/* hpTime.c

   Standalone test of the hires timer code from _lsprof4.c.

   Usage: hpTime [<seconds>]

   where <seconds> is a optional sleep time in seconds.  Use
   0 to see the overhead of hpTimer function calls.

   The _lsprof4.c code has been duplicated and slightly
   modified to exercise various builds and use an unsigned
   64-bit type for hires timer values.

   Sample build and output on:

   1. RHEL 3 update 7

     $ gcc -v
     ...
     gcc version 4.1.2
     $ gcc -O3 -o hpTime hpTime.c
     $ hpTime
     test: sleeping for 1 secs (x86) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 2623443435 ticks or 1.003745268 secs
     $ hpTime 4
     test: sleeping for 4 secs (x86) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 10471388967 ticks or 4.006422886 secs
     $ hpTime 0
     test: sleeping for 0 secs (x86) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 19 ticks or 0.000000007 secs

   2. RHEL 3 update 7, 64-bit

     $ gcc -v
     ...
     gcc version 4.1.2
     $ gcc -O3 -m64 -o hpTime hpTime.c
     $ hpTime
     test: sleeping for 1 secs (x86) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 2618620695 ticks or 1.001902958 secs
     $ hpTime 5
     test: sleeping for 5 secs (x86) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 13091450026 ticks or 5.008889191 secs
     [jbrouwers@localhost ~]$ hpTime 0
     test: sleeping for 0 secs (x86) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 21 ticks or 0.000000008 secs

   3. RHEL 3 update 7, rdtscll

     $ gcc -v
     ...
     gcc version 4.1.2
     $ gcc -DRDTSCLL -O3 -o hpTime hpTime.c 
     $ hpTime
     test: sleeping for 1 secs (rdtscll) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 2638513000 ticks or 1.009563887 secs
     $ hpTime 5
     test: sleeping for 5 secs (rdtscll) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 13069212486 ticks or 5.000604024 secs
     $ hpTime 0
     test: sleeping for 0 secs (rdtscll) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 18 ticks or 0.000000007 secs

   4. RHEL 3 update 7, 64-bit rdtscll

     $ gcc -v
     ...
     gcc version 4.1.2
     $ gcc -DRDTSCLL -m64 -O3 -o hpTime hpTime.c
     $ hpTime
     test: sleeping for 1 secs (rdtscll) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 2637057449 ticks or 1.008977073 secs
     $ hpTime 6
     test: sleeping for 6 secs (rdtscll) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 15690711956 ticks or 6.003505371 secs
     $ hpTime 0
     test: sleeping for 0 secs (rdtscll) ...
     resolution 0.383 nsecs/tick (2.614 GHz)
     sleep time 18 ticks or 0.000000007 secs

   5. CentOS 4.5 (Final)

     $ gcc -v
     ...
     gcc version 3.4.6 20060404 (Red Hat 3.4.6-8)
     $ gcc -O3 -o hpTime hpTime.c
     $ hpTime
     test: sleeping for 1 secs (x86) ...
     resolution 0.834 nsecs/tick (1.199 GHz)
     sleep time 1200391272 ticks or 1.001197414 secs
     $ hpTime 5
     test: sleeping for 5 secs (x86) ...
     resolution 0.834 nsecs/tick (1.199 GHz)
     sleep time 5995248528 ticks or 5.000414272 secs
     $ hpTime 0
     test: sleeping for 0 secs (x86) ...
     resolution 0.834 nsecs/tick (1.199 GHz)
     sleep time 128 ticks or 0.000000107 secs

