void kmp_stats_output_module::printTimerStats(FILE *statsOut,
statistic const *theStats,
statistic const *totalStats) {
fprintf(statsOut, "Timer, SampleCount, Min, "
"Mean, Max, Total, SD\n");
for (timer_e s = timer_e(0); s < TIMER_LAST; s = timer_e(s + 1)) {
statistic const *stat = &theStats[s];
char tag = timeStat::noUnits(s) ? ' ' : 'T';
fprintf(statsOut, "%-28s, %s\n", timeStat::name(s),
stat->format(tag, true).c_str());
}
// Also print the Total_ versions of times.
for (timer_e s = timer_e(0); s < TIMER_LAST; s = timer_e(s + 1)) {
char tag = timeStat::noUnits(s) ? ' ' : 'T';
if (totalStats && !timeStat::noTotal(s))
fprintf(statsOut, "Total_%-22s, %s\n", timeStat::name(s),
totalStats[s].format(tag, true).c_str());
}
}
void kmp_stats_output_module::printStats(FILE *statsOut, statistic const * theStats, bool areTimers)
{
if (areTimers)
{
// Check if we have useful timers, since we don't print zero value timers we need to avoid
// printing a header and then no data.
bool haveTimers = false;
for (int s = 0; s<TIMER_LAST; s++)
{
if (theStats[s].getCount() != 0)
{
haveTimers = true;
break;
}
}
if (!haveTimers)
return;
}
// Print
const char * title = areTimers ? "Timer, SampleCount," : "Counter, ThreadCount,";
fprintf (statsOut, "%s Min, Mean, Max, Total, SD\n", title);
if (areTimers) {
for (int s = 0; s<TIMER_LAST; s++) {
statistic const * stat = &theStats[s];
if (stat->getCount() != 0) {
char tag = timeStat::noUnits(timer_e(s)) ? ' ' : 'T';
fprintf (statsOut, "%-25s, %s\n", timeStat::name(timer_e(s)), stat->format(tag, true).c_str());
}
}
} else { // Counters
for (int s = 0; s<COUNTER_LAST; s++) {
statistic const * stat = &theStats[s];
fprintf (statsOut, "%-25s, %s\n", counter::name(counter_e(s)), stat->format(' ', true).c_str());
}
}
}
void kmp_stats_output_module::outputStats(const char* heading)
{
statistic allStats[TIMER_LAST];
statistic allCounters[COUNTER_LAST];
// stop all the explicit timers for all threads
windupExplicitTimers();
FILE * eventsOut;
FILE * statsOut = outputFileName ? fopen (outputFileName, "a+") : stderr;
if (eventPrintingEnabled()) {
eventsOut = fopen(eventsFileName, "w+");
}
if (!statsOut)
statsOut = stderr;
fprintf(statsOut, "%s\n",heading);
// Accumulate across threads.
kmp_stats_list::iterator it;
for (it = __kmp_stats_list.begin(); it != __kmp_stats_list.end(); it++) {
int t = (*it)->getGtid();
// Output per thread stats if requested.
if (perThreadPrintingEnabled()) {
fprintf (statsOut, "Thread %d\n", t);
printStats(statsOut, (*it)->getTimers(), true);
printCounters(statsOut, (*it)->getCounters());
fprintf(statsOut,"\n");
}
// Output per thread events if requested.
if (eventPrintingEnabled()) {
kmp_stats_event_vector events = (*it)->getEventVector();
printEvents(eventsOut, &events, t);
}
for (int s = 0; s<TIMER_LAST; s++) {
// See if we should ignore this timer when aggregating
if ((timeStat::masterOnly(timer_e(s)) && (t != 0)) || // Timer is only valid on the master and this thread is a worker
(timeStat::workerOnly(timer_e(s)) && (t == 0)) || // Timer is only valid on a worker and this thread is the master
timeStat::synthesized(timer_e(s)) // It's a synthesized stat, so there's no raw data for it.
