int gptlstop (char *name, int nc1)
{
char cname[MAX_CHARS+1];
int numchars;
numchars = MIN (nc1, MAX_CHARS);
strncpy (cname, name, numchars);
cname[numchars] = '\0';
return GPTLstop (cname);
}
double sub (int iter)
{
unsigned long usec;
unsigned long looplen = iam*iter*100000;
unsigned long i;
double sum;
int ret;
ret = GPTLstart ("sub");
/* Sleep msec is mpi rank + thread number */
usec = 1000 * (iam * iter);
ret = GPTLstart ("sleep");
usleep (usec);
ret = GPTLstop ("sleep");
ret = GPTLstart ("work");
sum = 0.;
ret = GPTLstart ("add");
for (i = 0; i < looplen; ++i) {
sum += i;
}
ret = GPTLstop ("add");
ret = GPTLstart ("madd");
for (i = 0; i < looplen; ++i) {
sum += i*1.1;
}
ret = GPTLstop ("madd");
ret = GPTLstart ("div");
for (i = 0; i < looplen; ++i) {
sum /= 1.1;
}
ret = GPTLstop ("div");
ret = GPTLstop ("work");
ret = GPTLstop ("sub");
return sum;
}
int main (int argc, char **argv)
{
char pname[MPI_MAX_PROCESSOR_NAME];
int iter;
int counter;
int c;
int tnum = 0;
int resultlen;
int ret;
double value;
extern char *optarg;
while ((c = getopt (argc, argv, "p:")) != -1) {
switch (c) {
case 'p':
if ((ret = GPTLevent_name_to_code (optarg, &counter)) != 0) {
printf ("Failure from GPTLevent_name_to_code\n");
return 1;
}
if (GPTLsetoption (counter, 1) < 0) {
printf ("Failure from GPTLsetoption (%s,1)\n", optarg);
return 1;
}
break;
default:
printf ("unknown option %c\n", c);
printf ("Usage: %s [-p option_name]\n", argv[0]);
return 2;
}
}
ret = GPTLsetoption (GPTLabort_on_error, 1);
ret = GPTLsetoption (GPTLoverhead, 1);
ret = GPTLsetoption (GPTLnarrowprint, 1);
if (MPI_Init (&argc, &argv) != MPI_SUCCESS) {
printf ("Failure from MPI_Init\n");
return 1;
}
ret = GPTLinitialize ();
ret = GPTLstart ("total");
ret = MPI_Comm_rank (MPI_COMM_WORLD, &iam);
ret = MPI_Comm_size (MPI_COMM_WORLD, &nproc);
ret = MPI_Get_processor_name (pname, &resultlen);
printf ("Rank %d is running on processor %s\n", iam, pname);
#ifdef THREADED_OMP
nthreads = omp_get_max_threads ();
#pragma omp parallel for private (iter, ret, tnum)
#endif
for (iter = 1; iter <= nthreads; iter++) {
#ifdef THREADED_OMP
tnum = omp_get_thread_num ();
#endif
printf ("Thread %d of rank %d on processor %s\n", tnum, iam, pname);
value = sub (iter);
}
ret = GPTLstop ("total");
ret = GPTLpr (iam);
if (iam == 0) {
printf ("summary: testing GPTLpr_summary...\n");
printf ("Number of threads was %d\n", nthreads);
printf ("Number of tasks was %d\n", nproc);
}
// NOTE: if ENABLE_PMPI is set, 2nd pr call below will show some extra send/recv calls
// due to MPI calls from within GPTLpr_summary_file
if (GPTLpr_summary (MPI_COMM_WORLD) != 0)
return 1;
if (GPTLpr_summary_file (MPI_COMM_WORLD, "timing.summary.duplicate") != 0)
return 1;
ret = MPI_Finalize ();
if (GPTLfinalize () != 0)
return 1;
return 0;
}
int main (int argc, char **argv)
