/* local or global wall-clock time */
uint64_t vt_pform_wtime() {
#if TIMER == TIMER_CYCLE_COUNTER
# ifdef _CRAYC
return (uint64_t)_rtc();
# else
uint64_t clock_value;
uint32_t low = 0;
uint32_t high = 0;
asm volatile ("rdtsc" : "=a" (low), "=d" (high));
clock_value = ((uint64_t)high << 32) | (uint64_t)low;
return clock_value;
# endif
#elif TIMER == TIMER_CLOCK_GETTIME
struct timespec tp;
clock_gettime(CLOCK_REALTIME, &tp);
return ((tp.tv_sec - vt_time_base) * 1000000000LL) + tp.tv_nsec;
#elif TIMER == TIMER_GETTIMEOFDAY
struct timeval tp;
gettimeofday(&tp, 0);
return ((tp.tv_sec - vt_time_base) * 1000000LL) + tp.tv_usec;
#elif TIMER == TIMER_PAPI_REAL_CYC
return vt_metric_real_cyc();
#elif TIMER == TIMER_PAPI_REAL_USEC
return vt_metric_real_usec() - vt_time_base;
#endif
}
/* platform specific initialization */
void vt_pform_init() {
struct timeval tp;
static ApTeam_t app;
int i;
#if TIMER == TIMER_RTC
vt_ticks_per_sec = (uint64_t)sysconf(_SC_SV2_USER_TIME_RATE);
# pragma omp parallel
{
gettimeofday(&tp, 0);
vt_rtc_base = _rtc();
}
#elif TIMER == TIMER_GETTIMEOFDAY
gettimeofday(&tp, 0);
vt_time_base = tp.tv_sec - (tp.tv_sec & 0xFFFF);
#elif TIMER == TIMER_PAPI_REAL_USEC
vt_time_base = vt_metric_real_usec();
#endif
if (apteamctl(ApTeam_Status, 0, 0, &app) == 1) {
vt_mspmode = (app.flags & APTEAM_MSP);
app.pes = (ApPe_t*)malloc(app.pecount * sizeof(ApPe_t));
if (apteamctl(ApTeam_Status, 0, 0, &app) == 1) {
for (i=0; i<_num_pes(); i++) {
if (_my_pe() == app.pes[i].lpe) vt_nodeid = (long)app.pes[i].place;
}
}
}
}
timef_(void)
#endif
{
static int64 initial_rt = -1; /* Clock value from initial call */
int64 rt, rtdif;
double retval;
#ifdef _UNICOS
if (_sec_per_clock == 0.0)
_sec_per_clock = 1.0 / (double) sysconf(_SC_CLK_TCK);
rt = _rtc();
#else
{
struct timeval buf;
struct timezone buf2;
(void) gettimeofday (&buf, &buf2);
rt = (long long)buf.tv_sec * 1000000LL + buf.tv_usec;
}
#endif
if (initial_rt < 0) {
initial_rt = rt;
rtdif = 0;
/*
* force rtdif to 0 to prevent anomalies due to possible
* race conditions between 2 or more tasks calling TIMEF
* concurrently on the initial call.
*/
}
else
rtdif = rt - initial_rt;
/*
* On pre-7.0 UNICOS CX/CEA systems and on all CRAY-2 systems the
* real-time hardware clock is set to 0 on reboot. If a restarted
* process had called TIMEF before the system was brought down
* and then after reboot, a negative difference in the real-time
* clock value would be observed. To minimize the impact of
* wrong timings being returned, we return 0 when this situation is
* detected.
*/
if (rtdif < 0) {
initial_rt = rt;
rtdif = 0;
}
retval = (double) rtdif * MSECPERCLK;
#ifdef _UNICOS
if (_numargs() > 0)
*time = (_f_real) retval;
#endif
return( (_f_real) retval);
}
/* cray timers */
void metric_read_craytimers(int tid, int idx, double values[]) {
#ifdef CRAY_TIMERS
#ifdef TAU_CATAMOUNT /* for Cray XT3 */
values[idx] = dclock() * 1.0e6;
#else /* for Cray X1 */
long long tick = _rtc();
values[idx] = (double)tick / HZ;
#endif /* TAU_CATAMOUNT */
#endif /* CRAY_TIMERS */
}
/* local or global wall-clock time */
uint64_t vt_pform_wtime() {
#if TIMER == TIMER_RTC
return (uint64_t)_rtc() - vt_rtc_base;
#elif TIMER == TIMER_GETTIMEOFDAY
struct timeval tp;
gettimeofday(&tp, 0);
return ((tp.tv_sec - vt_time_base) * 1e6) + tp.tv_usec;
#elif TIMER == TIMER_PAPI_REAL_CYC
return vt_metric_real_cyc();
#elif TIMER == TIMER_PAPI_REAL_USEC
return vt_metric_real_usec() - vt_time_base;
#endif
}
/*
* _glio_delay
*
* Delay for nano nanoseconds. This is useful in polling loops.
*/
void
_glio_delay(int nano)
{
#ifdef __mips
struct timespec ts;
ts.tv_sec = 0;
ts.tv_nsec = nano;
nanosleep(&ts, NULL);
#else
float hertz;
long ticks, start;
hertz = IRTC_RATE();
ticks = hertz * (float)nano / 1000000000.0;
start = _rtc();
/* if the clock goes backwards, stop spinning */
while (_rtc() < (start+ticks) && _rtc() >= start)
continue;
#endif
return;
}
_f_real
SECONDR(_f_real *time)
{
double timeval;
if (_sec_per_clock == 0.0)
_sec_per_clock = (double)1.0 / (double)__hertz;
timeval = (double) _rtc() * _sec_per_clock;
if (_numargs() > 0)
*time = (_f_real) timeval;
return( (_f_real) timeval);
}
void
_mtcpu()
{
struct stkstat stats; /* Stack statistics struct */
double waltime, clks, percent;
/* Print total clocks run, wallclock time, and % of machine used */
waltime = (double) (_rtc() - _rtstart) / (double) CLK_TCK;
clks = _second();
percent = (waltime > 0.0) ?
(100.0 * clks / (waltime * (double) ZQMAXCPU)) : 0.0;
if (waltime >= 0.0) {
if (percent > 0.05 && percent < 99.9999)
(void) fprintf(stderr,
" CPU: %.3fs, Wallclock: %.3fs, %.1f%% of %d-CPU Machine\n",
clks, waltime, percent, ZQMAXCPU);
else
(void) fprintf(stderr,
" CPU: %.3fs, Wallclock: %.3fs\n", clks,
waltime);
}
else
/*
* If system reboot after checkpoint, value of wallclock time
* may be negative. Kernel mod only fixes CX/CEA systems.
* Problem can still occur for Cray-2 systems. Do not give
* bad wallclock value.
*/
(void) fprintf(stderr, " CPU: %.3fs\n", clks);
/* Print memory/stack hi-water marks and number of overflows */
STKSTAT(&stats);
(void) fprintf(stderr,
" Memory HWM: %d, Stack HWM: %d, Stack segment expansions: %d\n",
$memhwm, MSWM, SOVFL);
return;
}
