
2 seconds
3 seconds
75 [1/30]seconds
75 [1/30]seconds
76 [1/30]seconds

I don't want to deal with writing duration_cast all over the place. I'm content with the precision of my floating-point representation.

Not a problem. When the destination of a conversion has floating-point representation, all conversions are allowed to happen implicitly.

typedef boost::chrono::duration<double, ratio<60> > dminutes;
dminutes dm4 = m3 + us3;  // dm4.count() == 5.000000083333333

If you were writing these conversions by hand, you could not make it more efficient. The use of ratio ensures that all conversion constants are simplified as much as possible at compile-time. This usually results in the numerator or denominator of the conversion factor simplifying to 1, and being subsequently ignored in converting the run-time values of the tick counts.

There are several options open to the user:

void f(boost::chrono::duration<double> d)  // accept floating-point seconds
{
    // d.count() == 3.e-6 when passed boost::chrono::microseconds(3)
}

f(boost::chrono::microseconds(3));

void f(boost::chrono::nanoseconds d)
{
    // d.count() == 3000 when passed boost::chrono::microseconds(3)
}

f(boost::chrono::microseconds(3));

In this design, if the client wants to pass in a floating-point duration, or a duration of finer precision than nanoseconds, then the client is responsible for choosing his own rounding mode in the conversion to nanoseconds.

boost::chrono::duration<double> s(1./3);  // 1/3 of a second
f(boost::chrono::duration_cast<boost::chrono::nanoseconds>(s));  // round towards zero in conversion to nanoseconds

In the example above, the client of f has chosen "round towards zero" as the desired rounding mode to nanoseconds. If the client has a duration that won't exactly convert to nanoseconds, and fails to choose how the conversion will take place, the compiler will refuse the call:

f(s);  // does not compile

template <class Rep, class Period>
void f(boost::chrono::duration<Rep, Period> d)
{
    // convert d to nanoseconds, rounding up if it is not an exact conversion
    boost::chrono::nanoseconds ns = boost::chrono::duration_cast<boost::chrono::nanoseconds>(d);
    if (ns < d)
        ++ns;
    // ns.count() == 333333334 when passed 1/3 of a floating-point second
}

  f(boost::chrono::duration<double>(1./3));

template <class Period>
void f(boost::chrono::duration<long long, Period> d)
{
    // convert d to nanoseconds, rounding up if it is not an exact conversion
    boost::chrono::nanoseconds ns = boost::chrono::duration_cast<nanoseconds>(d);
    if (ns < d)
        ++ns;
    // ns.count() == 333333334 when passed 333333333333 picoseconds
}
// About 1/3 of a second worth of picoseconds
f(boost::chrono::duration<long long, boost::pico>(333333333333));

Clients with floating-point durations who want to use f will now have to convert to an integral duration themselves before passing the result to f.

In summary, the author of f has quite a bit of flexibility and control in the interface he wants to provide his clients with, and easy options for manipulating that duration internal to his function.

No. No matter which option the author of f chooses above, the following client code will not compile:

f(3);  // Will not compile, 3 is not implicitly convertible to any __duration

This depend on the representation. The default typedefs uses a representation that don't handle overflows. The user can define his own representation that manage overflow as required by its application.

A time_point has a clock and a duration.

template <class Clock, class Duration = typename Clock::duration> class time_point;

The time_point's clock is not stored. It is simply embedded into the time_point's type and serves two purposes:

A time_point's duration is stored as the only data member of the time_point. Thus time_points and their corresponding duration have exactly the same layout. But they have very different meanings. For example, it is one thing to say I want to sleep for 3 minutes. It is a completely different thing to say I want to sleep until 3 minutes past the time I started that timer (unless you just happened to start that timer now). Both meanings (and options for sleeping) have great practical value in common use cases for sleeping, waiting on a condition variable, and waiting for a mutex's lock. These same concepts and tools are found (for example) in Ada.

A timer example:

void f()
{
    boost::chrono::steady_clock::time_point start = boost::chrono::steady_clock::now();
    g();
    h();
    duration<double> sec = boost::chrono::steady_clock::now() - start;
    cout << "f() took " << sec.count() << " seconds\n";
}

Note that if one is using the duration between two clock time_points in a way where the precision of the duration matters, it is good practice to convert the clock's duration to a known duration. This insulates the code from future changes which may be made to the clock's precision in the future. For example steady_clock could easily be based on the clock speed of the cpu. When you upgrade to a faster machine, you do not want your code that assumed a certain tick period of this clock to start experiencing run-time failures because your timing code has silently changed meaning.

A delay loop example:

// delay for at least 500 nanoseconds:
auto go = boost::chrono::steady_clock::now() + boost::chrono::nanoseconds(500);
while (boost::chrono::steady_clock::now() < go)
    ;

The above code will delay as close as possible to half a microsecond, no matter what the precision of steady_clock is. The more precise steady_clock becomes, the more accurate will be the delay to 500 nanoseconds.

system_clock is useful when you need to correlate the time with a known epoch so you can convert it to a calendar time. Note the specific functions in the system_clock class.

steady_clock is useful when you need to wait for a specific amount of time. steady_clock time can not be reset. As other steady clocks, it is usually based on the processor tick.