   6. CentOS 4.7 (Final), xVM 2.2.2 VirtualBox

     $ gcc -v
     ...
     gcc version 3.4.6 20060404 (Red Hat 3.4.6-10)
     $ gcc -O3 -o hpTime hpTime.c 
     $ hpTime
     test: sleeping for 1 secs (x86) ...
     resolution 0.545 nsecs/tick (1.836 GHz)
     sleep time 1841023884 ticks or 1.002718521 secs
     $ hpTime 8
     test: sleeping for 8 secs (x86) ...
     resolution 0.543 nsecs/tick (1.840 GHz)
     sleep time 14725436693 ticks or 8.001291819 secs
     $ hpTime 0
     test: sleeping for 0 secs (x86) ...
     resolution 0.545 nsecs/tick (1.834 GHz)
     sleep time 1691 ticks or 0.000000922 secs

   7. MacOS X 10.4.11 Tiger (Core Duo Intel)

     % gcc -v
     ...
     Target: i686-apple-darwin8
     gcc version 4.0.1 (Apple Computer, Inc. build 5367)
     % gcc -O3 -o hpTime hpTime.c
     % hpTime
     test: sleeping for 1 secs (TB) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 1000045727 ticks or 1.000045727 secs
     % hpTime 4
     test: sleeping for 4 secs (TB) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 4000040117 ticks or 4.000040117 secs
     % hpTime 0
     test: sleeping for 0 secs (TB) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 96 ticks or 0.000000096 secs

     % sysctl hw.tbfrequency
     hw.tbfrequency: 1000000000

   8. MacOS X 10.3.9 Panther (PowerPC)

     % gcc -v
     ...
     gcc version 3.3 20030304 (Apple Computer, Inc. build 1671)
     % gcc -O3 -o hpTime hpTime.c
     % hpTime
     test: sleeping for 1 secs (TB) ...
     resolution 40.145 nsecs/tick (0.025 GHz)
     sleep time 24911770 ticks or 1.000085830 secs
     % hpTime 5
     test: sleeping for 5 secs (TB) ...
     resolution 40.145 nsecs/tick (0.025 GHz)
     sleep time 124549562 ticks or 5.000056283 secs
     % hpTime 0
     test: sleeping for 0 secs (TB) ...
     resolution 40.145 nsecs/tick (0.025 GHz)
     sleep time 6 ticks or 0.000000241 secs

     % sysctl hw.tbfrequency
     hw.tbfrequency: 24909600

   9. MacOS X 10.3.9 Panther (PowerPC asm)

     % gcc -v
     ...
     gcc version 3.3 20030304 (Apple Computer, Inc. build 1671)
     % gcc -U__APPLE__ -O5 -o hpTime hpTime.c
     % hpTime
     test: sleeping for 1 secs (ppc) ...
     resolution 40.146 nsecs/tick (0.025 GHz)
     sleep time 24911901 ticks or 1.000122135 secs
     % hpTime 7
     test: sleeping for 7 secs (ppc) ...
     resolution 40.145 nsecs/tick (0.025 GHz)
     sleep time 174368953 ticks or 7.000065822 secs
     % hpTime 0
     test: sleeping for 0 secs (ppc) ...
     resolution 40.144 nsecs/tick (0.025 GHz)
     sleep time 2 ticks or 0.000000161 secs

   10. Solaris 10 (Opteron)

     > cc -V
     cc: Sun C 5.8 2005/10/13
     > cc -DSOLARIS -xO3 -o hpTime hpTime.c
     > hpTime
     test: sleeping for 1 secs (hrtime) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 1007477233 ticks or 1.007477233 secs
     > hpTime 6
     test: sleeping for 6 secs (hrtime) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 6000725228 ticks or 6.000725228 secs
     > hpTime 0
     test: sleeping for 0 secs (hrtime) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 178 ticks or 0.000000178 secs

   11. Solaris 10 (Opteron), 64-bit

     > cc -V
     cc: Sun C 5.8 2005/10/13
     > cc -xarch=amd64 -DSOLARIS -xO3 -o hpTime hpTime.c
     > hpTime
     test: sleeping for 1 secs (hrtime) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 1009813958 ticks or 1.009813958 secs
     > hpTime 6
     test: sleeping for 6 secs (hrtime) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 6001179611 ticks or 6.001179611 secs
     > hpTime 0
     test: sleeping for 0 secs (hrtime) ...
     resolution 1.000 nsecs/tick (1.000 GHz)
     sleep time 171 ticks or 0.000000171 secs
 */

