void event_create_eventList(int *eventSet, int eventCodeSetSize, int *eventCodeSet, int threadID) {
int retval, i, maxNumberHwCounters, eventCodeSetMaxSize;
PAPI_event_info_t info;
maxNumberHwCounters = PAPI_get_opt( PAPI_MAX_HWCTRS, NULL );
printf("Max number of hardware counters = %d \n", maxNumberHwCounters);
eventCodeSetMaxSize = PAPI_get_opt( PAPI_MAX_MPX_CTRS, NULL );
printf("Max number of multiplexed counters = %d \n", eventCodeSetMaxSize);
if ( eventCodeSetMaxSize < eventCodeSetSize)
test_fail( __FILE__, __LINE__, "eventCodeSetMaxSize < eventCodeSetSize, too many performance events defined! ", retval );
retval = PAPI_register_thread();
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_register_thread", retval );
retval = PAPI_create_eventset( eventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_create_eventset", retval );
retval = PAPI_assign_eventset_component( *eventSet, 0 );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_assign_eventset_component", retval );
retval = PAPI_set_multiplex( *eventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_set_multiplex", retval );
for (i = 0; i < eventCodeSetSize; i++) {
retval = PAPI_get_event_info(eventCodeSet[i], &info);
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_get_event_info", retval );
retval = PAPI_add_event( *eventSet, info.event_code);
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_add_event", retval );
else
printf("Adding %s \n", info.symbol);
}
printf("event_create_eventList done \n");
}
void
init_test(int SoftwareMPX, int KernelMPX, int* Events)
{
int i;
int retval;
PAPI_option_t option, itimer;
if ( ( retval = PAPI_assign_eventset_component( SoftwareMPX, 0 ) ) != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_assign_eventset_component", retval);
if ( ( retval = PAPI_assign_eventset_component( KernelMPX, 0 ) ) != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_assign_eventset_component", retval);
if ( ( retval = PAPI_set_multiplex( KernelMPX ) ) != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_set_multiplex", retval );
PAPI_get_opt(PAPI_DEF_ITIMER,&itimer);
memset(&option,0x0,sizeof(option));
option.multiplex.flags = PAPI_MULTIPLEX_FORCE_SW;
option.multiplex.eventset = SoftwareMPX;
option.multiplex.ns = itimer.itimer.ns;
PAPI_set_opt( PAPI_MULTIPLEX, &option );
for (i = 0; i < options.min - 1; i++) {
if ( options.kernel_mpx ) {
if ( ( retval = PAPI_add_event( KernelMPX, Events[i]) ) != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_add_event", retval );
}
}
if ( options.force_sw ) {
if ( ( retval = PAPI_add_event( SoftwareMPX, Events[i]) ) != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_add_event", retval );
}
}
}
}
/*
*______________________________________________
*
*InitCmd --
*
* This function is to add events to papi event set.
* A list of events is taken as argument.
*
* Results:
* The eventset is initialized.
*________________________________________________
*/
int
InitCmd(ClientData cdata, Tcl_Interp *interp, int objc, Tcl_Obj *const objv[])
{
if(objc!=2) {
fprintf(stderr,"Usage: perft::init listOfEvents");
return 1;
}
int retval,eventCode;
char* arg=Tcl_GetString(objv[1]);
int argc, eventList;
const char **argv;
eventList= Tcl_SplitList(interp, arg, &argc, &argv);
char **index=(char **)argv;
if(papiState.eventsData==NULL) {
AllocateStringArray(PAPI_MAX_EVENTS,PAPI_MAX_STRING_LENGTH,&papiState.eventsData);
} else {
retval=PAPI_cleanup_eventset(papiState.eventSet);
if ( retval!= PAPI_OK) {
HandlePapiError(retval);
return 1;
}
// Free all memory and data structures, EventSet must be empty
retval=PAPI_destroy_eventset(&papiState.eventSet);
if ( retval!= PAPI_OK) {
HandlePapiError(retval);
return 1;
}
}
papiState.eventSet=PAPI_NULL;
papiState.numberOfEvents=0;
retval=InitPapi();
if(retval==1) {
return 1;
}
retval=InitPapiMultiplexing();
if(retval==1) {
return 1;
}
retval=PAPI_assign_eventset_component(papiState.eventSet, 0);
if ( retval!= PAPI_OK) {
printf("Error in subroutine PAPI_assign_eventset_component");
HandlePapiError(retval);
return 1;
}
retval=PAPI_set_multiplex(papiState.eventSet);
if ( retval!= PAPI_OK) {
printf("Error in subroutine PAPI_set_multiplex");
HandlePapiError(retval);
return 1;
}
while(papiState.numberOfEvents<argc) {
retval=PAPI_event_name_to_code(*index,&eventCode);
if ( retval!= PAPI_OK) {
printf("Error in getting information about event name: %s\n",*index);
HandlePapiError(retval);
return 1;
}
papiState.eventsData[papiState.numberOfEvents]=*index;
retval=PAPI_add_event(papiState.eventSet, eventCode);
if(retval!=PAPI_OK) {
printf("Error in adding event: %s\n",*index);
HandlePapiError(retval);
return 1;
}
index+=1;
papiState.numberOfEvents+=1;
}
Tcl_SetObjResult(interp, Tcl_NewStringObj("Hardware counters initialized", -1));
return TCL_OK;
}
int main(int argc, char **argv)
{
PAPI_event_info_t info;
int i, j, retval;
int iters = 10000000;
double x = 1.1, y;
long long t1, t2;
long long values[MAXEVENTS], refvalues[MAXEVENTS];
int sleep_time = SLEEPTIME;
double valsqsum[MAXEVENTS];
double valsum[MAXEVENTS];
double spread[MAXEVENTS];
int nevents = MAXEVENTS;
int eventset = PAPI_NULL;
int events[MAXEVENTS];
events[0] = PAPI_BR_NTK;
events[1] = PAPI_BR_PRC;
events[2] = PAPI_BR_INS;
events[3] = PAPI_BR_MSP;
/*
events[3]=PAPI_BR_CN;
events[4]=PAPI_BR_UCN;*/
/*events[5]=PAPI_BR_TKN; */
for (i = 0; i < MAXEVENTS; i++) {
values[i] = 0;
valsqsum[i] = 0;
valsum[i] = 0;
}
if (argc > 1) {
if (!strcmp(argv[1], "TESTS_QUIET"))
tests_quiet(argc, argv);
else {
sleep_time = atoi(argv[1]);
if (sleep_time <= 0)
sleep_time = SLEEPTIME;
}
}
if (!TESTS_QUIET) {
printf("\nAccuracy check of branch presets.\n");
printf("Comparing a measurement with separate measurements.\n\n");
}
if ((retval = PAPI_library_init(PAPI_VER_CURRENT)) != PAPI_VER_CURRENT)
test_fail(__FILE__, __LINE__, "PAPI_library_init", retval);
if ((retval = PAPI_create_eventset(&eventset)))
test_fail(__FILE__, __LINE__, "PAPI_create_eventset", retval);
#ifdef MPX
if ((retval = PAPI_multiplex_init()))
test_fail(__FILE__, __LINE__, "PAPI_multiplex_init", retval);
if ((retval = PAPI_set_multiplex(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_set_multiplex", retval);
#endif
nevents = 0;
for (i = 0; i < MAXEVENTS; i++) {
if (PAPI_query_event(events[i]) != PAPI_OK)
continue;
if (PAPI_add_event(eventset, events[i]) == PAPI_OK) {
events[nevents] = events[i];
nevents++;
}
}
if (nevents < 1)
test_skip(__FILE__, __LINE__, "Not enough events left...", 0);
/* Find a reasonable number of iterations (each
* event active 20 times) during the measurement
*/
t2 = 10000 * 20 * nevents; /* Target: 10000 usec/multiplex, 20 repeats */
if (t2 > 30e6)
test_skip(__FILE__, __LINE__, "This test takes too much time", retval);
/* Measure one run */
t1 = PAPI_get_real_usec();
y = dummy3(x, iters);
t1 = PAPI_get_real_usec() - t1;
if (t2 > t1) /* Scale up execution time to match t2 */
iters = iters * (int)(t2 / t1);
else if (t1 > 30e6) /* Make sure execution time is < 30s per repeated test */
test_skip(__FILE__, __LINE__, "This test takes too much time", retval);
x = 1.0;
if (!TESTS_QUIET)
printf("\nFirst run: Together.\n");
t1 = PAPI_get_real_usec();
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
y = dummy3(x, iters);
if ((retval = PAPI_stop(eventset, values)))
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
t2 = PAPI_get_real_usec();
if (!TESTS_QUIET) {
printf("\tOperations= %.1f Mflop", y * 1e-6);
printf("\t(%g Mflop/s)\n\n", ((float) y / (t2 - t1)));
printf("PAPI grouped measurement:\n");
}
for (j = 0; j < nevents; j++) {
PAPI_get_event_info(events[j], &info);
if (!TESTS_QUIET) {
printf("%20s = ", info.short_descr);
printf(LLDFMT, values[j]);
printf("\n");
}
}
if (!TESTS_QUIET)
printf("\n");
if ((retval = PAPI_remove_events(eventset, events, nevents)))
test_fail(__FILE__, __LINE__, "PAPI_remove_events", retval);
if ((retval = PAPI_destroy_eventset(&eventset)))
test_fail(__FILE__, __LINE__, "PAPI_destroy_eventset", retval);
eventset = PAPI_NULL;
if ((retval = PAPI_create_eventset(&eventset)))
test_fail(__FILE__, __LINE__, "PAPI_create_eventset", retval);
for (i = 0; i < nevents; i++) {
if ((retval = PAPI_cleanup_eventset(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_cleanup_eventset", retval);
if ((retval = PAPI_add_event(eventset, events[i])))
test_fail(__FILE__, __LINE__, "PAPI_add_event", retval);
x = 1.0;
if (!TESTS_QUIET)
printf("\nReference measurement %d (of %d):\n", i + 1, nevents);
t1 = PAPI_get_real_usec();
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
y = dummy3(x, iters);
if ((retval = PAPI_stop(eventset, &refvalues[i])))
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
t2 = PAPI_get_real_usec();
if (!TESTS_QUIET) {
printf("\tOperations= %.1f Mflop", y * 1e-6);
printf("\t(%g Mflop/s)\n\n", ((float) y / (t2 - t1)));
}
PAPI_get_event_info(events[i], &info);
if (!TESTS_QUIET) {
printf("PAPI results:\n%20s = ", info.short_descr);
printf(LLDFMT, refvalues[i]);
printf("\n");
}
}
if (!TESTS_QUIET)
printf("\n");
if (!TESTS_QUIET) {
printf("\n\nRelative accuracy:\n");
for (j = 0; j < nevents; j++)
printf(" Event %.2d", j);
printf("\n");
}
for (j = 0; j < nevents; j++) {
spread[j] = abs((int)(refvalues[j] - values[j]));
if (values[j])
spread[j] /= (double) values[j];
if (!TESTS_QUIET)
printf("%10.3g ", spread[j]);
/* Make sure that NaN get counted as errors */
if (spread[j] < MPX_TOLERANCE)
i--;
else if (refvalues[j] < MINCOUNTS) /* Neglect inprecise results with low counts */
i--;
}
if (!TESTS_QUIET)
printf("\n\n");
if (i)
test_fail(__FILE__, __LINE__, "Values outside threshold", i);
else
test_pass(__FILE__, NULL, 0);
return 0;
}
int main(int argc, char **argv)
{
PAPI_event_info_t info;
char name2[PAPI_MAX_STR_LEN];
int i, j, retval, idx, repeats;
int iters = NUM_FLOPS;
double x = 1.1, y, dtmp;
long long t1, t2;
long long values[MAXEVENTS], refvals[MAXEVENTS];
int nsamples[MAXEVENTS], truelist[MAXEVENTS], ntrue;
#ifdef STARTSTOP
long long dummies[MAXEVENTS];
#endif
int sleep_time = SLEEPTIME;
double valsample[MAXEVENTS][REPEATS];
double valsum[MAXEVENTS];
double avg[MAXEVENTS];
double spread[MAXEVENTS];
int nevents = MAXEVENTS, nev1;
int eventset = PAPI_NULL;
int events[MAXEVENTS];
int eventidx[MAXEVENTS];
int eventmap[MAXEVENTS];
int fails;
events[0] = PAPI_FP_INS;
events[1] = PAPI_TOT_CYC;
events[2] = PAPI_TOT_INS;
events[3] = PAPI_TOT_IIS;
events[4] = PAPI_INT_INS;
events[5] = PAPI_STL_CCY;
events[6] = PAPI_BR_INS;
events[7] = PAPI_SR_INS;
events[8] = PAPI_LD_INS;
for (i = 0; i < MAXEVENTS; i++) {
values[i] = 0;
valsum[i] = 0;
nsamples[i] = 0;
}
if (argc > 1) {
if (!strcmp(argv[1], "TESTS_QUIET"))
tests_quiet(argc, argv);
else {
sleep_time = atoi(argv[1]);
if (sleep_time <= 0)
sleep_time = SLEEPTIME;
}
}
if (!TESTS_QUIET) {
printf("\nFunctional check of multiplexing routines.\n");
printf("Adding and removing events from an event set.\n\n");
}
if ((retval = PAPI_library_init(PAPI_VER_CURRENT)) != PAPI_VER_CURRENT)
test_fail(__FILE__, __LINE__, "PAPI_library_init", retval);
#ifdef MPX
init_multiplex();
#endif
if ((retval = PAPI_create_eventset(&eventset)))
test_fail(__FILE__, __LINE__, "PAPI_create_eventset", retval);
#ifdef MPX
/* In Component PAPI, EventSets must be assigned a component index
before you can fiddle with their internals.