Here is a polling solution, but it will probably be too inefficient:

boost::chrono::steady_clock::time_point start= chrono::steady_clock::now();
boost::chrono::steady_clock::duration delay= chrono::seconds(5);
while (boost::chrono::steady_clock::now() - start <= delay) {}

When available, high_resolution_clock is usually more expensive than the other system-wide clocks, so they are used only when the provided resolution is required to the application.

Process and thread clocks are used usually to measure the time spent by code blocks, as a basic time-spent profiling of different blocks of code (Boost.Chrono.Stopwatch is a clear example of this use).

You can use thread_clock whenever you want to measure the time spent by the current thread. For example:

boost::chrono::thread_clock::time_point start=boost::chrono::thread_clock::now();
// ... do something ...

typedef boost::chrono::milliseconds ms;
ms d = boost::chrono::thread_clock::now() - start;
// d now holds the number of milliseconds from start to end.
std::cout << ms.count() << "ms\n";

If you need seconds with a floating-point representation you can do:

typedef boost::chrono::duration<double> sec;  // seconds, stored with a double.
sec d = end - start;
std::cout << sec.count() << "s\n";

If you would like to programmatically inspect thread_clock::duration, you can get the representation type with thread_clock::rep, and the tick period with thread_clock::period (which should be a type ratio which has nested values ratio::num and ratio::den). The tick period of thread_clock is thread_clock::period::num / thread_clock::period::den seconds: 1/1000000000 in this case (1 billionth of a second), stored in a long long.

Any duration can be streamed out to a basic_ostream. The run-time value of the duration is formatted according to the rules and current format settings for duration::rep get_duration_style and the durationpunct facet.

the format is either

<value> <unit>

or

<unit> <value>

Warning

Need to be changed This is followed by a single space and then the compile-time unit name of the duration. This unit name is built on the string returned from ratio_string<> and the data used to construct the duration_punct which was inserted into the stream's locale. If a duration_punct has not been inserted into the stream's locale, a default constructed duration_punct will be added to the stream's locale. duration unit names come in two varieties: long(prefix) and short(symbol). The default constructed duration_punct provides names in the long(prefix) format. These names are English descriptions. Other languages are supported by constructing a duration_punct with the proper spellings for "hours", "minutes" and "seconds", and their abbreviations (for the short format). The short or long format can be easily chosen by streaming a duration_short() or duration_long() manipulator respectively or using the parameterized manipulator duration_fmt(duration_style::prefix) or duration_fmt(duration_style::symbol).

Example:

#include <iostream>
#include <boost/chrono/chrono_io.hpp>

int main()
{
    using namespace std;
    using namespace boost;

    cout << "milliseconds(1) = "
         <<  boost::chrono::milliseconds(1) << '\n';

    cout << "milliseconds(3) + microseconds(10) = "
         <<  boost::chrono::milliseconds(3) + boost::chrono::microseconds(10) << '\n';

    cout << "hours(3) + minutes(10) = "
         <<  boost::chrono::hours(3) + boost::chrono::minutes(10) << '\n';

    typedef boost::chrono::duration<long long, boost::ratio<1, 2500000000> > ClockTick;
    cout << "ClockTick(3) + boost::chrono::nanoseconds(10) = "
         <<  ClockTick(3) + boost::chrono::nanoseconds(10) << '\n';

   // ...
    return 0;
}

The output could be

milliseconds(1) = 1 microsecond
milliseconds(3) + microseconds(10) = 3010 microseconds
hours(3) + minutes(10) = 190 minutes
ClockTick(3) + nanoseconds(10) = 56 [1/5000000000]seconds

Set cout to use short names:
milliseconds(3) + microseconds(10) = 3010 μs
hours(3) + minutes(10) = 190 m
ClockTick(3) + nanoseconds(10) = 56 [1/5000000000]s

system_clock::now() = 129387415616250000 [1/10000000]s since Jan 1, 1970
monotonic_clock::now() = 37297387636417 ns since boot

Set cout to use long names:
high_resolution_clock::now() = 37297387655134 nanoseconds since boot

As can be seen, each duration type can be streamed without having to manually stream the compile-time units after the run-time value. And when the compile-time unit is known to be a "common unit", English names are used. For "uncommon units" a unit name is composed from the reduced numerator and denominator of the associated ratio. Whatever stream/locale settings are set for duration::rep are used for the value. Additionally, when the value is 1, singular forms for the units are used.

Sometimes it is desired to shorten these names by using the SI symbols instead of SI prefixes. This can be accomplished with the use of the symbol_format manipulator ^[1]:

cout << "\nSet cout to use short names:\n";
cout << boost::chrono::symbol_format;

cout << "milliseconds(3) + microseconds(10) = "
     <<  boost::chrono::milliseconds(3) + boost::chrono::microseconds(10) << '\n';

cout << "hours(3) + minutes(10) = "
     <<  boost::chrono::hours(3) + boost::chrono::minutes(10) << '\n';

cout << "ClockTick(3) + nanoseconds(10) = "
     <<  ClockTick(3) + boost::chrono::nanoseconds(10) << '\n';

The output could be

Set cout to use short names:
milliseconds(3) + microseconds(10) = 3010 μs
hours(3) + minutes(10) = 190 m
ClockTick(3) + nanoseconds(10) = 56 [1/5000000000]s

system_clock::now() = 129387415616250000 [1/10000000]s since Jan 1, 1970
monotonic_clock::now() = 37297387636417 ns since boot

Set cout to use long names:
high_resolution_clock::now() = 37297387655134 nanoseconds since boot

The μ for microsecond is specified to be U+00B5, encoded as UTF-8, UTF-16 or UTF-32 as appropriate for the stream's character size.