)
{
continue;
}
statistic * threadStat = (*it)->getTimer(timer_e(s));
allStats[s] += *threadStat;
}
// Special handling for synthesized statistics.
// These just have to be coded specially here for now.
// At present we only have a few:
// The total parallel work done in each thread.
// The variance here makes it easy to see load imbalance over the whole program (though, of course,
// it's possible to have a code with awful load balance in every parallel region but perfect load
// balance oever the whole program.)
// The time spent in barriers in each thread.
allStats[TIMER_Total_work].addSample ((*it)->getTimer(TIMER_OMP_work)->getTotal());
// Time in explicit barriers.
allStats[TIMER_Total_barrier].addSample ((*it)->getTimer(TIMER_OMP_barrier)->getTotal());
for (int c = 0; c<COUNTER_LAST; c++) {
if (counter::masterOnly(counter_e(c)) && t != 0)
continue;
allCounters[c].addSample ((*it)->getCounter(counter_e(c))->getValue());
}
}
if (eventPrintingEnabled()) {
printPloticusFile();
fclose(eventsOut);
}
fprintf (statsOut, "Aggregate for all threads\n");
printStats (statsOut, &allStats[0], true);
fprintf (statsOut, "\n");
printStats (statsOut, &allCounters[0], false);
if (statsOut != stderr)
fclose(statsOut);
}
void kmp_stats_output_module::outputStats(const char *heading) {
// Stop all the explicit timers in all threads
// Do this before declaring the local statistics because thay have
// constructors so will take time to create.
windupExplicitTimers();
statistic allStats[TIMER_LAST];
statistic totalStats[TIMER_LAST]; /* Synthesized, cross threads versions of
normal timer stats */
statistic allCounters[COUNTER_LAST];
FILE *statsOut =
!outputFileName.empty() ? fopen(outputFileName.c_str(), "a+") : stderr;
if (!statsOut)
statsOut = stderr;
FILE *eventsOut;
if (eventPrintingEnabled()) {
eventsOut = fopen(eventsFileName, "w+");
}
printHeaderInfo(statsOut);
fprintf(statsOut, "%s\n", heading);
// Accumulate across threads.
kmp_stats_list::iterator it;
for (it = __kmp_stats_list->begin(); it != __kmp_stats_list->end(); it++) {
int t = (*it)->getGtid();
// Output per thread stats if requested.
if (printPerThreadFlag) {
fprintf(statsOut, "Thread %d\n", t);
printTimerStats(statsOut, (*it)->getTimers(), 0);
printCounters(statsOut, (*it)->getCounters());
fprintf(statsOut, "\n");
}
// Output per thread events if requested.
if (eventPrintingEnabled()) {
kmp_stats_event_vector events = (*it)->getEventVector();
printEvents(eventsOut, &events, t);
}
// Accumulate timers.
for (timer_e s = timer_e(0); s < TIMER_LAST; s = timer_e(s + 1)) {
// See if we should ignore this timer when aggregating
if ((timeStat::masterOnly(s) && (t != 0)) || // Timer only valid on master
// and this thread is worker
(timeStat::workerOnly(s) && (t == 0)) // Timer only valid on worker
// and this thread is the master
) {
continue;
}
statistic *threadStat = (*it)->getTimer(s);
allStats[s] += *threadStat;
// Add Total stats for timers that are valid in more than one thread
if (!timeStat::noTotal(s))
totalStats[s].addSample(threadStat->getTotal());
}
// Accumulate counters.
for (counter_e c = counter_e(0); c < COUNTER_LAST; c = counter_e(c + 1)) {
if (counter::masterOnly(c) && t != 0)
continue;
allCounters[c].addSample((*it)->getCounter(c)->getValue());
}
}
if (eventPrintingEnabled()) {
printPloticusFile();
fclose(eventsOut);
}
fprintf(statsOut, "Aggregate for all threads\n");
printTimerStats(statsOut, &allStats[0], &totalStats[0]);
fprintf(statsOut, "\n");
printCounterStats(statsOut, &allCounters[0]);
if (statsOut != stderr)
fclose(statsOut);
}