int main( int argc, char **argv)
{
MP_INIT(argc,argv);
MP_MYID(&me);
MP_PROCS(&nproc);
if(nproc < 2) {
if(me == 0)
fprintf(stderr,
"USAGE: 2 <= processes < %d\n", nproc);
MP_BARRIER();
MP_FINALIZE();
exit(0);
}
if(me == 0){
printf("Test of ARMCI Wrappers to Basic Message Passing Operations\n");
fflush(stdout);
}
/* initialize ARMCI */
ARMCI_Init();
MP_BARRIER();
TestGlobals();
/* done */
ARMCI_Finalize();
MP_FINALIZE();
return(0);
}
int main(int argc, char **argv)
{
ARMCI_NetInit();
MP_INIT(argc,argv);
MP_MYID(&me);
MP_PROCS(&nproc);
if(nproc < 2 || nproc> MAXPROC) {
if(me == 0)
fprintf(stderr,
"USAGE: 2 <= processes < %d - got %d\n", MAXPROC, nproc);
MP_BARRIER();
MP_FINALIZE();
exit(0);
}
/* initialize ARMCI */
ARMCI_Init();
if(!me)printf("\n Performance of Basic Blocking Communication Operations\n");
MP_BARRIER();
CHECK_RESULT=1;
if(!me)printf("\n\t\t\tContiguous Data Transfer\n");
test_1D();
CHECK_RESULT=0;
/* test 1 dimension array */
if(!me)printf("\n\t\t\tContiguous Data Transfer\n");
test_1D();
/* test 2 dimension array */
if(!me)printf("\n\t\t\tStrided Data Transfer\n");
test_2D();
MP_BARRIER();
if(me == 0){
if(warn_accuracy)
printf("\nWARNING: Your timer does not have sufficient accuracy for this test (%d)\n",warn_accuracy);
printf("\n\n------------ Testing the same data transfer for correctness ----------\n");
fflush(stdout);
}
MP_BARRIER();
CHECK_RESULT=1;
if(!me)printf("\n\t\t\tContiguous Data Transfer\n");
test_1D();
if(me == 0) printf("OK\n");
MP_BARRIER();
if(!me)printf("\n\t\t\tStrided Data Transfer\n");
test_2D();
if(me == 0) printf("OK\n\n\nTests Completed.\n");
MP_BARRIER();
/* done */
ARMCI_Finalize();
MP_FINALIZE();
return(0);
}
int main(int argc, char* argv[])
{
int ndim;
MP_INIT(argc, argv);
MP_PROCS(&nproc);
MP_MYID(&me);
if(me==0){
printf("ARMCI test program for lock(%d processes)\n",nproc);
fflush(stdout);
sleep(1);
}
ARMCI_Init();
test_lock();
MP_BARRIER();
if(me==0){printf("test passed\n"); fflush(stdout);}
sleep(2);
MP_BARRIER();
ARMCI_Finalize();
MP_FINALIZE();
return(0);
}
void test_one_group(ARMCI_Group *group, int *pid_list) {
int grp_me, grp_size;
int i,j,src_proc,dst_proc;
double *ddst_put[MAXPROC];
double dsrc[ELEMS];
int elems[2] = {MAXPROC,ELEMS};
int value = -1, bytes, world_me;
MP_MYID(&world_me);
ARMCI_Group_rank(group, &grp_me);
ARMCI_Group_size(group, &grp_size);
if(grp_me==0) printf("GROUP SIZE = %d\n", grp_size);
printf("%d:group rank = %d\n", me, grp_me);
src_proc = 0; dst_proc = grp_size-1;
bytes = ELEMS*sizeof(double);
ARMCI_Malloc_group((void **)ddst_put, bytes, group);
for(i=0; i<ELEMS; i++) dsrc[i]=i*1.001*(grp_me+1);
for(i=0; i<ELEMS; i++) ddst_put[grp_me][i]=-1.0;
armci_msg_group_barrier(group);
if(grp_me==src_proc) {
/* NOTE: make sure to specify absolute ids in ARMCI calls */