When the format decision is taken at runtime, it could be better to use the parameterized manipulator duration_fmt as in

duration_style style;
//...
cout << duration_fmt(style);

Parsing a duration follows rules analogous to the duration converting constructor. A value and a unit (SI symbol or prefixed) are read from the basic_istream. If the duration has an integral representation, then the value parsed must be exactly representable in the target duration (after conversion to the target duration units), else failbit is set. durations based on floating-point representations can be parsed using any units that do not cause overflow.

For example a stream containing "5000 milliseconds" can be parsed into seconds, but if the stream contains "3001 ms", parsing into seconds will cause failbit to be set.

Example:

#include <boost/chrono/chrono_io.hpp>
#include <sstream>
#include <cassert>

int main()
{
    using namespace std;

    istringstream in("5000 milliseconds 4000 ms 3001 ms");
    boost::chrono::seconds d(0);
    in >> d;
    assert(in.good());
    assert(d == seconds(5));
    in >> d;
    assert(in.good());
    assert(d == seconds(4));
    in >> d;
    assert(in.fail());
    assert(d == seconds(4));

    return 0;
}

Note that a duration failure may occur late in the parsing process. This means that the characters making up the failed parse in the stream are usually consumed despite the failure to successfully parse.

Sometimes in templated code it is difficult to know what the unit of your duration is. It is all deterministic, and inspect-able. But it can be inconvenient to do so, especially if you just need to print out a "debugging" statement. For example:

// round to nearest, to even on tie
template <class To, class Rep, class Period>
To
round(const duration<Rep, Period>& d)
{
    To t0 = duration_cast<To>(d);
    To t1 = t0;
    ++t1;
    auto diff0 = d - t0;
    cout << "diff0 = " << diff0 << '\n';
    auto diff1 = t1 - d;
    cout << "diff1 = " << diff1 << '\n';
    if (diff0 == diff1)
    {
        if (t0.count() & 1)
            return t1;
        return t0;
    }
    else if (diff0 < diff1)
        return t0;
    return t1;
}

This is where I/O for duration really shines. The compiler knows what the type of diff0 is and with this proposal that type (with proper units) will automatically be printed out for you. For example:

milliseconds ms = round<milliseconds>(nanoseconds(123));  // diff0 = 123 nanoseconds
                                                          // diff1 = 999877 nanoseconds
milliseconds ms = round<milliseconds>(Ticks(44));         // diff0 = 2 [1/3000]seconds
                                                          // diff1 = 1 [1/3000]second

This simple I/O will make duration so much more accessible to programmers.

system_clock is special. It is the only clock that has conversions between its time_point and time_t. C subsequently relates time_t to the Gregorian calendar via ctime, gmtime, localtime, and strftime. Neither C, nor POSIX relate time_t to any calendar other than the Gregorian calendar. ISO 8601 is specified only in terms of the Gregorian calendar.

Boost.Chrono provides system_clock::time_point I/O in terms of the Gregorian calendar, and no other calendar. However as system_clock::time_point remains convertible with time_t, it is possible for clients to create other calendars which interoperate with time_t and subsequently system_clock::time_point.

Furthermore, it is existing practice for all major hosted operating systems to store system time in a format which facilitates display as Coordinated Universal Time (UTC). Therefore Boost.Chrono provides that the default output for system_clock::time_point be in a format that represents a point in time with respect to UTC.

cout << system_clock::now() << '\n';

could output

2011-09-15 18:36:59.325132 +0000

This format is strongly influenced by ISO 8601, but places a ' ' between the date and time instead of a 'T'. The former appears to more accurately represent existing practice. A fully numeric format was chosen so as to be understandable to as large a group of human readers as possible. A 24 hour format was chosen for the same reasons.

Of the referenced standards, only ISO 8601 discusses the output of fractional seconds. Neither C nor POSIX have built-in functionality for this. However it appears to be universal (as of this writing) that system_clock::period is sub-second. And it seems desirable that if you stream out a system_clock::time_point, you ought to be able to stream it back in and get the same value. Therefore the streaming of fractional seconds (at least by default) appears to be unavoidable.

Finally the trailing " +0000" disambiguates the UTC-formatted system_clock::time_point from one formatted with respect to the local time zone of the computer. The latter can easily be achieved with:

cout << time_fmt(local) << system_clock::now() << '\n';

that could result in

2011-09-15 14:36:59.325132 -0400

Note that system_clock::time_point itself is neither UTC, nor the local time. However in practice, system_clock::time_point is a count of ticks beyond some epoch which is synchronized with UTC. So as a mobile computer moves across time zones, the time zone traversal does not impact the value of a system_clock::time_point produced by system_clock::now(). And it is only in formatting it for human consumption that one can choose UTC or the local time zone. C and POSIX treat time_t just as Boost.Chrono treats system_clock::time_point:

tm* gmtime(const time_t* timer) -> UTC
tm* localtime(const time_t* timer) -> local time

This proposal simply extends the C/POSIX time_t functionality to C++ syntax and system_clock::time_point.