#include <stdio.h>
#include <stdlib.h>

#ifdef MS_WINDOWS
#   define HPTIME   "PC"
#   define hpTime_t  long long  /* ? */

#   include <windows.h>

    static hpTime_t
    hpTimer(void)
    {
        LARGE_INTEGER li;
        QueryPerformanceCounter(&li);
        return li.QuadPart;
    }

    static double
    hpTimerUnit(void)
    {
        LARGE_INTEGER li;
        if (QueryPerformanceFrequency(&li))
            return 1.0 / li.QuadPart;
        else
            return 0.000001;  /* unlikely */
    }

#elif defined(__APPLE__)

    /*  Use MacOS X time base (TB) available for Intel and PowerPC.

        See example 7.41 timebase_demo.c from the book "MacOS X Internals"
        by Amit Singh, <http://www.osxbook.com/book/src/singh-src.tar.bz2>,
        <http://books.google.com/books?id=K8vUkpOXhN4C&pg=PA782#PPA783>
        and <http://developer.apple.com/qa/qa2004/qa1398.html>.
     */
#   define HPTIME   "TB"
#   define hpTime_t  u_int64_t
#   include <mach/mach.h>
#   include <mach/mach_time.h>

    static hpTime_t  /* return TB as ticks */
    hpTimer(void)
    {
        return (hpTime_t)mach_absolute_time();
    }

    static double  /* return TB resolution as secs/tick */
    hpTimerUnit(void)
    {
        mach_timebase_info_data_t info;
        info.denom = info.numer = 0;
        if (mach_timebase_info(&info) == KERN_SUCCESS && info.denom && info.numer) {
            /* TB info ratio is nanosec ticks */
            return 1.0e-9 * (double)info.numer / (double)info.denom;
        } else
            return 1.0e-9;  /* unlikely */
    }

#elif defined(SOLARIS) || defined(__hppa) || defined(hppa)  /* means HP_UX? */

    /*  Solaris and HP-UX both provide a function  gethrtime
        returning a 64-bit time stamp with nanosec resolution
     */
#   define HPTIME   "hrtime"
#   define hpTime_t  hrtime_t
#   include <unistd.h>  /* sleep */
#   include <time.h>

    static hpTime_t
    hpTimer(void)
    {
        return (hpTime_t)gethrtime();
    }

    static double
    hpTimerUnit(void)
    {
        return 1.0e-9;  /* nanosec ticks */
    }

#else  /* !MS_WINDOWS && !__APPLE__ && !SOLARIS && !HP_UX */
#   define hpTime_t  unsigned long long

#   include <sys/time.h>  /* <sys/resource.h>? */
#   include <sys/times.h>
# ifdef RDTSCLL  /* some Linux distros */
#   include <asm/msr.h>
# endif

# ifdef GETTIMEOFDAY_NO_TZ
#   define _LSPROF_TV(tv)  (!gettimeofday(&tv) \
                            ? (((hpTime_t)tv.tv_sec * 1000000) + \
                                (hpTime_t)tv.tv_usec) \
                            : (hpTime_t)++tv.tv_sec) /* unlikely */
# else
#   define _LSPROF_TV(tv)  (!gettimeofday(&tv, (struct timezone *)NULL) \
                            ? (((hpTime_t)tv.tv_sec * 1000000) + \
                                (hpTime_t)tv.tv_usec) \
                            : (hpTime_t)++tv.tv_sec) /* unlikely */
# endif