0 is always the cpu component */
retval = PAPI_assign_eventset_component(eventset, 0);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_assign_eventset_component", retval);
if ((retval = PAPI_set_multiplex(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_set_multiplex", retval);
#endif
nevents = MAXEVENTS;
for (i = 0; i < nevents; i++) {
if ((retval = PAPI_add_event(eventset, events[i]))) {
for (j = i; j < MAXEVENTS; j++)
events[j] = events[j + 1];
nevents--;
i--;
}
}
if (nevents < 3)
test_skip(__FILE__, __LINE__, "Not enough events left...", 0);
/* Find a reasonable number of iterations (each
* event active 20 times) during the measurement
*/
t2 = 10000 * 20 * nevents; /* Target: 10000 usec/multiplex, 20 repeats */
if (t2 > 30e6)
test_skip(__FILE__, __LINE__, "This test takes too much time", retval);
/* Measure one run */
t1 = PAPI_get_real_usec();
y = dummy3(x, iters);
t1 = PAPI_get_real_usec() - t1;
if (t2 > t1) /* Scale up execution time to match t2 */
iters = iters * (int)(t2 / t1);
else if (t1 > 30e6) /* Make sure execution time is < 30s per repeated test */
test_skip(__FILE__, __LINE__, "This test takes too much time", retval);
j = nevents;
for (i = 1; i < nevents; i = i + 2)
eventidx[--j] = i;
for (i = 0; i < nevents; i = i + 2)
eventidx[--j] = i;
assert(j == 0);
for (i = 0; i < nevents; i++)
eventmap[i] = i;
x = 1.0;
if (!TESTS_QUIET)
printf("\nReference run:\n");
t1 = PAPI_get_real_usec();
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
y = dummy3(x, iters);
PAPI_read(eventset, refvals);
t2 = PAPI_get_real_usec();
ntrue = nevents;
PAPI_list_events(eventset, truelist, &ntrue);
if (!TESTS_QUIET) {
printf("\tOperations= %.1f Mflop", y * 1e-6);
printf("\t(%g Mflop/s)\n\n", ((float) y / (t2 - t1)));
printf("%20s %16s %-15s %-15s\n", "PAPI measurement:",
"Acquired count", "Expected event", "PAPI_list_events");
}
if (!TESTS_QUIET) {
for (j = 0; j < nevents; j++) {
PAPI_get_event_info(events[j], &info);
PAPI_event_code_to_name(truelist[j], name2);
if (!TESTS_QUIET)
printf("%20s = %16lld %-15s %-15s %s\n", info.short_descr, refvals[j],
info.symbol, name2, strcmp(info.symbol,
name2) ? "*** MISMATCH ***" : "");
}
printf("\n");
}
nev1 = nevents;
repeats = nevents * 4;
for (i = 0; i < repeats; i++) {
if ((i % nevents) + 1 == nevents)
continue;
if (!TESTS_QUIET)
printf("\nTest %d (of %d):\n", i + 1 - i / nevents, repeats - 4);
if ((retval = PAPI_stop(eventset, values)))
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
j = eventidx[i % nevents];
if ((i / nevents) % 2 == 0) {
PAPI_get_event_info(events[j], &info);
if (!TESTS_QUIET)
printf("Removing event[%d]: %s\n", j, info.short_descr);
if ((retval = PAPI_remove_event(eventset, events[j])))
test_fail(__FILE__, __LINE__, "PAPI_remove_event", retval);
nev1--;
for (idx = 0; eventmap[idx] != j; idx++);
for (j = idx; j < nev1; j++)
eventmap[j] = eventmap[j + 1];
} else {
PAPI_get_event_info(events[j], &info);
if (!TESTS_QUIET)
printf("Adding event[%d]: %s\n", j, info.short_descr);
if ((retval = PAPI_add_event(eventset, events[j])))
test_fail(__FILE__, __LINE__, "PAPI_add_event", retval);
eventmap[nev1] = j;
nev1++;
}
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
x = 1.0;
#ifndef STARTSTOP
if ((retval = PAPI_reset(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_reset", retval);
#else
if ((retval = PAPI_stop(eventset, dummies)))
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
#endif
t1 = PAPI_get_real_usec();
y = dummy3(x, iters);
PAPI_read(eventset, values);
t2 = PAPI_get_real_usec();
if (!TESTS_QUIET) {
printf("\n(calculated independent of PAPI)\n");
printf("\tOperations= %.1f Mflop", y * 1e-6);
printf("\t(%g Mflop/s)\n\n", ((float) y / (t2 - t1)));
printf("%20s %16s %-15s %-15s\n", "PAPI measurement:",
"Acquired count", "Expected event", "PAPI_list_events");
}
ntrue = nev1;
PAPI_list_events(eventset, truelist, &ntrue);
for (j = 0; j < nev1; j++) {
idx = eventmap[j];
/* printf("Mapping: Counter %d -> slot %d.\n",j,idx); */
PAPI_get_event_info(events[idx], &info);
PAPI_event_code_to_name(truelist[j], name2);
if (!TESTS_QUIET)
printf("%20s = %16lld %-15s %-15s %s\n", info.short_descr, values[j],
info.symbol, name2, strcmp(info.symbol,
name2) ? "*** MISMATCH ***" : "");
dtmp = (double) values[j];
valsum[idx] += dtmp;
valsample[idx][nsamples[idx]] = dtmp;
nsamples[idx]++;
}
if (!TESTS_QUIET)
printf("\n");
}
if (!TESTS_QUIET) {
printf("\n\nEstimated variance relative to average counts:\n");
for (j = 0; j < nev1; j++)
printf(" Event %.2d", j);
printf("\n");
}
fails = nevents;
/* Due to limited precision of floating point cannot really use
typical standard deviation compuation for large numbers with
very small variations. Instead compute the std devation
problems with precision.
*/
for (j = 0; j < nev1; j++) {
avg[j] = valsum[j] / nsamples[j];
spread[j] = 0;
for (i=0; i < nsamples[j]; ++i) {
double diff = (valsample[j][i] - avg[j]);
spread[j] += diff * diff;
}
spread[j] = sqrt(spread[j] / nsamples[j]) / avg[j];
if (!TESTS_QUIET)
printf("%9.2g ", spread[j]);
/* Make sure that NaN get counted as errors */
if (spread[j] < MPX_TOLERANCE)
fails--;
else if (values[j] < MINCOUNTS) /* Neglect inprecise results with low counts */
fails--;
}
if (!TESTS_QUIET) {
printf("\n\n");
for (j = 0; j < nev1; j++) {
PAPI_get_event_info(events[j], &info);
printf("Event %.2d: mean=%10.0f, sdev/mean=%7.2g nrpt=%2d -- %s\n",
j, avg[j], spread[j], nsamples[j], info.short_descr);
}
printf("\n\n");
}
if (fails)
test_fail(__FILE__, __LINE__, "Values differ from reference", fails);
else
test_pass(__FILE__, NULL, 0);
return 0;
}
int main(int argc, char **argv)
{
PAPI_event_info_t info;
int i, j, retval;
int iters = NUM_FLOPS;
double x = 1.1, y;
long long t1, t2;
long long values[MAXEVENTS], refvalues[MAXEVENTS];
int sleep_time = SLEEPTIME;
double valsqsum[MAXEVENTS];
double valsum[MAXEVENTS];
int nevents = MAXEVENTS;
int eventset = PAPI_NULL;
int events[MAXEVENTS];
events[0] = PAPI_FP_INS;
events[1] = PAPI_TOT_INS;
events[2] = PAPI_INT_INS;
events[3] = PAPI_TOT_CYC;
events[4] = PAPI_STL_CCY;
events[5] = PAPI_BR_INS;
events[6] = PAPI_SR_INS;
events[7] = PAPI_LD_INS;
events[8] = PAPI_TOT_IIS;
events[9] = PAPI_FAD_INS;
events[10] = PAPI_BR_TKN;
events[11] = PAPI_BR_MSP;
events[12] = PAPI_L1_ICA;
events[13] = PAPI_L1_DCA;
for (i = 0; i < MAXEVENTS; i++) {
values[i] = 0;
valsqsum[i] = 0;
valsum[i] = 0;
}
if (argc > 1) {
if (!strcmp(argv[1], "TESTS_QUIET"))
tests_quiet(argc, argv);
else {
sleep_time = atoi(argv[1]);
if (sleep_time <= 0)
sleep_time = SLEEPTIME;
}
}
if (!TESTS_QUIET) {
printf("\nAccuracy check of multiplexing routines.\n");
printf("Comparing a multiplex measurement with separate measurements.\n\n");
}
if ((retval = PAPI_library_init(PAPI_VER_CURRENT)) != PAPI_VER_CURRENT)
test_fail(__FILE__, __LINE__, "PAPI_library_init", retval);
decide_which_events(events, &nevents);
init_multiplex();
/* Find a reasonable number of iterations (each
* event active 20 times) during the measurement
*/
t2 = 10000 * 20 * nevents; /* Target: 10000 usec/multiplex, 20 repeats */
if (t2 > 30e6)
test_skip(__FILE__, __LINE__, "This test takes too much time", retval);
y = dummy3(x, iters);
/* Measure one run */
t1 = PAPI_get_real_usec();
y = dummy3(x, iters);
t1 = PAPI_get_real_usec() - t1;
if (t1 < 1000000) /* Scale up execution time to match t2 */
{
iters = iters * (1000000/t1);
printf("Modified iteration count to %d\n\n",iters);
}
/* Now loop through the items one at a time */
ref_measurements(iters, &eventset, events, nevents, refvalues);
/* Now check multiplexed */
if ((retval = PAPI_create_eventset(&eventset)))
test_fail(__FILE__, __LINE__, "PAPI_create_eventset", retval);
/* In Component PAPI, EventSets must be assigned a component index
before you can fiddle with their internals.
