void * fight(void * input)
{
int tid = get_thread_id();
int count = 100000;
if (debug)
printf("%d: before L2_Barrier \n", tid);
uint64_t t0 = GetTimeBase();
for (int i=0; i<count; i++)
L2_Barrier(&barrier, num_threads);
uint64_t t1 = GetTimeBase();
if (debug) {
printf("%d: after L2_Barrier \n", tid);
fflush(stdout);
}
uint64_t dt = t1-t0;
printf("%2d: %d calls to %s took %llu cycles per call \n",
tid, count, "L2_Barrier", dt/count);
fflush(stdout);
pthread_exit(NULL);
return NULL;
}
void * slowfight(void * input)
{
int tid = get_thread_id();
if (debug)
printf("%d: before L2_Barrier 1 \n", tid);
L2_Barrier(&barrier, num_threads);
if (debug) {
printf("%d: after L2_Barrier 1 \n", tid);
fflush(stdout);
}
int count = 1000000;
uint64_t rval;
uint64_t t0 = GetTimeBase();
for (int i=0; i<count; i++)
rval = Fetch_and_Add(&(slowcounter.atom), 1);
uint64_t t1 = GetTimeBase();
if (debug)
printf("%d: before L2_Barrier 2 \n", tid);
L2_Barrier(&barrier, num_threads);
if (debug) {
printf("%d: after L2_Barrier 2 \n", tid);
fflush(stdout);
}
uint64_t dt = t1-t0;
printf("%2d: %d calls to %s took %llu cycles per call \n",
tid, count, "Fetch_and_Add", dt/count);
fflush(stdout);
pthread_exit(NULL);
return NULL;
}
void * fight(void * input)
{
int tid = get_thread_id();
if (debug)
printf("%d: before L2_Barrier 1 \n", tid);
L2_Barrier(&barrier, num_threads);
if (debug) {
printf("%d: after L2_Barrier 1 \n", tid);
fflush(stdout);
}
int count = 1000000;
uint64_t t0 = GetTimeBase();
for (int i=0; i<count; i++) {
volatile uint64_t rval = L2_AtomicLoadIncrement(&(counter.atom));
}
uint64_t t1 = GetTimeBase();
if (debug)
printf("%d: before L2_Barrier 2 \n", tid);
L2_Barrier(&barrier, num_threads);
if (debug) {
printf("%d: after L2_Barrier 2 \n", tid);
fflush(stdout);
}
uint64_t dt = t1-t0;
printf("%2d: %d calls to %s took %llu cycles per call \n",
tid, count, "L2_AtomicLoadIncrement", dt/count);
fflush(stdout);
pthread_exit(NULL);
return NULL;
}
void EMON_Init_t(int numthreads)
{
unsigned int tid, pid, cid,node_id;
tid = PhysicalThreadID(); // between 0 and 3
pid = PhysicalThreadIndex(); // between 0 and 67
cid = pid/4;
node_id = EMON_rank_on_card();
// printf("node id = %d\n", node_id);
// EMON set up is done only by one thread of one node
if (pid == 0)// && (node_id == 0) )
{
int rc = EMON_SetupPowerMeasurement();
if (rc) {
printf("ERROR : EMON_SetupPowerMeasurement failed with rc=%d\n", rc);
test_exit(rc);
}
}
L2_Barrier(&id_barrier, numthreads);
}
/*========================================================================*/
void HPM_Print_t(int numthreads)
{
int i, j, k, nblocks;
// uint64_t counts, counts_0, counts_1, counts_2, counts_3;
long long counts, counts_0, counts_1, counts_2, counts_3;
// uint64_t l1_misses, l1p_misses, node_l1p_misses, node_l2_misses, loads, node_loads, node_l1_misses;
// double cycles, ipc, stall_cycles, node_stall_cycles, node_cycles;
// double node_fxu_instructions, node_fpu_instructions, node_l1_hits, node_l1p_hits, ddr_hit_fraction;
// double fxu_fraction, fpu_fraction, fxu_instructions, fpu_instructions;
// double cores_per_process, max_fraction, percent_max_issue_rate;
// double l1_hits, l1_hit_fraction, l1p_hits, l1p_hit_fraction, node_l2_hits, l2_hit_fraction;