ARMCI_Put(dsrc, &ddst_put[dst_proc][0], bytes,
ARMCI_Absolute_id(group,dst_proc));
}
armci_msg_group_barrier(group);
/* NOTE: make sure to specify absolute ids in ARMCI calls */
ARMCI_Fence(ARMCI_Absolute_id(group,dst_proc));
sleep(1);
/* Verify*/
if(grp_me==dst_proc) {
for(j=0; j<ELEMS; j++) {
if(ARMCI_ABS(ddst_put[grp_me][j]-j*1.001*(src_proc+1)) > 0.1) {
printf("\t%d: ddst_put[%d][%d] = %lf and expected value is %lf\n",
me, grp_me, j, ddst_put[grp_me][j], j*1.001*(src_proc+1));
ARMCI_Error("groups: armci put failed...1", 0);
}
}
printf("\n%d(%d): Test O.K. Verified\n", dst_proc, world_me);
}
armci_msg_group_barrier(group);
ARMCI_Free_group(ddst_put[grp_me], group);
}
void test_groups_noncollective() {
int *pid_lists[MAX_GROUPS];
int pids[MAXPROC];
int i, nprocs, world_me;
ARMCI_Group group;
int *my_pid_list=NULL, my_grp_size=0;
int ngrps;
MP_BARRIER();
MP_PROCS(&nprocs);
MP_MYID(&world_me);
random_permute(pids, nproc);
ngrps = nprocs/GROUP_SIZE;
for(i=0; i<nprocs/GROUP_SIZE; i++) {
pid_lists[i] = pids + (i*GROUP_SIZE);
}
for(i=0; i<nprocs; i++) {
if(pids[i] == world_me) {
int grp_id = ARMCI_MIN(i/GROUP_SIZE, ngrps-1);
my_pid_list = pid_lists[grp_id];
if(grp_id == ngrps-1)
my_grp_size = GROUP_SIZE + (nprocs%GROUP_SIZE);
else
my_grp_size = GROUP_SIZE;
}
}
qsort(my_pid_list, my_grp_size, sizeof(int), int_compare);
MP_BARRIER();
/*now create all these disjoint groups and test them in parallel*/
ARMCI_Group_create(my_grp_size, my_pid_list, &group);
test_one_group(&group, my_pid_list);
ARMCI_Group_free(&group);
ARMCI_AllFence();
MP_BARRIER();
if(world_me==0){printf("O.K.\n"); fflush(stdout);}
}
int main(int argc, char* argv[])
{
MP_INIT(argc, argv);
MP_PROCS(&nproc);
MP_MYID(&me);
/* printf("nproc = %d, me = %d\n", nproc, me);*/
if( (nproc<MINPROC || nproc>MAXPROC) && me==0)
ARMCI_Error("Test works for up to %d processors\n",MAXPROC);
if(me==0){
printf("ARMCI test program (%d processes)\n",nproc);
fflush(stdout);
sleep(1);
}
ARMCI_Init();
if(me==0){
printf("\n Testing ARMCI Groups!\n\n");
fflush(stdout);
}
test_groups();
ARMCI_AllFence();
MP_BARRIER();
if(me==0){printf("\n Collective groups: Success!!\n"); fflush(stdout);}
sleep(2);
#ifdef ARMCI_GROUP
test_groups_noncollective();
ARMCI_AllFence();
MP_BARRIER();
if(me==0){printf("\n Non-collective groups: Success!!\n"); fflush(stdout);}
sleep(2);
#endif
MP_BARRIER();
ARMCI_Finalize();
MP_FINALIZE();
return(0);
}
int main(int argc, char* argv[])
{
ARMCI_NetInit();
MP_INIT(argc, argv);
MP_PROCS(&nproc);
MP_MYID(&me);
if(nproc < 2 || nproc> MAXPROC) {
if(me == 0)
fprintf(stderr,
"USAGE: 2 <= processes < %d - got %d\n", MAXPROC, nproc);
MP_BARRIER();
MP_FINALIZE();
exit(0);
}
if(me==0){
printf("ARMCI test program (%d processes)\n",nproc);
fflush(stdout);