0 is always the cpu component */
retval = PAPI_assign_eventset_component(eventset, 0);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_assign_eventset_component", retval);
if ((retval = PAPI_set_multiplex(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_set_multiplex", retval);
if ((retval = PAPI_add_events(eventset, events, nevents)))
test_fail(__FILE__, __LINE__, "PAPI_add_events", retval);
printf("\nPAPI multiplexed measurements:\n");
x = 1.0;
t1 = PAPI_get_real_usec();
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
y = dummy3(x, iters);
if ((retval = PAPI_stop(eventset, values)))
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
t2 = PAPI_get_real_usec();
for (j = 0; j < nevents; j++) {
PAPI_get_event_info(events[j], &info);
if (!TESTS_QUIET) {
printf("%20s = ", info.short_descr);
printf(LLDFMT, values[j]);
printf("\n");
}
}
check_values(eventset, events, nevents, values, refvalues);
if ((retval = PAPI_remove_events(eventset, events, nevents)))
test_fail(__FILE__, __LINE__, "PAPI_remove_events", retval);
if ((retval = PAPI_cleanup_eventset(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_cleanup_eventset", retval);
if ((retval = PAPI_destroy_eventset(&eventset)))
test_fail(__FILE__, __LINE__, "PAPI_destroy_eventset", retval);
eventset = PAPI_NULL;
/* Now loop through the items one at a time */
ref_measurements(iters, &eventset, events, nevents, refvalues);
check_values(eventset, events, nevents, values, refvalues);
test_pass(__FILE__, NULL, 0);
return 0;
}
int main()
{
double *a;
double *b;
double *c;
int i = 0, j = 0, k = 0;
int *events; // Array of events
long long *values; // Array of values events
int EventSet = PAPI_NULL; // Handle for a PAPI event set as created by PAPI_create_eventset (3)
int retval; // Test fail function
int num_event = 0; // Number of events
int max_event; // Number of available events
int EventCode = 0; // Event code
PAPI_event_info_t pset; // PAPI_event_info_t Struct Reference
char evname[PAPI_MAX_STR_LEN]; // Symbol event
/* Memory asignament to matrixs*/
if((a = (double *)malloc(mrows * ncolumns * sizeof(double))) == NULL)
printf("Error malloc matrix a[%d]\n",mrows * ncolumns);
if((b = (double *)malloc(ncolumns * pcolumns * sizeof(double))) == NULL)
printf("Error malloc matrix b[%d]\n",mrows * ncolumns);
if((c = (double *)malloc(mrows * pcolumns * sizeof(double))) == NULL)
printf("Error malloc matrix c[%d]\n",mrows * ncolumns);
/* Initialize the Matrix arrays */
initmat(a, b, mrows, ncolumns, pcolumns);
/* Initialize the PAPI library */
retval = PAPI_library_init(PAPI_VER_CURRENT);
if (retval != PAPI_VER_CURRENT)
test_fail( __FILE__, __LINE__, "PAPI_library_init", retval );
/* Enable and initialize multiplex support */
retval = PAPI_multiplex_init();
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_multiplex_init", retval );
/* Create an EventSet */
retval = PAPI_create_eventset(&EventSet);
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_create_eventset", retval );
/* Assign it to the CPU component */
retval = PAPI_assign_eventset_component(EventSet, 0);
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_assign_eventset_component", retval );
/* Convert the EventSet to a multiplexed event set */
retval = PAPI_set_multiplex(EventSet);
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_set_multiplex", retval );
/* Obtaining the number of available events */
max_event = PAPI_get_opt( PAPI_MAX_MPX_CTRS, NULL );
printf("\nNumber of available events: %d", max_event );
/* Fill up the event set with as many non-derived events as we can */
EventCode = PAPI_PRESET_MASK;
do {
if ( PAPI_get_event_info( EventCode, &pset ) == PAPI_OK ) {
if ( pset.count && ( strcmp( pset.derived, "NOT_DERIVED" ) == 0 ) ) {
retval = PAPI_add_event( EventSet, ( int ) pset.event_code );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_add_event", retval );
else {
//printf( "Added %s\n", pset.symbol );
num_event++;
}
}
}
} while ( ( PAPI_enum_event( &EventCode, PAPI_PRESET_ENUM_AVAIL ) == PAPI_OK ) && ( num_event < max_event ) );
/* Memory asignament to values and events*/
events = ( int * ) malloc( ( size_t ) num_event * sizeof ( int ) );
if ( events == NULL )
test_fail( __FILE__, __LINE__, "Error malloc events", 0 );
values = ( long long * ) malloc( ( size_t ) num_event * sizeof ( long long ) );
if ( values == NULL )
test_fail( __FILE__, __LINE__, "Erro malloc values", 0 );
/* Start counting events */
if ((retval=PAPI_start(EventSet)) != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
/* Matrix-Matrix multiply */
matmul(a, b, c, mrows, ncolumns, pcolumns);
/* Read the counters */
if ((retval=PAPI_read( EventSet, values )) != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_read_counters", retval);
/* Stop counting events */
if ((retval=PAPI_stop( EventSet, values )) != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_stop_counters", retval);
/* List the events in the event set */
retval = PAPI_list_events( EventSet, events, &num_event );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_list_events", retval );
/* Print results */
printf("\nNumber of non-zero events: %d\n", num_event );
printf( "\nCounts of non-zero available events........................................................\n" );
printf("Name: \t\t\t Value: \t Description:\n");
for ( i = 0; i < num_event; i++ ) {
PAPI_event_code_to_name( events[i], evname ); // Obtaining name of available events
PAPI_get_event_info(events[i], &pset);
if ( values[i] != 0 ) printf("%s \t %15lld \t %s\n", evname, values[i], pset.long_descr);
}
printf( "\nCounts of zero available events............................................................\n" );
printf("Name: \t\t\t Value: \t Description:\n");
for ( i = 0; i < num_event; i++ ) {
PAPI_event_code_to_name( events[i], evname ); // Obtaining name of available events
PAPI_get_event_info(events[i], &pset);
if ( values[i] == 0 ) printf("%s \t %15lld \t %s\n", evname, values[i], pset.long_descr);
}
/* Check if counter pair(s) had identical values */
for ( i = 0; i < num_event; i++ ) {
for ( i = j+1; j < num_event; j++ ) {
if ( ( i != j ) && ( values[i] == values[j] ) ) k++;
}
}
if ( k != 0 ) {
printf( "\nCaution: %d counter pair(s) had identical values\n", k );
}
printf("\n");
/* Free memory */
free( events );
free( values );
free( a );
free( b );
free( c );
/* Cleaning events */
retval = PAPI_cleanup_eventset( EventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_cleanup_eventset", retval );
/* Destroying events */
retval = PAPI_destroy_eventset( &EventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_destroy_eventset", retval );
return 0;
}
int GPTLcreate_and_start_events (const int t) /* thread number */
{
int ret; /* return code */
int n; /* loop index over events */
char eventname[PAPI_MAX_STR_LEN]; /* returned from PAPI_event_code_to_name */
/* Create the event set */
if ((ret = PAPI_create_eventset (&EventSet[t])) != PAPI_OK)
return GPTLerror ("GPTLcreate_and_start_events: thread %d failure creating eventset: %s\n",
t, PAPI_strerror (ret));
if (verbose)
printf ("GPTLcreate_and_start_events: successfully created eventset for thread %d\n", t);
/* Add requested events to the event set */
for (n = 0; n < npapievents; n++) {
if ((ret = PAPI_add_event (EventSet[t], papieventlist[n])) != PAPI_OK) {
if (verbose) {
fprintf (stderr, "%s\n", PAPI_strerror (ret));
ret = PAPI_event_code_to_name (papieventlist[n], eventname);
fprintf (stderr, "GPTLcreate_and_start_events: failure adding event:%s\n",
eventname);
}
if (enable_multiplexing) {
if (verbose)
printf ("Trying multiplexing...\n");
is_multiplexed = true;
break;
} else
return GPTLerror ("enable_multiplexing is false: giving up\n");
}
}
if (is_multiplexed) {
/* Cleanup the eventset for multiplexing */
if ((ret = PAPI_cleanup_eventset (EventSet[t])) != PAPI_OK)
return GPTLerror ("GPTLcreate_and_start_events: %s\n", PAPI_strerror (ret));
if ((ret = PAPI_destroy_eventset (&EventSet[t])) != PAPI_OK)
return GPTLerror ("GPTLcreate_and_start_events: %s\n", PAPI_strerror (ret));
if ((ret = PAPI_create_eventset (&EventSet[t])) != PAPI_OK)
return GPTLerror ("GPTLcreate_and_start_events: failure creating eventset: %s\n",
PAPI_strerror (ret));
if ((ret = PAPI_multiplex_init ()) != PAPI_OK)
return GPTLerror ("GPTLcreate_and_start_events: failure from PAPI_multiplex_init%s\n",
PAPI_strerror (ret));
if ((ret = PAPI_set_multiplex (EventSet[t])) != PAPI_OK)
return GPTLerror ("GPTLcreate_and_start_events: failure from PAPI_set_multiplex: %s\n",
PAPI_strerror (ret));
for (n = 0; n < npapievents; n++) {
if ((ret = PAPI_add_event (EventSet[t], papieventlist[n])) != PAPI_OK) {
ret = PAPI_event_code_to_name (papieventlist[n], eventname);
return GPTLerror ("GPTLcreate_and_start_events: failure adding event:%s\n"
" Error was: %s\n", eventname, PAPI_strerror (ret));
}
}
}
/* Start the event set. It will only be read from now on--never stopped */
if ((ret = PAPI_start (EventSet[t])) != PAPI_OK)
return GPTLerror ("GPTLcreate_and_start_events: failed to start event set: %s\n",
PAPI_strerror (ret));
return 0;
}
int main(int argc, char **argv)
{
PAPI_event_info_t info;
int i, j, retval;
int iters = NUM_FLOPS;
double x = 1.1, y, dtmp;
long long t1, t2;
long long values[MAXEVENTS];
int sleep_time = SLEEPTIME;
#ifdef STARTSTOP
long long dummies[MAXEVENTS];
#endif
double valsample[MAXEVENTS][REPEATS];
double valsum[MAXEVENTS];
double avg[MAXEVENTS];
double spread[MAXEVENTS];
int nevents = MAXEVENTS;
int eventset = PAPI_NULL;
int events[MAXEVENTS];
int fails;
events[0] = PAPI_FP_INS;
events[1] = PAPI_TOT_INS;
events[2] = PAPI_INT_INS;
events[3] = PAPI_TOT_CYC;
events[4] = PAPI_STL_CCY;
events[5] = PAPI_BR_INS;
events[6] = PAPI_SR_INS;
events[7] = PAPI_LD_INS;
events[8] = PAPI_TOT_IIS;
for (i = 0; i < MAXEVENTS; i++) {
values[i] = 0;
valsum[i] = 0;
}
if (argc > 1) {
if (!strcmp(argv[1], "TESTS_QUIET"))
tests_quiet(argc, argv);
else {
sleep_time = atoi(argv[1]);
if (sleep_time <= 0)
sleep_time = SLEEPTIME;
}
}
if (!TESTS_QUIET) {
printf("\nAccuracy check of multiplexing routines.\n");
printf("Investigating the variance of multiplexed measurements.\n\n");
}
if ((retval = PAPI_library_init(PAPI_VER_CURRENT)) != PAPI_VER_CURRENT)
test_fail(__FILE__, __LINE__, "PAPI_library_init", retval);
#ifdef MPX
init_multiplex();
#endif
if ((retval = PAPI_create_eventset(&eventset)))
test_fail(__FILE__, __LINE__, "PAPI_create_eventset", retval);
#ifdef MPX
/* In Component PAPI, EventSets must be assigned a component index
before you can fiddle with their internals.