// double ld_bytes_per_cycle, st_bytes_per_cycle;
// uint64_t node_punit_counts[MAX_COUNTERS];
// long long node_timebase_sum;
// int Ax, Bx, Cx, Dx, Ex;
// Personality_t personality;
// char filename[132];
// FILE * fp;
unsigned int tid, pid, cid;
int node_id;
tid = PhysicalThreadID(); // between 0 and 3
pid = PhysicalThreadIndex(); // between 0 and 67
cid = pid/4;
node_id = EMON_rank_on_card();
L2_Barrier(&id_barrier, numthreads);
if ((pid != 0) || (node_id != 0))
return;
set_labels();
set_aggregation_mask(); // sets a mask to aggregate by sum (0) or by max (1)
if (code_block >= MAX_CODE_BLOCKS) nblocks = MAX_CODE_BLOCKS;
else nblocks = code_block;
// fp = stderr;
// print counts for each thread and the aggregate for every core
printf( "\n");
printf( "======================================================================\n");
printf( "Hardware counter report for BGQ - thread and core-specific values.\n");
printf( "======================================================================\n");
for (k=0; k<MAX_CORES; k++) {
if (coremask[k]) {
printf ("core %d\n", k);
for (j=0; j<nblocks; j++) {
if (block_starts[j] == block_stops[j]) {
printf( "----------------------------------------------------------------\n");
printf( "%s, call count = %d, cycles = %lld :\n",
code_block_label[j], block_starts[j], timebase_sum[j]);
printf( " -- Processor counters (thread specific) --------------\n");
if (hpm_threads == 1) {
for (i=0; i<num_events; i++) {
counts = counter_sum[k][j][i];
printf( "%-d %14lld %s\n", hpm_group, counts, label[counter_index[i]]);
}
}
else if (hpm_threads == 2) {
printf( " thread0 counts thread2 counts net counts label\n");
for (i=0; i<num_events; i++) {
counts_0 = counter_sum[k][j][i];
counts_2 = counter_sum[k][j][i+num_events];
if (mask[counter_index[i]]) {
counts = (counts_0 > counts_2) ? counts_0 : counts_2;
}
else counts = counts_0 + counts_2;
printf( "%-d %14lld %14lld %14lld %s\n",
hpm_group, counts_0, counts_2, counts, label[counter_index[i]]);
}
}
else if (hpm_threads == 3) {
printf( " thread0 counts thread1 counts thread2 counts net counts label\n");
for (i=0; i<num_events; i++) {
counts_0 = counter_sum[k][j][i];
counts_2 = counter_sum[k][j][i+num_events];
counts_1 = counter_sum[k][j][i+2*num_events];
if (mask[counter_index[i]]) {
counts = (counts_0 > counts_2) ? counts_0 : counts_2;
counts = (counts > counts_1) ? counts : counts_1;
}
else counts = counts_0 + counts_1 + counts_2;
printf( "%-d %14lld %14lld %14lld %14lld %s\n",
hpm_group, counts_0, counts_1, counts_2, counts, label[counter_index[i]]);
}
}
else if (hpm_threads == 4) {
printf( " thread0 counts thread1 counts thread2 counts thread3 counts net counts label\n");
for (i=0; i<num_events; i++) {
counts_0 = counter_sum[k][j][i];
counts_2 = counter_sum[k][j][i+num_events];
counts_1 = counter_sum[k][j][i+2*num_events];
counts_3 = counter_sum[k][j][i+3*num_events];
if (mask[counter_index[i]]) {
counts = (counts_0 > counts_2) ? counts_0 : counts_2;
counts = (counts > counts_1) ? counts : counts_1;
counts = (counts > counts_3) ? counts : counts_3;
}
else counts = counts_0 + counts_1 + counts_2 + counts_3;
printf( "%-d %14lld %14lld %14lld %14lld %14lld %s\n",
hpm_group, counts_0, counts_1, counts_2, counts_3, counts, label[counter_index[i]]);