sleep(1);
}
ARMCI_Init();
if(me==0){
printf("\n put/get/acc requests (Time in secs)\n\n");
fflush(stdout);
}
test_perf_nb(1);
test_perf_nb(0);
ARMCI_AllFence();
MP_BARRIER();
if(me==0){printf("\nSuccess!!\n"); fflush(stdout);}
sleep(2);
MP_BARRIER();
ARMCI_Finalize();
MP_FINALIZE();
return(0);
}
int main(int argc, char* argv[])
{
MP_INIT(argc, argv);
MP_PROCS(&nproc);
MP_MYID(&me);
/* printf("nproc = %d, me = %d\n", nproc, me);*/
if(nproc>MAXPROC && me==0)
ARMCI_Error("Test works for up to %d processors\n",MAXPROC);
if(me==0){
printf("ARMCI test program (%d processes)\n",nproc);
fflush(stdout);
sleep(1);
}
ARMCI_Init();
if(me==0){
printf("\nAggregate put/get requests\n\n");
fflush(stdout);
}
test_aggregate(1); /* cold start */
test_aggregate(0); /* warm start */
ARMCI_AllFence();
MP_BARRIER();
if(me==0){printf("\nSuccess!!\n"); fflush(stdout);}
sleep(2);
MP_BARRIER();
ARMCI_Finalize();
MP_FINALIZE();
return(0);
}
main(int argc, char *argv[])
{
int i, j;
int ch;
extern char *optarg;
int edge;
int size;
int nloop=5;
double **ptr_loc;
MP_INIT(arc,argv);
MP_PROCS(&nproc);
MP_MYID(&me);
while ((ch = getopt(argc, argv, "n:b:p:h")) != -1) {
switch(ch) {
case 'n':
n = atoi(optarg);
break;
case 'b':
block_size = atoi(optarg);
break;
case 'p':
nproc = atoi(optarg);
break;
case 'h': {
printf("Usage: LU, or \n");
printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC\n");
MP_BARRIER();
MP_FINALIZE();
exit(0);
}
}
}
if(me == 0) {
printf("\n Blocked Dense LU Factorization\n");
printf(" %d by %d Matrix\n", n, n);
printf(" %d Processors\n", nproc);
printf(" %d by %d Element Blocks\n", block_size, block_size);
printf("\n");
}
num_rows = (int) sqrt((double) nproc);
for (;;) {
num_cols = nproc/num_rows;
if (num_rows*num_cols == nproc)
break;
num_rows--;
}
nblocks = n/block_size;
if (block_size * nblocks != n) {
nblocks++;
}
edge = n%block_size;
if (edge == 0) {
edge = block_size;
}
#ifdef DEBUG
if(me == 0)
for (i=0; i<nblocks; i++) {
for (j=0; j<nblocks; j++)
printf("%d ", block_owner(i, j));
printf("\n");
}
MP_BARRIER();
MP_FINALIZE();
exit(0);
#endif
for (i=0; i<nblocks; i++) {
for (j=0; j<nblocks; j++) {
if(block_owner(i,j) == me) {
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
}
else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
}
else {
size = block_size*block_size;
}
proc_bytes += size*sizeof(double);
}
}
}
ptr = (void **)malloc(nproc * sizeof(void *));
#ifdef MPI2_ONESIDED
MPI_Alloc_mem(proc_bytes, MPI_INFO_NULL, &ptr[me]);
MPI_Win_create((void*)ptr[me], proc_bytes, 1, MPI_INFO_NULL,
MPI_COMM_WORLD, &win);
for(i=0; i<nproc; i++) ptr[i] = (double *)ptr[me];
MPI_Barrier(MPI_COMM_WORLD);
#else
/* initialize ARMCI */
ARMCI_Init();
ARMCI_Malloc(ptr, proc_bytes);
#endif
a = (double **)malloc(nblocks*nblocks*sizeof(double *));
