int
main( int argc, char **argv )
{
int i;
tests_quiet( argc, argv ); /* Set TESTS_QUIET variable */
// if ( ( retval =
// PAPI_library_init( PAPI_VER_CURRENT ) ) != PAPI_VER_CURRENT )
// test_fail( __FILE__, __LINE__, "PAPI_library_init", retval );
printf( "\n----------------------------------\n" );
printf( "For 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 ) );
printf( "There are %d error codes defined\n", PAPI_NUM_ERRORS );
printf( "----------------------------------\n" );
for (i = 0; i < PAPI_NUM_ERRORS; i++) {
char *errstr, *errdescr;
errstr = PAPI_strerror( -i );
errdescr = PAPI_descr_error( -i );
printf( "Error code %4d: %-15s | %s\n", -i, errstr, errdescr );
}
printf( "----------------------------------\n\n" );
exit( 1 );
}
/* Support routine to display header information to the screen
from the hardware info data structure. The same code was duplicated
in a number of tests and utilities. Seems to make sense to refactor.
This may not be the best place for it to live, but it works for now.
*/
int
papi_print_header( char *prompt, const PAPI_hw_info_t ** hwinfo )
{
if ( ( *hwinfo = PAPI_get_hardware_info( ) ) == NULL )
return ( PAPI_ESBSTR );
printf( "%s", prompt );
printf
( "--------------------------------------------------------------------------------\n" );
printf( "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 ) );
printf( "Vendor string and code : %s (%d)\n", ( *hwinfo )->vendor_string,
( *hwinfo )->vendor );
printf( "Model string and code : %s (%d)\n", ( *hwinfo )->model_string,
( *hwinfo )->model );
printf( "CPU Revision : %f\n", ( *hwinfo )->revision );
if ( ( *hwinfo )->cpuid_family > 0 )
printf
( "CPUID Info : Family: %d Model: %d Stepping: %d\n",
( *hwinfo )->cpuid_family, ( *hwinfo )->cpuid_model,
( *hwinfo )->cpuid_stepping );
printf( "CPU Megahertz : %f\n", ( *hwinfo )->mhz );
printf( "CPU Clock Megahertz : %d\n", ( *hwinfo )->clock_mhz );
if ( ( *hwinfo )->threads > 0 )
printf( "Hdw Threads per core : %d\n", ( *hwinfo )->threads );
if ( ( *hwinfo )->cores > 0 )
printf( "Cores per Socket : %d\n", ( *hwinfo )->cores );
if ( ( *hwinfo )->sockets > 0 )
printf( "Sockets : %d\n", ( *hwinfo )->sockets );
if ( ( *hwinfo )->nnodes > 0 )
printf( "NUMA Nodes : %d\n", ( *hwinfo )->nnodes );
printf( "CPUs per Node : %d\n", ( *hwinfo )->ncpu );
printf( "Total CPUs : %d\n", ( *hwinfo )->totalcpus );
printf( "Running in a VM : %s\n", ( *hwinfo )->virtualized?
"yes":"no");
if ( (*hwinfo)->virtualized) {
printf( "VM Vendor: : %s\n", (*hwinfo)->virtual_vendor_string);
}
printf( "Number Hardware Counters : %d\n",
PAPI_get_opt( PAPI_MAX_HWCTRS, NULL ) );
printf( "Max Multiplex Counters : %d\n",
PAPI_get_opt( PAPI_MAX_MPX_CTRS, NULL ) );
printf
( "--------------------------------------------------------------------------------\n" );
printf( "\n" );
return PAPI_OK;
}
/* Support routine to display header information to the screen
from the hardware info data structure. The same code was duplicated
in a number of tests and utilities. Seems to make sense to refactor.
This may not be the best place for it to live, but it works for now.
*/
int
papi_print_header( char *prompt, int event_flag,
const PAPI_hw_info_t ** hwinfo )
{
if ( ( *hwinfo = PAPI_get_hardware_info( ) ) == NULL )
return ( PAPI_ESBSTR );
printf( "%s", prompt );
printf
( "--------------------------------------------------------------------------------\n" );
printf( "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 ) );
printf( "Vendor string and code : %s (%d)\n", ( *hwinfo )->vendor_string,
( *hwinfo )->vendor );
printf( "Model string and code : %s (%d)\n", ( *hwinfo )->model_string,
( *hwinfo )->model );
printf( "CPU Revision : %f\n", ( *hwinfo )->revision );
if ( ( *hwinfo )->cpuid_family > 0 )
printf
( "CPUID Info : Family: %d Model: %d Stepping: %d\n",
( *hwinfo )->cpuid_family, ( *hwinfo )->cpuid_model,
( *hwinfo )->cpuid_stepping );
printf( "CPU Megahertz : %f\n", ( *hwinfo )->mhz );
printf( "CPU Clock Megahertz : %d\n", ( *hwinfo )->clock_mhz );
if ( ( *hwinfo )->threads > 0 )
printf( "Hdw Threads per core : %d\n", ( *hwinfo )->threads );
if ( ( *hwinfo )->cores > 0 )
printf( "Cores per Socket : %d\n", ( *hwinfo )->cores );
if ( ( *hwinfo )->sockets > 0 )
printf( "Sockets : %d\n", ( *hwinfo )->sockets );
if ( ( *hwinfo )->nnodes > 0 )
printf( "NUMA Nodes : %d\n", ( *hwinfo )->nnodes );
printf( "CPU's per Node : %d\n", ( *hwinfo )->ncpu );
printf( "Total CPU's : %d\n", ( *hwinfo )->totalcpus );
printf( "Number Hardware Counters : %d\n",
PAPI_get_opt( PAPI_MAX_HWCTRS, NULL ) );
printf( "Max Multiplex Counters : %d\n",
PAPI_get_opt( PAPI_MAX_MPX_CTRS, NULL ) );
printf
( "--------------------------------------------------------------------------------\n" );
if ( event_flag )
printf
( "The following correspond to fields in the PAPI_event_info_t structure.\n" );
printf( "\n" );
return ( PAPI_OK );
}
/**
* 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);
}