}
}
printf( " -- L2 counters (shared for the node) -----------------\n");
printf( "%-d %14lld L2 Hits\n", 100, L2_sum[j][0]);
printf( "%-d %14lld L2 Misses\n", 100, L2_sum[j][1]);
printf( "%-d %14lld L2 lines prefetched\n", 100, L2_sum[j][2]);
printf( "%-d %14lld L2 lines loaded from memory\n", 100, L2_sum[j][3]);
printf( "%-d %14lld L2 full lines stored to mem\n", 100, L2_sum[j][4]);
printf( "%-d %14lld L2 partial lines stored to mem\n", 100, L2_sum[j][5]);
printf( "\n");
}
else {
printf( "mismatch in starts/stops for code block '%s'\n", code_block_label[j]);
printf( " starts = %d\n", block_starts[j]);
printf( " stops = %d\n", block_stops[j]);
}
}
printf( "\n");
}
}
return;
}
/*=================================================================*/
void HPM_Init_t(int numthreads)
{
int i, j, k, core;
// int threads_per_core;
// int * eventSet;
char * ptr;
unsigned int tid, pid, cid;
unsigned int lock_status;
int rc;
// Upci_Mode_t Mode;
tid = PhysicalThreadID(); // between 0 and 3
pid = PhysicalThreadIndex(); // between 0 and 67
cid = pid/4;
if (pid == 0)
{
// set the initial cumulative counter values to zero
for (k=0; k<MAX_CORES; k++)
for (j=0; j<MAX_CODE_BLOCKS; j++)
for (i=0; i<MAX_COUNTERS; i++)
counter_sum[k][j][i] = 0LL;
for (j=0; j<MAX_CODE_BLOCKS; j++) timebase_sum[j] = 0LL;
for (j=0; j<MAX_CODE_BLOCKS; j++)
for (i=0; i<6; i++)
L2_sum[j][i] = 0LL;
// keep track of code block starts and stops
for (j=0; j<MAX_CODE_BLOCKS; j++) {
block_starts[j] = 0;
block_stops[j] = 0;
}
// set mask used for thread and core aggregation
for (i=0; i<MAX_EVENTS; i++) mask[i] = 0;
// check env variables
// fixme
ptr = fwext_getenv("HPM_GROUP");
if (ptr == NULL) {
hpm_group = 0;
}
else hpm_group = hpm_atoi(ptr);
// printf("hpm_group = %d\n", hpm_group);
// hpm_group = 82;
if (hpm_group < -1) hpm_group = 0;
if (hpm_group > 99) hpm_group = 0;
// fixme
// ptr = fwext_getenv("HPM_SCOPE"); if (pid !=0) return;
// if (ptr != NULL) {
// if (strncasecmp(ptr,"process", 7) == 0) process_scope = 1;
// if (strncasecmp(ptr,"node", 4) == 0) node_scope = 1;
// }
// fixme
// ptr = fwext_getenv("HPM_METRICS");
// if (ptr != NULL) {
// if (strncasecmp(ptr,"yes", 3) == 0) derived_metrics = 1;
// }
for (i=0; i<MAX_CORES; i++) coremask[i] = 1;
// find the number of cores used by this process
// fixme
// numcores = 0;
// for (i=0; i<MAX_CORES; i++) numcores += coremask[i];
numcores = 17;
// determine the number of threads per core
// numthreads = BgGetNumThreads();
// numthreads = 68;
// threads_per_core = numthreads / numcores;
// hpm_threads = threads_per_core;
// fixme
hpm_threads = 4;
// optionally reset the number of threads per core that will be counted
// fixme
// ptr = fwext_getenv("HPM_THREADS");
// if (ptr != NULL) {
// hpm_threads = fwext_atoi(ptr);
// if (hpm_threads < 1) hpm_threads = 1;
// if (hpm_threads > 4) hpm_threads = 4;
// }
// set num_events and num_counters based on hpm_group and hpm_threads
switch (hpm_group) {
case -1:
num_events = 6;
eventSet = exptSet;
break;
case 0:
num_events = 6;
eventSet = mySet;
break;
case 1:
num_events = 12;
if (hpm_threads > 2) hpm_threads = 2;
eventSet = ldSet;
break;
case 2:
num_events = 24;
if (hpm_threads > 1) hpm_threads = 1;
eventSet = fpuSet;
break;
case 3:
num_events = 12;
if (hpm_threads > 2) hpm_threads = 2;
eventSet = fpSet0;
break;
case 30:
num_events = 6;
eventSet = fpSet00;
break;
case 31:
num_events = 6;
eventSet = fpSet01;
break;
case 4:
num_events = 12;
if (hpm_threads > 2) hpm_threads = 2;
eventSet = fpSet1;
break;
case 40:
num_events = 6;
eventSet = fpSet10;
break;
case 41:
num_events = 6;
eventSet = fpSet11;
break;
case 5:
num_events = 24;
if (hpm_threads > 1) hpm_threads = 1;
eventSet = fxuSet;
break;
case 6:
num_events = 12;
if (hpm_threads > 2) hpm_threads = 2;
eventSet = fxSet0;
break;
case 60:
num_events = 6;
eventSet = fxSet00;
break;
case 61:
num_events = 6;
eventSet = fxSet01;
break;
case 7:
num_events = 12;
if (hpm_threads > 2) hpm_threads = 2;
eventSet = fxSet1;
break;
case 70:
num_events = 6;
eventSet = fxSet10;
break;
case 71:
num_events = 6;
eventSet = fxSet11;
break;
case 81:
num_events = 6;
eventSet = l1pset0;
break;
case 82:
num_events = 6;
eventSet = l1pset1;
break;
case 83:
num_events = 6;
eventSet = l1pset2;
break;
default:
break;
}
num_counters = num_events * hpm_threads;
ppc_msync();
Upci_Mode_Init(&Mode[0], UPC_DISTRIB_MODE, UPC_CM_INDEP, 0);
initialized = 1;
ppc_msync();
}
while ((initialized == 0) && (tid == 0))
{
;
}
if (tid == 0) {
lock_status = 0;
while (lock_status == 0)
{
lock_status = hpm_lock_acquire();
}
core = cid;
// initialize hardware counters
// Upci_Mode_Init(&Mode[core], UPC_DISTRIB_MODE, UPC_CM_INDEP, core);
Upci_Punit_Init(&Punit[core], &Mode[core], core);
// UPC_L1p_SetMode(core, L1P_CFG_UPC_SWITCH);
// use one thread per core to enable 24 different punit counters
// add events to count, save hwthread in one of the reserved event handle slots
k = 0;
for (i=0; i<num_events; i++) { // hwthread 0
rc = Upci_Punit_AddEvent(&Punit[core], eventSet[i], 0, &eventHandle[core][k]);
if (rc != 0) printf("failed to add event %d\n", eventSet[i]);
if (pid == 0)
counter_index[k] = eventSet[i];
eventHandle[core][k].rsv[0] = 0;
k++;
}
if (hpm_threads > 1) {
for (i=0; i<num_events; i++) { // hwthread 2
rc = Upci_Punit_AddEvent(&Punit[core], eventSet[i], 2, &eventHandle[core][k]);
if (rc != 0) printf("failed to add event %d\n", eventSet[i]);
if (pid == 0)
counter_index[k] = eventSet[i];
eventHandle[core][k].rsv[0] = 2;
k++;
}
}
if (hpm_threads > 2) {
for (i=0; i<num_events; i++) { // hwthread 1
rc = Upci_Punit_AddEvent(&Punit[core], eventSet[i], 1, &eventHandle[core][k]);
if (rc != 0) printf("failed to add event %d\n", eventSet[i]);
if (pid == 0)
counter_index[k] = eventSet[i];
eventHandle[core][k].rsv[0] = 1;
k++;
}
}
if (hpm_threads > 3) {
for (i=0; i<num_events; i++) { // hwthread 3
rc = Upci_Punit_AddEvent(&Punit[core], eventSet[i], 3, &eventHandle[core][k]);
if (rc != 0) printf("failed to add event %d\n", eventSet[i]);
if (pid == 0)
counter_index[k] = eventSet[i];
eventHandle[core][k].rsv[0] = 3;
k++;
}
}
rc = Upci_Punit_Apply(&Punit[core]);
if (rc != 0) printf("Upci_Punit_Apply failed\n");
Upci_Punit_Start(&Punit[core], (UPCI_CTL_RESET | UPCI_CTL_DELAY));
// printf("Initialised upc by core = %d\n", cid);
// Upci_Punit_Dump(2, &Punit[core]);
lock_val = 0;
ppc_msync();
}
if (pid == 0)
{
UPC_L2_EnableUPC(1, 1);
UPC_L2_Start();
}
// PMPI_Barrier(local_comm);
L2_Barrier(&id_barrier, numthreads);
return;
}