0 is always the cpu component */
retval = PAPI_assign_eventset_component(eventset, 0);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_assign_eventset_component", retval);
if ((retval = PAPI_set_multiplex(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_set_multiplex", retval);
#endif
nevents = MAXEVENTS;
for (i = 0; i < nevents; i++) {
if ((retval = PAPI_add_event(eventset, events[i]))) {
for (j = i; j < MAXEVENTS; j++)
events[j] = events[j + 1];
nevents--;
i--;
}
}
if (nevents < 2)
test_skip(__FILE__, __LINE__, "Not enough events left...", 0);
/* Find a reasonable number of iterations (each
* event active 20 times) during the measurement
*/
t2 = 10000 * 20 * nevents; /* Target: 10000 usec/multiplex, 20 repeats */
if (t2 > 30e6)
test_skip(__FILE__, __LINE__, "This test takes too much time", retval);
/* Measure one run */
t1 = PAPI_get_real_usec();
y = dummy3(x, iters);
t1 = PAPI_get_real_usec() - t1;
if (t2 > t1) /* Scale up execution time to match t2 */
iters = iters * (int)(t2 / t1);
else if (t1 > 30e6) /* Make sure execution time is < 30s per repeated test */
test_skip(__FILE__, __LINE__, "This test takes too much time", retval);
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
for (i = 1; i <= REPEATS; i++) {
x = 1.0;
#ifndef STARTSTOP
if ((retval = PAPI_reset(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_reset", retval);
#else
if ((retval = PAPI_stop(eventset, dummies)))
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
if ((retval = PAPI_start(eventset)))
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
#endif
if (!TESTS_QUIET)
printf("\nTest %d (of %d):\n", i, REPEATS);
t1 = PAPI_get_real_usec();
y = dummy3(x, iters);
PAPI_read(eventset, values);
t2 = PAPI_get_real_usec();
if (!TESTS_QUIET) {
printf("\n(calculated independent of PAPI)\n");
printf("\tOperations= %.1f Mflop", y * 1e-6);
printf("\t(%g Mflop/s)\n\n", ((float) y / (t2 - t1)));
printf("PAPI measurements:\n");
}
for (j = 0; j < nevents; j++) {
PAPI_get_event_info(events[j], &info);
if (!TESTS_QUIET) {
printf("%20s = ", info.short_descr);
printf(LLDFMT, values[j]);
printf("\n");
}
dtmp = (double) values[j];
valsum[j] += dtmp;
valsample[j][i-1] = dtmp;
}
if (!TESTS_QUIET)
printf("\n");
}
if (!TESTS_QUIET) {
printf("\n\nEstimated variance relative to average counts:\n");
for (j = 0; j < nevents; j++)
printf(" Event %.2d", j);
printf("\n");
}
fails = nevents;
/* Due to limited precision of floating point cannot really use
typical standard deviation compuation for large numbers with
very small variations. Instead compute the std devation
problems with precision.
*/
for (j = 0; j < nevents; j++) {
avg[j] = valsum[j] / REPEATS;
spread[j] = 0;
for (i=0; i < REPEATS; ++i) {
double diff = (valsample[j][i] - avg[j]);
spread[j] += diff * diff;
}
spread[j] = sqrt(spread[j] / REPEATS) / avg[j];
if (!TESTS_QUIET)
printf("%9.2g ", spread[j]);
/* Make sure that NaN get counted as errors */
if (spread[j] < MPX_TOLERANCE)
--fails;
else if (valsum[j] < MINCOUNTS) /* Neglect inprecise results with low counts */
--fails;
}
if (!TESTS_QUIET) {
printf("\n\n");
for (j = 0; j < nevents; j++) {
PAPI_get_event_info(events[j], &info);
printf("Event %.2d: mean=%10.0f, sdev/mean=%7.2g nrpt=%2d -- %s\n",
j, avg[j], spread[j], REPEATS, info.short_descr);
}
printf("\n\n");
}
if (fails)
test_fail(__FILE__, __LINE__, "Values outside threshold", fails);
else
test_pass(__FILE__, NULL, 0);
return 0;
}
/**
* Called by the CBTF collector service in order to start data collection.
*/
void cbtf_collector_start(const CBTF_DataHeader* header)
{
/**
* Start sampling.
*
* Starts hardware counter (HWC) sampling for the thread executing this
* function. Initializes the appropriate thread-local data structures and
* then enables the sampling counter.
*
* @param arguments Encoded function arguments.
*/
/* Create and access our thread-local storage */
#ifdef USE_EXPLICIT_TLS
TLS* tls = malloc(sizeof(TLS));
Assert(tls != NULL);
CBTF_SetTLS(TLSKey, tls);
#else
TLS* tls = &the_tls;
#endif
Assert(tls != NULL);
tls->defer_sampling=false;
#ifndef NDEBUG
IsCollectorDebugEnabled = (getenv("CBTF_DEBUG_COLLECTOR") != NULL);
IsCollectorDetailsDebugEnabled = (getenv("CBTF_DEBUG_COLLECTOR_DETAILS") != NULL);
#if defined (HAVE_OMPT)
IsOMPTDebugEnabled = (getenv("CBTF_DEBUG_COLLECTOR_OMPT") != NULL);
#endif
#endif
/* Decode the passed function arguments */
// Need to handle the arguments...
CBTF_hwcsamp_start_sampling_args args;
memset(&args, 0, sizeof(args));
args.sampling_rate = 100;
/* First set defaults */
int hwcsamp_rate = 100;
char* hwcsamp_papi_event = "PAPI_TOT_CYC,PAPI_TOT_INS";
#if defined (CBTF_SERVICE_USE_OFFLINE)
char* hwcsamp_event_param = getenv("CBTF_HWCSAMP_EVENTS");
if (hwcsamp_event_param != NULL) {
hwcsamp_papi_event=hwcsamp_event_param;
}
const char* sampling_rate = getenv("CBTF_HWCSAMP_RATE");
if (sampling_rate != NULL) {
hwcsamp_rate=atoi(sampling_rate);
}
args.collector = 1;
args.experiment = 0;
tls->data.interval = (uint64_t)(1000000000) / (uint64_t)(hwcsamp_rate);;
#endif
/* Initialize the actual data blob */
memcpy(&tls->header, header, sizeof(CBTF_DataHeader));
initialize_data(tls);
/* We can not assign mpi rank in the header at this point as it may not
* be set yet. assign an integer tid value. omp_tid is used regardless of
* whether the application is using openmp threads.
* libmonitor uses the same numbering scheme as openmp.
*/
tls->header.omp_tid = monitor_get_thread_num();
tls->header.id = strdup(cbtf_collector_unique_id);
tls->header.time_begin = CBTF_GetTime();
#ifndef NDEBUG
if (IsCollectorDebugEnabled) {
fprintf(stderr,"[%ld,%d] ENTER cbtf_collector_start\n",tls->header.pid,tls->header.omp_tid);
}
#endif
if(hwcsamp_papi_init_done == 0) {
#ifndef NDEBUG
if (IsCollectorDebugEnabled) {
fprintf(stderr,"[%ld,%d] cbtf_collector_start: initialize papi\n",tls->header.pid,tls->header.omp_tid);
}
#endif
CBTF_init_papi();
tls->EventSet = PAPI_NULL;
tls->data.clock_mhz = (float) hw_info->mhz;
hwcsamp_papi_init_done = 1;
} else {
tls->data.clock_mhz = (float) hw_info->mhz;
}
/* PAPI SETUP */
CBTF_Create_Eventset(&tls->EventSet);
int rval = PAPI_OK;
#ifndef NDEBUG
if (IsCollectorDebugEnabled) {
fprintf(stderr, "PAPI Version: %d.%d.%d.%d\n", PAPI_VERSION_MAJOR( PAPI_VERSION ),
PAPI_VERSION_MINOR( PAPI_VERSION ),
PAPI_VERSION_REVISION( PAPI_VERSION ),
PAPI_VERSION_INCREMENT( PAPI_VERSION ) );
fprintf(stderr,"System has %d hardware counters.\n", PAPI_num_counters());
}
#endif
/* In Component PAPI, EventSets must be assigned a component index
* before you can fiddle with their internals. 0 is always the cpu component */
#if (PAPI_VERSION_MAJOR(PAPI_VERSION)>=4)
rval = PAPI_assign_eventset_component( tls->EventSet, 0 );
if (rval != PAPI_OK) {
CBTF_PAPIerror(rval,"CBTF_Create_Eventset assign_eventset_component");
return;
}
#endif
/* NOTE: if multiplex is turned on, papi internaly uses a SIGPROF handler.