if (a == NULL) {
fprintf(stderr, "Could not malloc memory for a\n");
exit(-1);
}
ptr_loc = (double **)malloc(nproc*sizeof(double *));
for(i=0; i<nproc; i++) ptr_loc[i] = (double *)ptr[i];
for(i=0; i<nblocks; i ++) {
for(j=0; j<nblocks; j++) {
a[i+j*nblocks] = ptr_loc[block_owner(i, j)];
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
} else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
} else {
size = block_size*block_size;
}
ptr_loc[block_owner(i, j)] += size;
}
}
/* initialize the array */
init_array();
/* barrier to ensure all initialization is done */
MP_BARRIER();
/* to remove cold-start misses, all processors touch their own data */
touch_array(block_size, me);
MP_BARRIER();
if(doprint) {
if(me == 0) {
printf("Matrix before LU decomposition\n");
print_array(me);
}
MP_BARRIER();
}
lu(n, block_size, me); /* cold start */
/* Starting the timer */
MP_BARRIER();
if(me == 0) start_timer();
for(i=0; i<nloop; i++) lu(n, block_size, me);
MP_BARRIER();
/* Timer Stops here */
if(me == 0)
printf("\nRunning time = %lf milliseconds.\n\n", elapsed_time()/nloop);
printf("%d: (ngets=%d) Communication (get) time = %e milliseconds\n", me, get_cntr, comm_time*1000/nloop);
if(doprint) {
if(me == 0) {
printf("after LU\n");
print_array(me);
}
MP_BARRIER();
}
/* done */
#ifdef MPI2_ONESIDED
MPI_Win_free(&win);
MPI_Free_mem(ptr[me]);
#else
ARMCI_Free(ptr[me]);
ARMCI_Finalize();
#endif
MP_FINALIZE();
}
int main(int argc, char* argv[])
{
int i;
struct timeval start_time[14];
struct timeval stop_time[14];
/*
char * test_name[14] = {
"dim", "nbdim", "vec_small", "acc",
"vector", "vector_acc", "fetch_add",
"swap", "rput", "aggregate", "implicit",
"memlock", "acc_type", "collective"
};
int test_flags[14] = {
1, 1, 1, 1,
1, 1, 1,
1, 1, 0, 1,
1, 1, 1
};
*/
char * test_name[2] = { "acc_type", "collective" };
int test_flags[2] = { 1, 1 };
#define TEST_ACC_TYPE 0
#define TEST_COLLECTIVE 1
MP_INIT(argc, argv);
ARMCI_Init();
MP_PROCS(&nproc);
MP_MYID(&me);
if(nproc > MAXPROC && me == 0)
ARMCI_Error("Test works for up to %d processors\n",MAXPROC);
if(me == 0)
{
printf("ARMCI test program (%d processes)\n",nproc);
fflush(stdout);
sleep(1);
}
gettimeofday(&start_time[TEST_ACC_TYPE],NULL);
if(test_flags[TEST_ACC_TYPE] == 1)
{
if(me == 0)
{
printf("\nTesting Accumulate Types\n");
fflush(stdout);
}
MP_BARRIER();
if(me == 0)
{
printf("Test Accumulate ARMCI_ACC_INT\n");
fflush(stdout);
}
test_acc_type(ARMCI_ACC_INT);
ARMCI_AllFence();
MP_BARRIER();
if(me == 0)
{
printf("Test Accumulate ARMCI_ACC_LNG\n");
fflush(stdout);
}
test_acc_type(ARMCI_ACC_LNG);
ARMCI_AllFence();
MP_BARRIER();
if(me == 0)
{
printf("Test Accumulate ARMCI_ACC_FLT\n");
fflush(stdout);
}
test_acc_type(ARMCI_ACC_FLT);
ARMCI_AllFence();
MP_BARRIER();
if(me == 0)
{