The time_fmt() manipulator is "sticky". It will remain in effect until the stream destructs or until it is changed. The stream can be reset to its default state with:

cout << time_fmt(utc);

And the formatting can be further customized by using the time format sequences. For example:

cout << time_fmt(local, "%A %B %e, %Y %r");
cout << system_clock::now() << '\n';  // Sunday April 24, 2011 02:36:59 PM

When specifying formatting manipulators for wide streams, use wide strings.

You can use the same manipulators with istreams to specify parsing sequences.

Unfortunately there are no formatting/parsing sequences which indicate fractional seconds. Boost.Chrono does not provide such sequences. In the meantime, one can format and parse fractional seconds for system_clock::time_point by defaulting the format, or by using an empty string in time_fmt().

The stream's current locale may impact the parsing/format sequences supplied to the system_clock::time_point manipulators (e.g. names of days of the week, and names of months).

Unlike system_clock::time_point, the other clocks have no conversion with time_t. There is likely no relationship between steady_clock::time_point and UTC at all (UTC is not steady).

In general a time_point is formatted by outputting its internal duration followed by a string that describes the time_point::clock epoch. This string will vary for each distinct clock, and for each implementation of the supplied clocks.

#ifdef BOOST_CHRONO_HAS_CLOCK_STEADY
    cout << "steady_clock::now() = " << boost::chrono::steady_clock::now() << '\n';
#endif
    cout << "\nSet cout to use long names:\n"
            << boost::chrono::duration_long
            << "high_resolution_clock::now() = "
            << boost::chrono::high_resolution_clock::now() << '\n';

The output could be

steady_clock::now() = 37297387636417 ns since boot

Set cout to use long names:
high_resolution_clock::now() = 37297387655134 nanoseconds since boot

Parsing a time_point involves first parsing a duration and then parsing the epoch string. If the epoch string does not match that associated with time_point::clock then failbit will be set.

Example:

#include <boost/chrono/chrono_io.hpp>
#include <sstream>
#include <iostream>
#include <cassert>

int main()
{
    using namespace std;

    boost::chrono::high_resolution_clock::time_point t0 = boost::chrono::high_resolution_clock::now();
    stringstream io;
    io << t0;
    boost::chrono::high_resolution_clock::time_point t1;
    io >> t1;
    assert(!io.fail());
    cout << io.str() << '\n';
    cout << t0 << '\n';
    cout << t1 << '\n';
    boost::chrono::high_resolution_clock::time_point t = boost::chrono::high_resolution_clock::now();
    cout << t << '\n';

    cout << "That took " << t - t0 << '\n';
    cout << "That took " << t - t1 << '\n';

    return 0;
}

The output could be:

50908679121461 nanoseconds since boot
That took 649630 nanoseconds

Here's a simple example to find out how many hours the computer has been up (on this platform):

#include <boost/chrono/chrono_io.hpp>
#include <iostream>

int main()
{
    using namespace std;
    using namespace boost;

    typedef boost::chrono::time_point<boost::chrono::steady_clock, boost::chrono::duration<double, boost::ratio<3600> > > T;
    T tp = boost::chrono::steady_clock::now();
    std::cout << tp << '\n';
    return 0;
}

The output could be:

17.8666 hours since boot

The I/O interface described in the preceding I/O sections were at the user level. These services are based on low level services that are useful when writing libraries. The low level services are related to access to the associated ios state and locale facets. The design follows the C++ IOStreams standard design:

The library encapsulate the locale-dependent parsing and formatting of duration into a new facet class. Let's focus on formatting in this example. The concerned facet class is duration_put, analogous to time_put, money_put, etc.

The use of this facet is similar to the time_put facet.

Next we show how to override the duration's default constructor to do anything you want (in this case set it to zero). All we need to do is to change the representation

namespace I_dont_like_the_default_duration_behavior {

template <class R>
class zero_default
{
public:
    typedef R rep;

private:
    rep rep_;
public:
    zero_default(rep i = 0) : rep_(i) {}
    operator rep() const {return rep_;}

    zero_default& operator+=(zero_default x) {rep_ += x.rep_; return *this;}
    zero_default& operator-=(zero_default x) {rep_ -= x.rep_; return *this;}
    zero_default& operator*=(zero_default x) {rep_ *= x.rep_; return *this;}
    zero_default& operator/=(zero_default x) {rep_ /= x.rep_; return *this;}

    zero_default  operator+ () const {return *this;}
    zero_default  operator- () const {return zero_default(-rep_);}
    zero_default& operator++()       {++rep_; return *this;}
    zero_default  operator++(int)    {return zero_default(rep_++);}
    zero_default& operator--()       {--rep_; return *this;}
    zero_default  operator--(int)    {return zero_default(rep_--);}