* Since we are sampling potentially with SIGPROF or now SIGRTMIN and we
* prefer to limit our events to 6, we do not need multiplexing.
*/
if (getenv("CBTF_HWCSAMP_MULTIPLEX") != NULL) {
#if !defined(RUNTIME_PLATFORM_BGP)
rval = PAPI_set_multiplex( tls->EventSet );
if ( rval == PAPI_ENOSUPP) {
fprintf(stderr,"CBTF_Create_Eventset: Multiplex not supported\n");
} else if (rval != PAPI_OK) {
CBTF_PAPIerror(rval,"CBTF_Create_Eventset set_multiplex");
}
#endif
}
/* TODO: check return values of direct PAPI calls
* and handle them as needed.
*/
/* Rework the code here to call PAPI directly rather than
* call any OPENSS helper functions due to inconsitent
* behaviour seen on various lab systems
*/
int eventcode = 0;
rval = PAPI_OK;
if (hwcsamp_papi_event != NULL) {
char *tfptr, *saveptr=NULL, *tf_token;
tfptr = strdup(hwcsamp_papi_event);
for (tf_token = strtok_r(tfptr, ",", &saveptr);
tf_token != NULL;
tf_token = strtok_r(NULL, ",", &saveptr) ) {
PAPI_event_name_to_code(tf_token,&eventcode);
rval = PAPI_add_event(tls->EventSet,eventcode);
if (rval != PAPI_OK) {
CBTF_PAPIerror(rval,"CBTF_Create_Eventset PAPI_event_name_to_code");
}
}
if (tfptr) free(tfptr);
} else {
PAPI_event_name_to_code("PAPI_TOT_CYC",&eventcode);
rval = PAPI_add_event(tls->EventSet,eventcode);
PAPI_event_name_to_code("PAPI_TOT_INS",&eventcode);
rval = PAPI_add_event(tls->EventSet,eventcode);
}
#if defined (HAVE_OMPT)
/* these are ompt specific.*/
/* initialize the flags and counts for idle,wait_barrier. */
tls->thread_idle = tls->thread_wait_barrier = tls->thread_barrier = false;
#endif
/* Begin sampling */
tls->header.time_begin = CBTF_GetTime();
CBTF_Start(tls->EventSet);
CBTF_Timer(tls->data.interval, hwcsampTimerHandler);
}
void *watch_process(void* _pid) {
pid_t pid = *((pid_t*)_pid);
printf("%d\n", pid);
int retval, EventSet = PAPI_NULL;
//char events_name[NB_EVENTS][PAPI_MAX_STR_LEN] = {"UNHALTED_CORE_CYCLES", "INSTRUCTION_RETIRED", "MEM_LOAD_RETIRED:L1D_HIT", "MEM_LOAD_RETIRED:L2_HIT", "LLC_REFERENCES"};
//char events_name[NB_EVENTS][PAPI_MAX_STR_LEN] = {"UNHALTED_CORE_CYCLES", "INSTRUCTION_RETIRED", "MEM_LOAD_RETIRED:L2_HIT", "L2_RQSTS:LOADS", "L2_DATA_RQSTS:ANY"};
//char events_name[NB_EVENTS][PAPI_MAX_STR_LEN] = {"UNHALTED_CORE_CYCLES", "INSTRUCTION_RETIRED", "L1I:READS", "L1D_CACHE_LD:MESI", "L1D_CACHE_ST:MESI"};
//char events_name[NB_EVENTS][PAPI_MAX_STR_LEN] = {"UNHALTED_CORE_CYCLES", "INSTRUCTION_RETIRED", "L1D_CACHE_LD:MESI", "L1D_CACHE_ST:MESI", "L2_DATA_RQSTS:ANY" };
char events_name[NB_EVENTS][PAPI_MAX_STR_LEN] = {"UNHALTED_CORE_CYCLES", "INSTRUCTION_RETIRED", "L1I:READS", "L1D_CACHE_LD:MESI", "L1D_CACHE_ST:MESI", "MEM_LOAD_RETIRED:L2_HIT", "L2_RQSTS:LOADS", "L2_DATA_RQSTS:ANY", "BR_INST_RETIRED:ALL_BRANCHES", "BR_INST_EXEC:ANY", "BR_MISP_EXEC:ANY" };
unsigned int native_events[NB_EVENTS];
long long values[NB_EVENTS];
int err;
/* Initialize the library */
if ( (err = PAPI_library_init(PAPI_VER_CURRENT)) != PAPI_VER_CURRENT) {
fprintf(stderr, "PAPI library init error! %d\n", err);
exit(1);
}
if ( (err = PAPI_multiplex_init()) != PAPI_OK) {
fprintf(stderr, "PAPI multiplex init error! %d\n", err);
exit(1);
}
if (PAPI_thread_init(pthread_self) != PAPI_OK) {
fprintf(stderr, "PAPI thread init error! %d\n", err);
exit(1);
}
//sched_setaffinity(0, 0);
//if ((retval = PAPI_set_granularity(PAPI_GRN_MAX)) != PAPI_OK) {
// fprintf(stderr, "PAPI set granurality error! %d, %d, %d, %d, %d\n", retval, PAPI_EINVAL, PAPI_ENOEVST, PAPI_ENOCMP, PAPI_EISRUN);
// exit(1);
//}
// Traduction des string en code
name_to_code(events_name, NB_EVENTS, native_events);
// Vérification des évènements
check_events(native_events, NB_EVENTS);
if (PAPI_create_eventset(&EventSet) != PAPI_OK) {
fprintf(stderr, "PAPI EventSet init error!\n");
exit(1);
}
if ( (err = PAPI_assign_eventset_component( EventSet, 0 ) ) != PAPI_OK ) {
fprintf(stderr, "PAPI assign component error!\n%s\n", PAPI_strerror(err));
exit(1);
}
if ( ( err = PAPI_set_multiplex(EventSet) ) != PAPI_OK) {
fprintf(stderr, "PAPI set multiplex error!\n%s\n", PAPI_strerror(err));
exit(1);
}
if ( (retval = PAPI_add_events(EventSet, native_events, NB_EVENTS) ) != PAPI_OK) {
fprintf(stderr, "PAPI add events error!\n%s\n", PAPI_strerror(retval) );
exit(1);
}
if (PAPI_attach(EventSet, pid) != PAPI_OK)
exit(1);
while(while_watch) {
/* Start counting */
if (PAPI_start(EventSet) != PAPI_OK) {
fprintf(stderr, "PAPI start error!\n");
exit(1);
}
usleep(1000000);
if (PAPI_stop(EventSet, values) != PAPI_OK) {
fprintf(stderr, "PAPI stop error!\n");
exit(1);
}
print_values(events_name, values, NB_EVENTS);
}
return 0;
}
int main(int argc, char **argv)
{
#if defined(PARALLEL) || defined(PARTIALPAR)
omp_set_num_threads(initParams::nprocs);
size_t threadID;
std::vector<size_t> neighbourThreadId;
std::vector<size_t> workingThreadId;
std::vector<cv::KeyPoint> globalKeypoints;
cv::Mat globalDescriptors;
std::vector<size_t> globalIdx(initParams::nprocs/2, 0);
std::vector<size_t> levelThreadIdx;
#endif
cv::Mat frame;
cv::VideoCapture cap;
class initParams params;
// Set motion field norm threshold. If intra-frame motion greater than this threshold carry out outlier classification, rotation compensation, FOE and TTC computations
#if !defined(PARALLEL)
const float OFThresh = 10.0f;
#endif
#ifdef USE_PAPI
int events[NUM_EVENTS] = {
PAPI_L1_DCM,
PAPI_L2_DCA,
PAPI_L2_DCM
};
long long values[NUM_EVENTS+1];
long long start_time = PAPI_get_real_usec();
double total_time = 0.0f;
#endif
// Initialization of parameters
params.set();
if (argc < 3)
{
std::cerr << "Run Configuration :\n\t./[program name] [input video file] [Total # of frames in video file, 798 for data set-1 and 977 for data set-2]\n";
exit(EXIT_FAILURE);
}
cap.open(argv[1]);
if(!cap.isOpened()){
std::cerr << "Could not access video file: " << argv[1] << std::endl;
exit(EXIT_FAILURE);
}
size_t frameTotal = (size_t) std::atoi(argv[2]);
class agast objExtractor;
class FindRobustFeatures rFeatures(&objExtractor, ¶ms);
cv::Mat tmpDescriptors;
std::vector<cv::KeyPoint> tmpKeypoints;
#ifdef WRITEDATA
analysis::TTCvector.resize((size_t)cap.get(CV_CAP_PROP_FRAME_COUNT), 0.0f);
analysis::kpVector.resize((size_t)cap.get(CV_CAP_PROP_FRAME_COUNT), 0);
analysis::outliersVector.resize((size_t)cap.get(CV_CAP_PROP_FRAME_COUNT), 0);
analysis::matchOutliersVector.resize((size_t)cap.get(CV_CAP_PROP_FRAME_COUNT), 0);
analysis::OFNorm.resize((size_t)cap.get(CV_CAP_PROP_FRAME_COUNT), 0);
analysis::keypointsVector.resize((size_t)cap.get(CV_CAP_PROP_FRAME_COUNT), 0);
#endif
#if defined(PROF) || defined(WRITEDATA)
time_t rawtime;
char buffer[80];
time(&rawtime);
struct tm * timeinfo = localtime(&rawtime);
strftime(buffer,80, "%d-%m-%y_%H-%M-%S", timeinfo);
analysis::timeStamp = std::string(buffer);
#endif
#ifdef PROF
analysis::perfTime.resize((size_t)cap.get(CV_CAP_PROP_FRAME_COUNT), std::vector<float>(3, 0.0f));
#endif
#ifdef USE_PAPI
//********** PAPI Module initialization for performance analysis***********//
// Setup PAPI library
int ret = PAPI_library_init(PAPI_VER_CURRENT);
if (ret != PAPI_VER_CURRENT) {
printf("Problem while initializing PAPI library: %s\n", PAPI_strerror(ret));
exit(EXIT_FAILURE);
}
// Init multiplexer
ret = PAPI_multiplex_init();
if (ret != PAPI_OK) {
printf("Problem while initializing PAPI multiplexer: %s\n", PAPI_strerror(ret));
exit(1);
}
// Create event set for the events that we want to measure
int eventset = PAPI_NULL;
ret = PAPI_create_eventset(&eventset);
if (ret != PAPI_OK) {
printf("Problem while creating PAPI event set: %s\n", PAPI_strerror(ret));
exit(1);
}
ret = PAPI_add_event(eventset, PAPI_L1_DCA);
if (ret != PAPI_OK) {
printf("Problem while adding first PAPI event: %s\n", PAPI_strerror(ret));
exit(1);
}
// set multiplexer for event set
ret = PAPI_set_multiplex(eventset);
if (ret != PAPI_OK) {
printf("Problem while setting PAPI multiplex: %s\n", PAPI_strerror(ret));
exit(1);
}
// Add events to event set
ret = PAPI_add_events(eventset, events, NUM_EVENTS);
if (ret != PAPI_OK) {
printf("Problem while adding PAPI events: %s\n", PAPI_strerror(ret));
exit(1);
}
//********** END INITIALIZATION **********//
// Start measuring computation runtime
start_time = PAPI_get_real_usec();
// Start measuring events in event set
ret = PAPI_start(eventset);
if (ret != PAPI_OK) {
printf("Problem while starting PAPI eventset: %s\n", PAPI_strerror(ret));
exit(1);
}
#endif
try
{
while(( (size_t)cap.get(CV_CAP_PROP_POS_FRAMES) + initParams::frameDelay ) < frameTotal)
{
/****************************PARALLEL-REGION************************************/
#if defined(PARALLEL) || defined(PARTIALPAR)
globalIdx.resize(initParams::nprocs/2, 0);
analysis::refFrameCount = cap.get(CV_CAP_PROP_POS_FRAMES);
analysis::frameCount = cap.get(CV_CAP_PROP_POS_FRAMES);
for (size_t i = 0; i < initParams::nprocs; ++i)
{
if (i % 2 == 0) { workingThreadId.push_back(i); }
else { neighbourThreadId.push_back(i); }
}
#pragma omp parallel num_threads(initParams::nprocs) private(threadID, frame) shared(globalIdx, params, neighbourThreadId, workingThreadId, objExtractor)
{
threadID = omp_get_thread_num();
// get index of the frame to be handled by a particular thread executed only once:
size_t myFrameID = 0;
#pragma omp for ordered schedule(static,1)
for (size_t i = 0; i < initParams::nprocs; ++i)
{
#pragma omp ordered
myFrameID = threadID + analysis::frameCount;
analysis::frameCount++;
}
class agast objExtractorLoc;
while(cap.get(CV_CAP_PROP_POS_FRAMES) != myFrameID) { std::this_thread::sleep_for(std::chrono::milliseconds(5)); }