{
int ret;
int i, code;
long long pc[1]; /* papi counters */
double sum;
printf ("testpapi: Testing PAPI interface...\n");
printf ("%s: testing getting event code for PAPI_TOT_CYC...\n", argv[0]);
if ((ret = GPTLevent_name_to_code ("PAPI_TOT_CYC", &code)) != 0) {
printf ("Failure\n");
return 2;
}
printf ("Success\n");
printf ("%s: testing GPTLsetoption(PAPI_TOT_CYC,1)...\n", argv[0]);
if (GPTLsetoption (code, 1) != 0) {
printf ("Failure\n");
return 3;
}
printf ("Success\n");
printf ("%s: testing GPTLinitialize\n", argv[0]);
if ((ret = GPTLinitialize ()) != 0) {
printf ("Failure\n");
return 3;
}
printf ("Success\n");
printf ("%s: testing GPTLstart\n", argv[0]);
if ((ret = GPTLstart ("sum")) != 0) {
printf ("Unexpected failure from GPTLstart(\"sum\")\n");
return 3;
}
printf ("Success\n");
sum = 0.;
for (i = 0; i < 1000000; ++i)
sum += (double) i;
printf ("%s: testing GPTLstop\n", argv[0]);
if ((ret = GPTLstop ("sum")) != 0) {
printf ("Unexpected failure from GPTLstop(\"sum\")\n");
return 3;
}
printf ("Success\n");
printf ("%s: testing GPTLquerycounters...\n", argv[0]);
if (GPTLquerycounters ("sum", 0, pc) != 0) {
printf ("Failure\n");
return 4;
}
printf ("sum=%g\n",sum);
printf ("%s: testing reasonableness of PAPI counters...\n", argv[0]);
if (pc[0] < 1 || pc[0] > 1.e8) {
printf ("Suspicious PAPI_TOT_CYC value=%lld for 1e6 additions\n", pc[0]);
return 5;
} else {
printf ("Success\n");
}
printf ("%s: all tests successful\n", argv[0]);
return 0;
}
/*
* If define GPUALL=1, then vector x and y are in GPU
* Otherwise, x and y are in CPU
* */
void aprod(int version, int procRank, int numProc, int mode, int m /*row*/,
int n /*column*/, MY_FLOAT_TYPE x[]/*v len=n*/,
MY_FLOAT_TYPE y[] /*u=perPorcVectorB len=m, assign NULL if PLSQRV3*/, void *UsrWork,
int isFinalIteration, int iteration)
{
#ifdef BARRIER
MPI_Barrier(MPI_COMM_WORLD);
#endif
#ifdef GPTL
GPTLstart("aprod");
#endif
int i;
///////////////////////////////////////PLSQRv3///////////////////////////////////////////////////////////////////////////////////////////////
PPDataV3_TRI * p = (PPDataV3_TRI *) UsrWork;
if (mode == 1) {//y <- y+Ax
#ifdef BARRIER
MPI_Barrier(MPI_COMM_WORLD);
#endif
#ifdef GPTL
GPTLstart("aprod1_spmv");
#endif
aprod_mode_1_tridiagonal(procRank, numProc, iteration, isFinalIteration, p);
#ifdef BARRIER
MPI_Barrier(MPI_COMM_WORLD);
#endif
#ifdef GPTL
GPTLstop("aprod1_spmv");
#endif
}//end of PLSQR3 mode=1
/*PLSQR3 Aprod mode = 2*/
else if (mode == 2){//x<-x+AT*y
#ifdef BARRIER
MPI_Barrier(MPI_COMM_WORLD);
#endif
#ifdef GPTL
GPTLstart("aprod2_spmvT");
#endif
aprod_mode_2_tridiagonal(procRank, numProc, iteration, isFinalIteration, p);
#ifdef BARRIER
MPI_Barrier(MPI_COMM_WORLD);
#endif
#ifdef GPTL
GPTLstop("aprod2_spmvT");
#endif
}//end of mode =2
#ifdef GPTL
GPTLstop("aprod");
#endif
}