    friend zero_default operator+(zero_default x, zero_default y) {return x += y;}
    friend zero_default operator-(zero_default x, zero_default y) {return x -= y;}
    friend zero_default operator*(zero_default x, zero_default y) {return x *= y;}
    friend zero_default operator/(zero_default x, zero_default y) {return x /= y;}

    friend bool operator==(zero_default x, zero_default y) {return x.rep_ == y.rep_;}
    friend bool operator!=(zero_default x, zero_default y) {return !(x == y);}
    friend bool operator< (zero_default x, zero_default y) {return x.rep_ < y.rep_;}
    friend bool operator<=(zero_default x, zero_default y) {return !(y < x);}
    friend bool operator> (zero_default x, zero_default y) {return y < x;}
    friend bool operator>=(zero_default x, zero_default y) {return !(x < y);}
};

typedef boost::chrono::duration<zero_default<long long>, boost::nano        > nanoseconds;
typedef boost::chrono::duration<zero_default<long long>, boost::micro       > microseconds;
typedef boost::chrono::duration<zero_default<long long>, boost::milli       > milliseconds;
typedef boost::chrono::duration<zero_default<long long>                      > seconds;
typedef boost::chrono::duration<zero_default<long long>, boost::ratio<60>   > minutes;
typedef boost::chrono::duration<zero_default<long long>, boost::ratio<3600> > hours;
}

Usage

using namespace I_dont_like_the_default_duration_behavior;

milliseconds ms;
std::cout << ms.count() << '\n';

See the source file example/i_dont_like_the_default_duration_behavior.cpp

A "saturating" signed integral type is developed. This type has +/- infinity and a NaN (like IEEE floating-point) but otherwise obeys signed integral arithmetic. This class is subsequently used as the template parameter Rep in boost::chrono::duration to demonstrate a duration class that does not silently ignore overflow.

See the source file example/saturating.cpp

Example round_up utility: converts d to To, rounding up for inexact conversions Being able to easily write this function is a major feature!

#include <boost/chrono.hpp>
#include <boost/type_traits.hpp>

#include <iostream>

template <class To, class Rep, class Period>
To
round_up(boost::chrono::duration<Rep, Period> d)
{
    To result = boost::chrono::duration_cast<To>(d);
    if (result < d)
        ++result;
    return result;
}

To demonstrate interaction with an xtime-like facility:

struct xtime
{
    long sec;
    unsigned long usec;
};

template <class Rep, class Period>
xtime
to_xtime_truncate(boost::chrono::duration<Rep, Period> d)
{
    xtime xt;
    xt.sec = static_cast<long>(boost::chrono::duration_cast<seconds>(d).count());
    xt.usec = static_cast<long>(boost::chrono::duration_cast<microseconds>(d - seconds(xt.sec)).count());
    return xt;
}

template <class Rep, class Period>
xtime
to_xtime_round_up(boost::chrono::duration<Rep, Period> d)
{
    xtime xt;
    xt.sec = static_cast<long>(boost::chrono::duration_cast<seconds>(d).count());
    xt.usec = static_cast<unsigned long>(round_up<boost::chrono::microseconds>(d - boost::chrono::seconds(xt.sec)).count());
    return xt;
}

microseconds
from_xtime(xtime xt)
{
    return boost::chrono::seconds(xt.sec) + boost::chrono::microseconds(xt.usec);
}

void print(xtime xt)
{
    std::cout << '{' << xt.sec << ',' << xt.usec << "}\n";
}

Usage

xtime xt = to_xtime_truncate(seconds(3) + boost::chrono::milliseconds(251));
print(xt);
boost::chrono::milliseconds ms = boost::chrono::duration_cast<boost::chrono::milliseconds>(from_xtime(xt));
std::cout << ms.count() << " milliseconds\n";
xt = to_xtime_round_up(ms);
print(xt);
xt = to_xtime_truncate(boost::chrono::seconds(3) + nanoseconds(999));
print(xt);
xt = to_xtime_round_up(boost::chrono::seconds(3) + nanoseconds(999));
print(xt);

See the source file xtime.cpp

Users can easily create their own clocks, with both points in time and time durations which have a representation and precision of their own choosing. For example if there is a hardware counter which simply increments a count with each cycle of the cpu, one can very easily build clocks, time points and durations on top of that, using only a few tens of lines of code. Such systems can be used to call the time-sensitive threading API's such as sleep, wait on a condition variable, or wait for a mutex lock. The API proposed herein is not sensitive as to whether this is a 300MHz clock (with a 3 1/3 nanosecond tick period) or a 3GHz clock (with a tick period of 1/3 of a nanosecond). And the resulting code will be just as efficient as if the user wrote a special purpose clock cycle counter.