// get first frame and extract new features:
cap >> objExtractorLoc.frame;
objExtractorLoc.extractFeatures();
if (threadID == initParams::nprocs-1) { objExtractorLoc.frame.copyTo(objExtractor.frame); }
#pragma omp master
{
// keep ref. of the first keypoints and their descriptors:
objExtractor.descriptorsOld = objExtractorLoc.descriptors.clone();
objExtractor.keypointsOld = objExtractorLoc.keypoints;
#ifdef VISUALIZE
objExtractorLoc.frame.copyTo(objExtractor.frameOld);
#endif
}
// follow keypoints over the next framedelay # of frames:
objExtractorLoc.swap();
cap >> objExtractorLoc.frame;
objExtractorLoc.extractFeatures();
objExtractorLoc.featureMatching();
#if ! defined(WRITEDATA)
objExtractorLoc.homographyRANSAC(3.0f);
#else
objExtractorLoc.homographyRANSAC(3.0f, analysis::outliers, analysis::frameCount, analysis::outliersVector);
#endif
if((size_t)objExtractorLoc.goodMatches.size() > 5)
objExtractorLoc.swapGM();
if (!globalIdx.empty())
{
helperFunc::storeGlobal(globalKeypoints, globalDescriptors, objExtractorLoc, threadID, globalIdx, neighbourThreadId);
#pragma omp barrier
helperFunc::recomputeInterThreadGoodMatches(globalKeypoints, globalDescriptors, objExtractorLoc, workingThreadId, globalIdx, threadID, neighbourThreadId);
}
#pragma omp master
{
// re-swaping into global extractors' new points, due to swaping carried out at the end of inter thread matching:
objExtractor.keypoints = objExtractorLoc.keypointsOld;
objExtractor.descriptors = objExtractorLoc.descriptorsOld.clone();
objExtractor.featureMatching();
rFeatures.homographyRANSAC(3.0f, true);
#ifdef VISUALIZE
analysis::frameCount = cap.get(CV_CAP_PROP_POS_FRAMES) - 1;
#endif
}
}
//Compute velocity projection vectors:
rFeatures.compute();
objExtractor.goodMatches.clear();
#ifdef VISUALIZE
//Analyze and verify data if necessary:
if((char)cv::waitKey(cap.get(CV_CAP_PROP_FPS))==27) break;
cv::waitKey(10);
#endif
/****************************SERIAL-REGION**************************************/
#else
//Get first frame and extract new features:
analysis::refFrameCount = cap.get(CV_CAP_PROP_POS_FRAMES);
#if defined(VISUALIZE) || defined(WRITEIMAGE)
cv::Mat refFrame;
#endif
cap >> objExtractor.frame;
#if defined(VISUALIZE) || defined(WRITEIMAGE)
objExtractor.frame.copyTo(refFrame);
#endif
objExtractor.extractFeatures();
#ifdef PROF
analysis::perfTime[analysis::frameCount][2] = objExtractor.elapsedTime.count();
#endif
// Keep ref. of the first keypoints and their descriptors:
objExtractor.descriptors.copyTo(tmpDescriptors);
tmpKeypoints = objExtractor.keypoints;
// Specify initial features as old features:
objExtractor.swap();
// Follow keypoints over the next frames:
while((size_t)cap.get(CV_CAP_PROP_POS_FRAMES) < frameTotal )
{
analysis::frameCount = cap.get(CV_CAP_PROP_POS_FRAMES);
analysis::outliers = 0;
cap >> objExtractor.frame;
objExtractor.extractFeatures();
#ifdef PROF
analysis::perfTime[analysis::frameCount][2] = objExtractor.elapsedTime.count();
analysis::start = std::chrono::high_resolution_clock::now();
#endif
// FLANN based matching of keypoints:
objExtractor.featureMatching();
#ifdef PROF
analysis::stop = std::chrono::high_resolution_clock::now();
analysis::elapsedTime = analysis::stop - analysis::start;
analysis::perfTime[analysis::frameCount][0] = analysis::elapsedTime.count();
#endif
#ifdef VISUALIZE
analysis::visMatchingPts(objExtractor);
cv::waitKey(5);
#endif
#ifdef WRITEDATA
analysis::keypointsVector[analysis::frameCount] = std::min((size_t)objExtractor.keypoints.size(), (size_t)objExtractor.keypointsOld.size());
analysis::matchOutliersVector[analysis::frameCount] = analysis::keypointsVector[analysis::frameCount] - (size_t)objExtractor.goodMatches.size();
#endif
#ifdef PROF
analysis::start = std::chrono::high_resolution_clock::now();
#endif
// RANSAC based outlier classification:
rFeatures.homographyRANSAC(1.2f, true);
#ifdef PROF
analysis::stop = std::chrono::high_resolution_clock::now();
analysis::elapsedTime = analysis::stop - analysis::start;
analysis::perfTime[analysis::frameCount][1] = analysis::elapsedTime.count();
#endif
#ifdef VISUALIZE
analysis::visMatchingPts(objExtractor);
cv::waitKey(5);
#endif
float OF = 0.0f;
// Calculation of motion field norm makes sense only if matching points are greater than 3
if ((size_t)objExtractor.goodMatches.size() > 3)
{
// Calculate intra-frame motion flow norm
OF = helperFunc::calcMeanOFNorm(objExtractor.keypoints, objExtractor.keypointsOld, objExtractor.goodMatches);
}
#ifdef WRITEDATA
analysis::OFNorm[analysis::frameCount] = helperFunc::calcMeanOFNorm(objExtractor.keypoints, objExtractor.keypointsOld, objExtractor.goodMatches);
#endif
// Different cases of computation:
if ((size_t)objExtractor.goodMatches.size() < 5 || OF >= OFThresh)
{
// Case 1: Matching(refFrame(n) , refFrame(n+1)) lead to no good matches
if (analysis::refFrameCount == (analysis::frameCount - 1) )
{
// If large motion field norm was the case then compute ttc:
if (OF >= OFThresh)
{
//Compute velocity projection vectors and ttc:
rFeatures.compute();
}
// Define latest frame as refFrame
analysis::refFrameCount = analysis::frameCount;
objExtractor.descriptors.copyTo(tmpDescriptors);
tmpKeypoints = objExtractor.keypoints;
#if defined(VISUALIZE) || defined(WRITEIMAGE)
refFrame = objExtractor.frame.clone();
#endif
// Move latest frame data to old
objExtractor.swap();
}
// Case 2: Matching(refFrame(n+i) , refFrame(n+i+1)) lead to no good matches
else
{
// In case of motion flow greater than threshold compute ttc for the given intra-frame motion otherwise move consistent features and compute motion flow
if (OF >= OFThresh)
{
// Compute velocity projection vectors:
rFeatures.compute();
// Define latest frame as refFrame
analysis::refFrameCount = analysis::frameCount;
objExtractor.descriptors.copyTo(tmpDescriptors);
tmpKeypoints = objExtractor.keypoints;
#if defined(VISUALIZE) || defined(WRITEIMAGE)
refFrame = objExtractor.frame.clone();
#endif
// Move latest frame data to old
objExtractor.swap();
}
else
{
// Move consistent good matching features to new
objExtractor.keypoints.swap(objExtractor.keypointsOld);
objExtractor.keypointsOld.swap(tmpKeypoints);
cv::Mat placeHolder(objExtractor.descriptors.size(), objExtractor.descriptors.type());
#ifdef OPTIMIZED
objExtractor.descriptors.copyTo(placeHolder);
#else
placeHolder = objExtractor.descriptors.clone();
#endif
objExtractor.descriptors.resize(0);
objExtractor.descriptorsOld.copyTo(objExtractor.descriptors);
tmpDescriptors.copyTo(objExtractor.descriptorsOld );
tmpDescriptors.resize(0);
#ifdef OPTIMIZED
placeHolder.copyTo(tmpDescriptors);
#else
tmpDescriptors = placeHolder.clone();
#endif
placeHolder.release();
#if defined(VISUALIZE) || defined(WRITEIMAGE)
cv::Mat tmpFrame;
tmpFrame = objExtractor.frame.clone();
objExtractor.frame = objExtractor.frameOld.clone();
objExtractor.frameOld = refFrame.clone();
refFrame = tmpFrame.clone();
tmpFrame.release();
#endif
// Match refFrame features with consistent good matching features uptil refFrame(n+i)
objExtractor.featureMatching();
#ifdef VISUALIZE
analysis::visMatchingPts(objExtractor, "Using Consistent Features");
cv::waitKey(10);
#endif
// Calculate Motion Flow norm
OF = helperFunc::calcMeanOFNorm(objExtractor.keypoints, objExtractor.keypointsOld, objExtractor.goodMatches);
if( OF > OFThresh )
{
// Compute motion flow and ttc:
rFeatures.compute();
}
objExtractor.goodMatches.clear();
objExtractor.keypointsOld = tmpKeypoints;
tmpDescriptors.copyTo(objExtractor.descriptorsOld);
// Define latest frame as refFrame
analysis::refFrameCount = analysis::frameCount;
}
}
}
else { objExtractor.swapGM(); }
#ifdef VISUALIZE
if((char)cv::waitKey(cap.get(CV_CAP_PROP_FPS))==27) break;
cv::waitKey(10);
#endif
}
#endif
}
}
catch(cv::Exception) { std::cerr << "Exception @ frame : " << cap.get(CV_CAP_PROP_POS_FRAMES) << "\n"; }
#ifdef USE_PAPI
// Collect data from the event set
if ( PAPI_stop(eventset, values) != PAPI_OK ) {
printf("Problem while reading PAPI event set!\n");
exit(EXIT_FAILURE);
}
// Print results
total_time = ((double)(PAPI_get_real_usec() - start_time))/1000000;
printf("Computation execution time: %lf seconds\n", total_time);
printf("L1 Cache miss rate = %lf \n",((double)values[1])/(values[0]));
printf("L2 Cache miss rate = %lf \n",((double)values[3])/(values[2]));
#endif
/*************************WRITE DATA TO FILE BLOCK********************************/
/*<---------------------------- REGQUIRES SETTING OF PERSONAL PATH TO SAVING GENERATED DATA -------------------------->*/
#ifdef WRITEDATA
std::ofstream file;
std::string fileName;
fileName = "/home/ragesam/gitThesis/testCase/poseFeatures/build/ttc_" + analysis::timeStamp + ".dat";
std::cout << "\nWRITING TTC DATA TO FILE \n";
file.open(fileName);
if(!file.is_open())
{
std::cout << "\nCan not write TTC to file\n";
std::exit(EXIT_FAILURE);
}
else
{
for (size_t i = 0; i < (size_t)analysis::TTCvector.size(); ++i)
{
file << i << "\t" << analysis::TTCvector[i] << "\t" << analysis::kpVector[i] <<"\n";
}
}
file.close();
fileName = "/home/ragesam/gitThesis/testCase/poseFeatures/build/outliers_" + analysis::timeStamp + ".dat";
std::cout << "\nWRITING OUTLIER DATA TO FILE \n";
file.open(fileName);
if(!file.is_open())
{
std::cout << "\nCan not write outliers to file\n";
std::exit(EXIT_FAILURE);
}
else
{
file << "Frame #\t# of Keypoints\tFLANN based outliers\tRANSAC based outliers\tIntra-fram projection vector norm\n";
for (size_t i = 0; i < (size_t)analysis::outliersVector.size(); ++i)
{