# ifdef __GNUC__  /* or PYCC_GCC? */
#   define _LSPROF_TSC 1  /* try time stamp count (TSC or TB) */
# else
#   undef  _LSPROF_TSC    /* no TSC */
# endif

    static hpTime_t
    hpTimer (void)
    {
        hpTime_t t;
# if defined(_LSPROF_TSC) && defined(rdtscll)
#       define HPTIME  "rdtscll"
        /* rdtscll() is defined in <asm/msr.h>
           of certain, recent Linux distros */
        rdtscll(t);
# elif defined(_LSPROF_TSC) && (defined(__i386__) || defined(__x86_64__))
#       define HPTIME  "x86"
        unsigned int hi, lo;
        __asm__ __volatile__ (
            "rdtsc \n\t"
            : "=a"(lo), "=d"(hi)
            :
        );
        t = (((hpTime_t)hi) << 32) | lo;
# elif defined(_LSPROF_TSC) && (defined(__ppc64__) || defined(__POWERPC64__))
#       define HPTIME  "ppc64"
        __asm__ __volatile__ (
            "mftb %0 \n\t"
            : "=r"(t)
            :
        );
# elif defined(_LSPROF_TSC) && (defined(__ppc__) || defined(__POWERPC__))
        unsigned int hi, lo, hj;
#       define HPTIME  "ppc"
        /* avoid label (causes duplicate symbols for
           -O3 and above with older GNU C compilers) */
        do {
            __asm__ __volatile__ (
                "mftbu %0 \n\t"
                "mftb  %1 \n\t"
                "mftbu %2 \n\t"
                : "=r"(hi), "=r"(lo), "=r"(hj)
                :
            );
        } while (hi != hj);
        t = (((hpTime_t)hi) << 32) | lo;
# else  /* gettimeofday as usec ticks */
#       define HPTIME  "GTOD"
        struct timeval tv;
        t = _LSPROF_TV(tv);
#       undef _LSPROF_TSC  /* no TSC*/
# endif
        return (t);
    }

# ifdef _LSPROF_TSC
    /* start ticks and time, set at
       initialzation in init_lsprof */
    static hpTime_t  hpTime0 = 0;
    static hpTime_t  usTime0 = 0;

    static double
    hpTimerUnit(void)
    {
        /* determine TSC resolution by waiting an
           add'l 250 usecs since initialization */
        struct timeval tv;
        hpTime_t hp, us, usTime1;

        usTime1 = usTime0 + 250000;
        do {
            hp = hpTimer();
            us = _LSPROF_TV(tv);
            /* reset on wrap around of any */
            if (hpTime0 > hp || usTime0 > us || usTime0 >= usTime1) {
                hpTime0 = hp;
                usTime0 = us;
                usTime1 = us + 250000;
            }
        } while (us <= usTime1 || hp <= hpTime0);
        return 1.0e-6 * (double)(us - usTime0) / (double)(hp - hpTime0);
    }
# else  /* !_LSPROF_TSC */
    static double
    hpTimerUnit(void)
    {
        return 1.0e-6;  /* usec ticks */
    }
# endif  /* _LSPROF_TSC */

#endif


int
main(int argc, char **argv)
{
    hpTime_t t, d;
    double   r, Hz;
    int      s = 1;  /* default sleep time */

#ifdef _LSPROF_TSC
    struct timeval tv;
    int w = 4;  /* warm up */
    while (w-- > 0)
        hpTime0 = hpTimer();
    usTime0 = _LSPROF_TV(tv);
#endif  /* _LSPROF_TSC */

    if (argc == 2) {  /* <seconds> */
        int a = atoi(argv[1]);
        if (a >= 0 && a <= 60)
            s = a;
    }
    printf("test: sleeping for %d secs (%s) ...\n", s, HPTIME);

    t = hpTimer();
    if (s > 0) sleep(s);
    d = hpTimer() - t;

    r = hpTimerUnit();
    Hz = r ? (1.0 / r) : 0.0;
    printf("resolution %.3f nsecs/tick (%.3f GHz)\n", r * 1.0e9, (double)Hz * 1.0e-9);
    printf("sleep time %llu ticks or %.9f secs\n", d, r * (double)d);

    exit(0);
}