#include <boost/chrono.hpp>
#include <boost/type_traits.hpp>
#include <iostream>

template <long long speed>
struct cycle_count
{
    typedef typename boost::__ratio_multiply__<boost::ratio<speed>, boost::mega>::type
        frequency;  // Mhz
    typedef typename boost::__ratio_divide__<boost::ratio<1>, frequency>::type period;
    typedef long long rep;
    typedef boost::chrono::duration<rep, period> duration;
    typedef boost::chrono::time_point<cycle_count> time_point;

    static time_point now()
    {
        static long long tick = 0;
        // return exact cycle count
        return time_point(duration(++tick));  // fake access to clock cycle count
    }
};

template <long long speed>
struct approx_cycle_count
{
    static const long long frequency = speed * 1000000;  // MHz
    typedef nanoseconds duration;
    typedef duration::rep rep;
    typedef duration::period period;
    static const long long nanosec_per_sec = period::den;
    typedef boost::chrono::time_point<approx_cycle_count> time_point;

    static time_point now()
    {
        static long long tick = 0;
        // return cycle count as an approximate number of nanoseconds
        // compute as if nanoseconds is only duration in the std::lib
        return time_point(duration(++tick * nanosec_per_sec / frequency));
    }
};

See the source file cycle_count.cpp

This example demonstrates the use of a timeval-like struct to be used as the representation type for both duration and time_point.

class xtime {
private:
    long tv_sec;
    long tv_usec;

    void fixup() {
        if (tv_usec < 0) {
            tv_usec += 1000000;
            --tv_sec;
        }
    }

public:
    explicit xtime(long sec, long usec) {
        tv_sec = sec;
        tv_usec = usec;
        if (tv_usec < 0 || tv_usec >= 1000000) {
            tv_sec += tv_usec / 1000000;
            tv_usec %= 1000000;
            fixup();
        }
    }

    explicit xtime(long long usec) {
        tv_usec = static_cast<long>(usec % 1000000);
        tv_sec  = static_cast<long>(usec / 1000000);
        fixup();
    }

    // explicit
    operator long long() const {return static_cast<long long>(tv_sec) * 1000000 + tv_usec;}

    xtime& operator += (xtime rhs) {
        tv_sec += rhs.tv_sec;
        tv_usec += rhs.tv_usec;
        if (tv_usec >= 1000000) {
            tv_usec -= 1000000;
            ++tv_sec;
        }
        return *this;
    }

    xtime& operator -= (xtime rhs) {
        tv_sec -= rhs.tv_sec;
        tv_usec -= rhs.tv_usec;
        fixup();
        return *this;
    }

    xtime& operator %= (xtime rhs) {
        long long t = tv_sec * 1000000 + tv_usec;
        long long r = rhs.tv_sec * 1000000 + rhs.tv_usec;
        t %= r;
        tv_sec = static_cast<long>(t / 1000000);
        tv_usec = static_cast<long>(t % 1000000);
        fixup();
        return *this;
    }

    friend xtime operator+(xtime x, xtime y) {return x += y;}
    friend xtime operator-(xtime x, xtime y) {return x -= y;}
    friend xtime operator%(xtime x, xtime y) {return x %= y;}

    friend bool operator==(xtime x, xtime y)
        { return (x.tv_sec == y.tv_sec && x.tv_usec == y.tv_usec); }

    friend bool operator<(xtime x, xtime y) {
        if (x.tv_sec == y.tv_sec)
            return (x.tv_usec < y.tv_usec);
        return (x.tv_sec < y.tv_sec);
    }

    friend bool operator!=(xtime x, xtime y) { return !(x == y); }
    friend bool operator> (xtime x, xtime y) { return y < x; }
    friend bool operator<=(xtime x, xtime y) { return !(y < x); }
    friend bool operator>=(xtime x, xtime y) { return !(x < y); }

    friend std::ostream& operator<<(std::ostream& os, xtime x)
        {return os << '{' << x.tv_sec << ',' << x.tv_usec << '}';}
};

Clock based on timeval-like struct.

class xtime_clock
{
public:
    typedef xtime                                  rep;
    typedef boost::micro                           period;
    typedef boost::chrono::duration<rep, period>   duration;
    typedef boost::chrono::time_point<xtime_clock> time_point;

    static time_point now()
    {
    #if defined(BOOST_CHRONO_WINDOWS_API)
        time_point t(duration(xtime(0)));
        gettimeofday((timeval*)&t, 0);
        return t;

    #elif defined(BOOST_CHRONO_MAC_API)

        time_point t(duration(xtime(0)));
        gettimeofday((timeval*)&t, 0);
        return t;

    #elif defined(BOOST_CHRONO_POSIX_API)
        //time_point t(0,0);

        timespec ts;
        ::clock_gettime( CLOCK_REALTIME, &ts );

        xtime xt( ts.tv_sec, ts.tv_nsec/1000);
        return time_point(duration(xt));

    #endif  // POSIX
    }
};

Usage of xtime_clock

std::cout << "sizeof xtime_clock::time_point = " << sizeof(xtime_clock::time_point) << '\n';
std::cout << "sizeof xtime_clock::duration = " << sizeof(xtime_clock::duration) << '\n';
std::cout << "sizeof xtime_clock::rep = " << sizeof(xtime_clock::rep) << '\n';
xtime_clock::duration delay(boost::chrono::milliseconds(5));
xtime_clock::time_point start = xtime_clock::now();
while (xtime_clock::now() - start <= delay) {}
xtime_clock::time_point stop = xtime_clock::now();
xtime_clock::duration elapsed = stop - start;
std::cout << "paused " << boost::chrono::::nanoseconds(elapsed).count() << " nanoseconds\n";

See the source file example/timeval_demo.cpp

The user can define a function returning the earliest time_point as follows:

template <class Clock, class Duration1, class Duration2>
typename boost::common_type<time_point<Clock, Duration1>,
                     time_point<Clock, Duration2> >::type
min(time_point<Clock, Duration1> t1, time_point<Clock, Duration2> t2)
{
    return t2 < t1 ? t2 : t1;
}

Being able to easily write this function is a major feature!