file << i << "\t" << analysis::keypointsVector[i] << "\t" << analysis::matchOutliersVector[i] << "\t" << analysis::outliersVector[i] << "\t" << analysis::OFNorm[i] <<"\n";
}
}
file.close();
#endif
#ifdef PROF
std::ofstream fileProf;
std::string fileNameProf;
fileNameProf = "/home/ragesam/gitThesis/testCase/poseFeatures/build/profile_" + analysis::timeStamp + ".dat";
std::cout << "\nWRITING PROFILING DATA TO FILE \n";
fileProf.open(fileNameProf);
if(!fileProf.is_open())
{
std::cout << "\nCan not write profiling data to file\n";
std::exit(EXIT_FAILURE);
}
else
{
fileProf << "Frame #\telapsedTime Matching\telapsedTime RANSAC\telapsedTimeDetection\n";
for (size_t i = 0; i < (size_t)analysis::perfTime.size(); ++i)
{
fileProf << i << "\t" << analysis::perfTime[i][0] << "\t" << analysis::perfTime[i][1] << "\t" << analysis::perfTime[i][2] << "\n";
}
}
fileProf.close();
#endif
cap.release();
return 0;
}
int case1(void)
{
int retval, i, EventSet = PAPI_NULL, j = 0, k = 0, allvalid = 1;
int max_mux, nev, *events;
long long *values;
PAPI_event_info_t pset;
char evname[PAPI_MAX_STR_LEN];
init_papi();
init_multiplex();
retval = PAPI_create_eventset(&EventSet);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_create_eventset", retval);
/* In Component PAPI, EventSets must be assigned a component index
before you can fiddle with their internals.
0 is always the cpu component */
retval = PAPI_assign_eventset_component(EventSet, 0);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_assign_eventset_component", retval);
retval = PAPI_set_multiplex(EventSet);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_set_multiplex", retval);
max_mux = PAPI_get_opt(PAPI_MAX_MPX_CTRS, NULL);
if (max_mux > 32) max_mux = 32;
/* Fill up the event set with as many non-derived events as we can */
printf("\nFilling the event set with as many non-derived events as we can...\n");
i = PAPI_PRESET_MASK;
do {
if (PAPI_get_event_info(i, &pset) == PAPI_OK)
{
if (pset.count && (strcmp(pset.derived,"NOT_DERIVED") == 0))
{
retval = PAPI_add_event(EventSet, pset.event_code);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_add_event", retval);
else
{
printf("Added %s\n", pset.symbol);
j++;
}
}
}
} while ((PAPI_enum_event(&i, PAPI_PRESET_ENUM_AVAIL) == PAPI_OK) && (j < max_mux));
events = (int *) malloc(j * sizeof(int));
if (events == NULL)
test_fail(__FILE__, __LINE__, "malloc events", 0);
values = (long long *) malloc(j * sizeof(long long));
if (values == NULL)
test_fail(__FILE__, __LINE__, "malloc values", 0);
do_stuff();
if (PAPI_start(EventSet) != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_start", retval);
do_stuff();
retval = PAPI_stop(EventSet, values);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_stop", retval);
nev =j;
retval = PAPI_list_events(EventSet, events, &nev);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_list_events", retval);
printf("\nEvent Counts:\n");
for (i = 0, allvalid = 0; i < j; i++) {
PAPI_event_code_to_name(events[i], evname);
printf(TAB1, evname, values[i]);
if (values[i] == 0)
allvalid++;
}
printf("\n");
if (allvalid){
printf("Caution: %d counters had zero values\n", allvalid);
}
for (i = 0, allvalid = 0; i < j; i++) {
for (k = i+1; k < j; k++) {
if ((i != k) && (values[i] == values[k]))
{
allvalid++;
break;
}
}
}
if (allvalid){
printf("Caution: %d counter pair(s) had identical values\n", allvalid);
}
free(events);
free(values);
retval = PAPI_cleanup_eventset(EventSet);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_cleanup_eventset", retval);
retval = PAPI_destroy_eventset(&EventSet);
if (retval != PAPI_OK)
test_fail(__FILE__, __LINE__, "PAPI_destroy_eventset", retval);
return (SUCCESS);
}
void
mainloop( int arg )
{
int allvalid;
long long *values;
int EventSet = PAPI_NULL;
int retval, i, j = 2, skipped_counters=0;
PAPI_event_info_t pset;
( void ) arg;
/* Initialize the library */
retval = PAPI_library_init( PAPI_VER_CURRENT );
if ( retval != PAPI_VER_CURRENT )
test_fail( __FILE__, __LINE__, "PAPI_library_init", retval );
init_multiplex( );
retval = PAPI_create_eventset( &EventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_create_eventset", retval );
/* In Component PAPI, EventSets must be assigned a component index
before you can fiddle with their internals.
0 is always the cpu component */
retval = PAPI_assign_eventset_component( EventSet, 0 );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_assign_eventset_component",
retval );
retval = PAPI_set_multiplex( EventSet );
if ( retval == PAPI_ENOSUPP) {
test_skip(__FILE__, __LINE__, "Multiplex not supported", 1);
} else if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_set_multiplex", retval );
if ( ( retval =
PAPI_thread_init( ( unsigned
long ( * )( void ) ) ( pthread_self ) ) ) !=
PAPI_OK ) {
if ( retval == PAPI_ECMP )
test_skip( __FILE__, __LINE__, "PAPI_thread_init", retval );
else
test_fail( __FILE__, __LINE__, "PAPI_thread_init", retval );
}
retval = PAPI_add_event( EventSet, PAPI_TOT_INS );
if ( ( retval != PAPI_OK ) && ( retval != PAPI_ECNFLCT ) )
test_fail( __FILE__, __LINE__, "PAPI_add_event", retval );
if ( !TESTS_QUIET ) {
printf( "Added %s\n", "PAPI_TOT_INS" );
}
retval = PAPI_add_event( EventSet, PAPI_TOT_CYC );
if ( ( retval != PAPI_OK ) && ( retval != PAPI_ECNFLCT ) )
test_fail( __FILE__, __LINE__, "PAPI_add_event", retval );
if ( !TESTS_QUIET ) {
printf( "Added %s\n", "PAPI_TOT_CYC" );
}
values = ( long long * ) malloc( MAX_TO_ADD * sizeof ( long long ) );
if ( values == NULL )
test_fail( __FILE__, __LINE__, "malloc", 0 );
for ( i = 0; i < PAPI_MAX_PRESET_EVENTS; i++ ) {
retval = PAPI_get_event_info( i | PAPI_PRESET_MASK, &pset );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_get_event_info", retval );
if ( pset.count ) {
printf( "Adding %s\n", pset.symbol );
retval = PAPI_add_event( EventSet, ( int ) pset.event_code );
if ( ( retval != PAPI_OK ) && ( retval != PAPI_ECNFLCT ) )
test_fail( __FILE__, __LINE__, "PAPI_add_event", retval );
if ( retval == PAPI_OK ) {
printf( "Added %s\n", pset.symbol );
} else {
printf( "Could not add %s\n", pset.symbol );
}
do_stuff( );
if ( retval == PAPI_OK ) {
retval = PAPI_start( EventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_start", retval );
do_stuff( );
retval = PAPI_stop( EventSet, values );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_stop", retval );
if ( values[j] ) {
if ( ++j >= MAX_TO_ADD )
break;
} else {
retval =
PAPI_remove_event( EventSet, ( int ) pset.event_code );
if ( retval == PAPI_OK )
printf( "Removed %s\n", pset.symbol );
/* This added because the test */
/* can take a long time if mplexing */
/* is broken and all values are 0 */
skipped_counters++;
if (skipped_counters>MAX_TO_ADD) break;
}
}
}
}
retval = PAPI_start( EventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_start", retval );
do_stuff( );
retval = PAPI_stop( EventSet, values );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_stop", retval );
test_print_event_header( "multiplex3_pthreads:\n", EventSet );
allvalid = 0;
for ( i = 0; i < MAX_TO_ADD; i++ ) {
printf( ONENUM, values[i] );
if ( values[i] != 0 )
allvalid++;
}
printf( "\n" );
if ( !allvalid )
test_fail( __FILE__, __LINE__, "all counter registered no counts", 1 );
retval = PAPI_cleanup_eventset( EventSet ); /* JT */
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_cleanup_eventset", retval );
retval = PAPI_destroy_eventset( &EventSet );
if ( retval != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_destroy_eventset", retval );
free( values );
PAPI_shutdown( );
}
int main(int argc, char **argv) {
int retval,max_multiplex,i,EventSet=PAPI_NULL;
PAPI_event_info_t info;
int added=0;
int events_tried=0;
/* Set TESTS_QUIET variable */
tests_quiet( argc, argv );
/* Initialize the library */
retval = PAPI_library_init( PAPI_VER_CURRENT );
if ( retval != PAPI_VER_CURRENT ) {
test_fail( __FILE__, __LINE__, "PAPI_library_init", retval );
}
retval = PAPI_multiplex_init( );
if ( retval != PAPI_OK) {
test_fail(__FILE__, __LINE__, "Multiplex not supported", 1);
}
max_multiplex=PAPI_get_opt( PAPI_MAX_MPX_CTRS, NULL );
if (!TESTS_QUIET) {
printf("Maximum multiplexed counters=%d\n",max_multiplex);
}
if (!TESTS_QUIET) {
printf("Trying to multiplex as many as possible:\n");
}
retval = PAPI_create_eventset( &EventSet );
if ( retval != PAPI_OK ) {
test_fail(__FILE__, __LINE__, "PAPI_create_eventset", retval );
}
retval = PAPI_assign_eventset_component( EventSet, 0 );
if ( retval != PAPI_OK ) {
test_fail(__FILE__, __LINE__, "PAPI_assign_eventset_component",
retval );
}
retval = PAPI_set_multiplex( EventSet );
if ( retval != PAPI_OK ) {
test_fail(__FILE__, __LINE__, "PAPI_create_multiplex", retval );
}
i = 0 | PAPI_PRESET_MASK;
PAPI_enum_event( &i, PAPI_ENUM_FIRST );
do {
retval = PAPI_get_event_info( i, &info );
if (retval==PAPI_OK) {
if (!TESTS_QUIET) printf("Adding %s: ",info.symbol);
}
retval = PAPI_add_event( EventSet, info.event_code );
if (retval!=PAPI_OK) {
if (!TESTS_QUIET) printf("Fail!\n");
}
else {
if (!TESTS_QUIET) printf("Success!\n");
added++;
}
events_tried++;
} while (PAPI_enum_event( &i, PAPI_PRESET_ENUM_AVAIL ) == PAPI_OK );
PAPI_shutdown( );
if (!TESTS_QUIET) {
printf("Added %d of theoretical max %d\n",added,max_multiplex);
}
if (events_tried<max_multiplex) {
if (!TESTS_QUIET) {
printf("Ran out of events before we ran out of room\n");
}
}
else if (added!=max_multiplex) {
test_fail(__FILE__, __LINE__,
"Couldn't max out multiplexed events", 1);
}
test_pass( __FILE__, NULL, 0 );
exit( 0 );
}
/*
** GPTLcreate_and_start_events: Create and start the PAPI eventset.
** Threaded routine to create the "event set" (PAPI terminology) and start
** the counters. This is only done once, and is called from get_thread_num
** for the first time for the thread.
**
** Input args:
** t: thread number
**
** Return value: 0 (success) or GPTLerror (failure)
*/
int GPTLcreate_and_start_events (const int t) /* thread number */
{
int ret; /* return code */
int n; /* loop index over events */
char eventname[PAPI_MAX_STR_LEN]; /* returned from PAPI_event_code_to_name */
static const char *thisfunc = "GPTLcreate_and_start_events";
/*
** Set the domain to count all contexts. Only needs to be set once for all threads
*/
if ((ret = PAPI_set_domain (PAPI_DOM_ALL)) != PAPI_OK)
return GPTLerror ("%s: thread %d failure setting PAPI domain: %s\n",
thisfunc, t, PAPI_strerror (ret));
/* Create the event set */
if ((ret = PAPI_create_eventset (&EventSet[t])) != PAPI_OK)
return GPTLerror ("%s: thread %d failure creating eventset: %s\n",
thisfunc, t, PAPI_strerror (ret));
if (verbose)
printf ("%s: successfully created eventset for thread %d\n", thisfunc, t);
/* Add requested events to the event set */
for (n = 0; n < npapievents; n++) {
if ((ret = PAPI_add_event (EventSet[t], papieventlist[n])) != PAPI_OK) {
if (verbose) {
fprintf (stderr, "%s\n", PAPI_strerror (ret));
ret = PAPI_event_code_to_name (papieventlist[n], eventname);
fprintf (stderr, "%s: failure adding event:%s\n", thisfunc, eventname);
}
if (enable_multiplexing) {
if (verbose)
printf ("Trying multiplexing...\n");
is_multiplexed = true;
break;
} else
return GPTLerror ("enable_multiplexing is false: giving up\n");
}
}
if (is_multiplexed) {
/* Cleanup the eventset for multiplexing */
if ((ret = PAPI_cleanup_eventset (EventSet[t])) != PAPI_OK)
return GPTLerror ("%s: %s\n", thisfunc, PAPI_strerror (ret));
if ((ret = PAPI_destroy_eventset (&EventSet[t])) != PAPI_OK)
return GPTLerror ("%s: %s\n", thisfunc, PAPI_strerror (ret));
if ((ret = PAPI_create_eventset (&EventSet[t])) != PAPI_OK)
return GPTLerror ("%s: failure creating eventset: %s\n", thisfunc, PAPI_strerror (ret));
/*
** Assign EventSet to component 0 (cpu). This step is MANDATORY in recent PAPI releases
** in order to enable event multiplexing
*/
if ((ret = PAPI_assign_eventset_component (EventSet[t], 0)) != PAPI_OK)
return GPTLerror ("%s: thread %d failure in PAPI_assign_eventset_component: %s\n",
thisfunc, t, PAPI_strerror (ret));
if ((ret = PAPI_multiplex_init ()) != PAPI_OK)
return GPTLerror ("%s: failure from PAPI_multiplex_init%s\n", thisfunc, PAPI_strerror (ret));
if ((ret = PAPI_set_multiplex (EventSet[t])) != PAPI_OK)
return GPTLerror ("%s: failure from PAPI_set_multiplex: %s\n", thisfunc, PAPI_strerror (ret));
for (n = 0; n < npapievents; n++) {
if ((ret = PAPI_add_event (EventSet[t], papieventlist[n])) != PAPI_OK) {
ret = PAPI_event_code_to_name (papieventlist[n], eventname);
return GPTLerror ("%s: failure adding event:%s Error was: %s\n",
thisfunc, eventname, PAPI_strerror (ret));
}
}
}
/* Start the event set. It will only be read from now on--never stopped */
if ((ret = PAPI_start (EventSet[t])) != PAPI_OK)
return GPTLerror ("%s: failed to start event set: %s\n", thisfunc, PAPI_strerror (ret));
return 0;
}
void *
thread( void *arg )
{
( void ) arg; /*unused */
int eventset = PAPI_NULL;
long long values[PAPI_MPX_DEF_DEG];
int ret = PAPI_register_thread( );
if ( ret != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_register_thread", ret );
ret = PAPI_create_eventset( &eventset );
if ( ret != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_create_eventset", ret );
printf( "Event set %d created\n", eventset );
/* In Component PAPI, EventSets must be assigned a component index
before you can fiddle with their internals.
0 is always the cpu component */
ret = PAPI_assign_eventset_component( eventset, 0 );
if ( ret != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_assign_eventset_component", ret );
}
ret = PAPI_set_multiplex( eventset );
if ( ret == PAPI_ENOSUPP) {
test_skip( __FILE__, __LINE__, "Multiplexing not supported", 1 );
}
else if ( ret != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_set_multiplex", ret );
}
ret = PAPI_add_events( eventset, events, numevents );
if ( ret < PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_add_events", ret );
}
ret = PAPI_start( eventset );
if ( ret != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_start", ret );
}
do_stuff( );
ret = PAPI_stop( eventset, values );
if ( ret != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_stop", ret );
}
ret = PAPI_cleanup_eventset( eventset );
if ( ret != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_cleanup_eventset", ret );
}
ret = PAPI_destroy_eventset( &eventset );
if ( ret != PAPI_OK ) {
test_fail( __FILE__, __LINE__, "PAPI_destroy_eventset", ret );
}
ret = PAPI_unregister_thread( );
if ( ret != PAPI_OK )
test_fail( __FILE__, __LINE__, "PAPI_unregister_thread", ret );
return ( NULL );
}
static int
refresh_metrics(int log)
{
int sts = 0;
int state = 0;
int i;
int number_of_active_counters = 0;
time_t now;
now = time(NULL);
/* Shut down, save previous state. */
state = check_papi_state();
if (state & PAPI_RUNNING) {
sts = PAPI_stop(EventSet, values);
if (sts != PAPI_OK) {
/* futile to continue */
handle_papi_error(sts, log);
return PM_ERR_VALUE;
}
/* Save previous values */
for (i = 0; i < number_of_events; i++){
if(papi_info[i].position >= 0) {
papi_info[i].prev_value += values[papi_info[i].position];
papi_info[i].position = -1;
}
}
/* Clean up eventset */
sts = PAPI_cleanup_eventset(EventSet);
if (sts != PAPI_OK) {
handle_papi_error(sts, log);
/* FALLTHROUGH */
}
sts = PAPI_destroy_eventset(&EventSet); /* sets EventSet=NULL */
if (sts != PAPI_OK) {
handle_papi_error(sts, log);
/* FALLTHROUGH */
}
}
/* Initialize new EventSet */
EventSet = PAPI_NULL;
if ((sts = PAPI_create_eventset(&EventSet)) != PAPI_OK) {
handle_papi_error(sts, log);
return PM_ERR_GENERIC;
}
if ((sts = PAPI_assign_eventset_component(EventSet, 0 /*CPU*/)) != PAPI_OK) {
handle_papi_error(sts, log);
return PM_ERR_GENERIC;
}
if (enable_multiplexing && (sts = PAPI_set_multiplex(EventSet)) != PAPI_OK) {
handle_papi_error(sts, log);
/* not fatal - FALLTHROUGH */
}
/* Add all survivor events to new EventSet */
number_of_active_counters = 0;
for (i = 0; i < number_of_events; i++) {
if (papi_info[i].metric_enabled == METRIC_ENABLED_FOREVER ||
papi_info[i].metric_enabled >= now) {
sts = PAPI_add_event(EventSet, papi_info[i].info.event_code);
if (sts != PAPI_OK) {
if (pmDebug & DBG_TRACE_APPL0) {
char eventname[PAPI_MAX_STR_LEN];
PAPI_event_code_to_name(papi_info[i].info.event_code, eventname);
__pmNotifyErr(LOG_DEBUG, "Unable to add: %s due to error: %s\n",
eventname, PAPI_strerror(sts));
}
handle_papi_error(sts, log);
/*
* We may be called to make drastic changes to the PAPI
* eventset. Partial successes/failures are quite reasonable,
* as a user may be asking to activate a mix of good and bad
* counters. Diagnosing the partial failure is left to the
* caller, by examining the papi_info[i].position value
* after this function returns. */
continue;
}
papi_info[i].position = number_of_active_counters++;
}
}
/* Restart counting. */
if (number_of_active_counters > 0) {
sts = PAPI_start(EventSet);
if (sts != PAPI_OK) {
handle_papi_error(sts, log);
return PM_ERR_VALUE;
}
}
return 0;
}