BOOST_AUTO(t1, system_clock::now() + seconds(3));
BOOST_AUTO(t2, system_clock::now() + nanoseconds(3));
BOOST_AUTO(t3, min(t1, t2));

See the source file example/min_time_point.cpp

#include <boost/chrono.hpp>
#include <iostream>
#include <iomanip>

using namespace boost::chrono;

template< class Clock >
class timer
{
  typename Clock::time_point start;
public:
  timer() : start( Clock::now() ) {}
  typename Clock::duration elapsed() const
  {
    return Clock::now() - start;
  }
  double seconds() const
  {
    return elapsed().count() * ((double)Clock::period::num/Clock::period::den);
  }
};

int main()
{
  timer<system_clock> t1;
  timer<steady_clock> t2;
  timer<high_resolution_clock> t3;

  std::cout << "Type the Enter key: ";
  std::cin.get();

  std::cout << std::fixed << std::setprecision(9);
  std::cout << "system_clock-----------: "
            << t1.seconds() << " seconds\n";
  std::cout << "steady_clock--------: "
            << t2.seconds() << " seconds\n";
  std::cout << "high_resolution_clock--: "
            << t3.seconds() << " seconds\n";

  system_clock::time_point d4 = system_clock::now();
  system_clock::time_point d5 = system_clock::now();

  std::cout << "\nsystem_clock latency-----------: " << (d5 - d4).count() << std::endl;

  steady_clock::time_point d6 = steady_clock::now();
  steady_clock::time_point d7 = steady_clock::now();

  std::cout << "steady_clock latency--------: " << (d7 - d6).count() << std::endl;

  high_resolution_clock::time_point d8 = high_resolution_clock::now();
  high_resolution_clock::time_point d9 = high_resolution_clock::now();

  std::cout << "high_resolution_clock latency--: " << (d9 - d8).count() << std::endl;

  std::time_t now = system_clock::to_time_t(system_clock::now());

  std::cout << "\nsystem_clock::now() reports UTC is "
    << std::asctime(std::gmtime(&now)) << "\n";

  return 0;
}

The output of this program run looks like this:

See the source file example/await_keystroke.cpp

In the example above we take advantage of the fact that time_points convert as long as they have the same clock, and as long as their internal durations convert. We also take advantage of the fact that a duration with a floating-point representation will convert from anything. Finally the I/O system discovers the more readable "hours" unit for our duration<double, ratio<3600>>.

There are many other ways to format durations and time_points. For example see ISO 8601. Instead of coding every possibility into operator<<, which would lead to significant code bloat for even the most trivial uses, this document seeks to inform the reader how to write custom I/O when desired.

As an example, the function below streams arbitrary durations to arbitrary basic_ostreams using the format:

[-]d/hh:mm:ss.cc

Where:

// format duration as [-]d/hh::mm::ss.cc
template <class CharT, class Traits, class Rep, class Period>
std::basic_ostream<CharT, Traits>&
display(std::basic_ostream<CharT, Traits>& os,
        boost::chrono::duration<Rep, Period> d)
{
    using namespace std;
    using namespace boost;

    typedef boost::chrono::duration<long long, boost::ratio<86400> > days;
    typedef boost::chrono::duration<long long, boost:centi> centiseconds;

    // if negative, print negative sign and negate
    if (d < boost::chrono::duration<Rep, Period>(0))
    {
        d = -d;
        os << '-';
    }
    // round d to nearest centiseconds, to even on tie
    centiseconds cs = boost::chrono::duration_cast<centiseconds>(d);
    if (d - cs > boost::chrono::milliseconds(5)
        || (d - cs == boost::chrono::milliseconds(5) && cs.count() & 1))
        ++cs;
    // separate seconds from centiseconds
    boost::chrono::seconds s = boost::chrono::duration_cast<boost::chrono::seconds>(cs);
    cs -= s;
    // separate minutes from seconds
    boost::chrono::minutes m = boost::chrono::duration_cast<boost::chrono::minutes>(s);
    s -= m;
    // separate hours from minutes
    boost::chrono::hours h = boost::chrono::duration_cast<boost::chrono::hours>(m);
    m -= h;
    // separate days from hours
    days dy = boost::chrono::duration_cast<days>(h);
    h -= dy;
    // print d/hh:mm:ss.cc
    os << dy.count() << '/';
    if (h < boost::chrono::hours(10))
        os << '0';
    os << h.count() << ':';
    if (m < boost::chrono::minutes(10))
        os << '0';
    os << m.count() << ':';
    if (s < boost::chrono::seconds(10))
        os << '0';
    os << s.count() << '.';
    if (cs < boost::chrono::centiseconds(10))
        os << '0';
    os << cs.count();
    return os;
}

int main()
{
    using namespace std;
    using namespace boost;

    display(cout, boost::chrono::steady_clock::now().time_since_epoch()
                  + boost::chrono::duration<long, boost::mega>(1)) << '\n';
    display(cout, -boost::chrono::milliseconds(6)) << '\n';
    display(cout, boost::chrono::duration<long, boost::mega>(1)) << '\n';
    display(cout, -boost::chrono::duration<long, boost::mega>(1)) << '\n';
}

The output could be:

12/06:03:22.95
-0/00:00:00.01
11/13:46:40.00
-11/13:46:40.00

The C++11 standard library's multi-threading library requires the ability to deal with the representation of time in a manner consistent with modern C++ practices. Next is a simulation of this interface.

The non-member sleep functions can be emulated as follows:

namespace boost { namespace this_thread {

template <class Rep, class Period>
void sleep_for(const chrono::duration<Rep, Period>& d) {
    chrono::microseconds t = chrono::duration_cast<chrono::microseconds>(d);
    if (t < d)
        ++t;
    if (t > chrono::microseconds(0))
        std::cout << "sleep_for " << t.count() << " microseconds\n";
}

template <class Clock, class Duration>
void sleep_until(const chrono::time_point<Clock, Duration>& t) {
    using namespace chrono;
    typedef time_point<Clock, Duration> Time;
    typedef system_clock::time_point SysTime;
    if (t > Clock::now()) {
        typedef typename common_type<typename Time::duration,
                                     typename SysTime::duration>::type D;
        /* auto */ D d = t - Clock::now();
        microseconds us = duration_cast<microseconds>(d);
        if (us < d)
            ++us;
        SysTime st = system_clock::now() + us;
        std::cout << "sleep_until    ";
        detail::print_time(st);
        std::cout << " which is " << (st - system_clock::now()).count() << " microseconds away\n";
    }
}

}}

Next is the boost::thread::timed_mutex modified functions

namespace boost {
struct timed_mutex {
    // ...

    template <class Rep, class Period>
    bool try_lock_for(const chrono::duration<Rep, Period>& d) {
        chrono::microseconds t = chrono::duration_cast<chrono::microseconds>(d);
        if (t <= chrono::microseconds(0))
            return try_lock();
        std::cout << "try_lock_for " << t.count() << " microseconds\n";
        return true;
    }

    template <class Clock, class Duration>
    bool try_lock_until(const chrono::time_point<Clock, Duration>& t)
    {
        using namespace chrono;
        typedef time_point<Clock, Duration> Time;
        typedef system_clock::time_point SysTime;
        if (t <= Clock::now())
            return try_lock();
        typedef typename common_type<typename Time::duration,
          typename Clock::duration>::type D;
        /* auto */ D d = t - Clock::now();
        microseconds us = duration_cast<microseconds>(d);
        SysTime st = system_clock::now() + us;
        std::cout << "try_lock_until ";
        detail::print_time(st);
        std::cout << " which is " << (st - system_clock::now()).count()
          << " microseconds away\n";
        return true;
    }
};
}

boost::thread::condition_variable time related function are modified as follows:

namespace boost {
struct condition_variable
{
    // ...

    template <class Rep, class Period>
    bool wait_for(mutex&, const chrono::duration<Rep, Period>& d) {
        chrono::microseconds t = chrono::duration_cast<chrono::microseconds>(d);
        std::cout << "wait_for " << t.count() << " microseconds\n";
        return true;
    }

    template <class Clock, class Duration>
    bool wait_until(mutex&, const chrono::time_point<Clock, Duration>& t) {
        using namespace boost::chrono;
        typedef time_point<Clock, Duration> Time;
        typedef system_clock::time_point SysTime;
        if (t <= Clock::now())
            return false;
        typedef typename common_type<typename Time::duration,
          typename Clock::duration>::type D;
        /* auto */ D d = t - Clock::now();
        microseconds us = duration_cast<microseconds>(d);
        SysTime st = system_clock::now() + us;
         std::cout << "wait_until     ";
        detail::print_time(st);
        std::cout << " which is " << (st - system_clock::now()).count()
          << " microseconds away\n";
        return true;
    }
};
}

Next follows how simple is the usage of this functions:

boost::mutex m;
boost::timed_mutex mut;
boost::condition_variable cv;

using namespace boost;

this_thread::sleep_for(chrono::seconds(3));
this_thread::sleep_for(chrono::nanoseconds(300));
chrono::system_clock::time_point time_limit = chrono::system_clock::now() + chrono::__seconds_(4) + chrono::milliseconds(500);
this_thread::sleep_until(time_limit);

mut.try_lock_for(chrono::milliseconds(30));
mut.try_lock_until(time_limit);

cv.wait_for(m, chrono::minutes(1));    // real code would put this in a loop
cv.wait_until(m, time_limit);  // real code would put this in a loop

// For those who prefer floating-point
this_thread::sleep_for(chrono::duration<double>(0.25));
this_thread::sleep_until(chrono::system_clock::now() + chrono::duration<double>(1.5));