printf("Test Accumulate ARMCI_ACC_DBL\n");
fflush(stdout);
}
test_acc_type(ARMCI_ACC_DBL);
ARMCI_AllFence();
MP_BARRIER();
if(me == 0)
{
printf("Test Accumulate ARMCI_ACC_CPL\n");
fflush(stdout);
}
test_acc_type(ARMCI_ACC_CPL);
ARMCI_AllFence();
MP_BARRIER();
if(me == 0)
{
printf("Test Accumulate ARMCI_ACC_DCP\n");
fflush(stdout);
}
test_acc_type(ARMCI_ACC_DCP);
ARMCI_AllFence();
MP_BARRIER();
}
gettimeofday(&stop_time[TEST_ACC_TYPE],NULL);
gettimeofday(&start_time[TEST_COLLECTIVE],NULL);
if(test_flags[TEST_COLLECTIVE] == 1)
{
if(me == 0)
{
printf("\nTesting Collective Types\n");
fflush(stdout);
}
if(me == 0)
{
printf("Test Collective ARMCI_INT\n");
fflush(stdout);
}
MP_BARRIER();
test_collective(ARMCI_INT);
MP_BARRIER();
if(me == 0)
{
printf("Test Collective ARMCI_LONG\n");
fflush(stdout);
}
MP_BARRIER();
test_collective(ARMCI_LONG);
MP_BARRIER();
if(me == 0)
{
printf("Test Collective ARMCI_FLOAT\n");
fflush(stdout);
}
MP_BARRIER();
test_collective(ARMCI_FLOAT);
MP_BARRIER();
if(me == 0)
{
printf("Test Collective ARMCI_DOUBLE\n");
fflush(stdout);
}
MP_BARRIER();
test_collective(ARMCI_DOUBLE);
MP_BARRIER();
}
gettimeofday(&stop_time[TEST_COLLECTIVE],NULL);
if(me == 0)
{
printf("Accumulate and Collective tests passed\n");
fflush(stdout);
}
if(me == 0)
{
printf("Testcase runtime\n");
printf("Name,Time(seconds)\n");
for(i = 0; i < 2; i++)
if(test_flags[i] == 1)
{
double time_spent = (stop_time[i].tv_sec - start_time[i].tv_sec) + ((double) stop_time[i].tv_usec - start_time[i].tv_usec) / 1E6;
printf("%s,%.6f\n", test_name[i], time_spent);
}
}
MP_BARRIER();
ARMCI_Finalize();
MP_FINALIZE();
return(0);
}
int main(int argc, char **argv) {
int i;
double **myptrs;
double t0, t1, tnbget=0, tnbwait=0, t2=0;
MP_INIT(argc,argv);
ARMCI_Init();
MP_PROCS(&nprocs);
MP_MYID(&me);
if (nprocs < 2)
ARMCI_Error("This program requires at least to processes", 1);
myptrs = (double **)malloc(sizeof(double *)*nprocs);
ARMCI_Malloc((void **)myptrs, LOOP*sizeof(double));
MP_BARRIER();
if(me == 0) {
for(i = 0; i < 10; i++) {
// This is a bug:
// ARMCI_Get(myptrs[me]+i,myptrs[me+1]+i,sizeof(double),me+1);
ARMCI_Get(myptrs[me+1]+i, myptrs[me]+i, sizeof(double), me+1);
}
t0 = MP_TIMER();
for(i = 0; i < LOOP; i++) {
// This is a bug:
// ARMCI_Get(myptrs[me]+i,myptrs[me+1]+i,sizeof(double),me+1);
ARMCI_Get(myptrs[me+1]+1, myptrs[me]+i, sizeof(double), me+1);
}
t1 = MP_TIMER();
printf("\nGet Latency=%lf\n", 1e6*(t1-t0)/LOOP);
fflush(stdout);
t1 = t0 = 0;
for(i = 0; i < LOOP; i++) {
armci_hdl_t nbh;
ARMCI_INIT_HANDLE(&nbh);
t0 = MP_TIMER();
//ARMCI_NbGet(myptrs[me]+i, myptrs[me+1]+i, sizeof(double), me+1, &nbh);
ARMCI_NbGet(myptrs[me+1]+i, myptrs[me]+i, sizeof(double), me+1, &nbh);
t1 = MP_TIMER();
ARMCI_Wait(&nbh);
t2 = MP_TIMER();
tnbget += (t1-t0);
tnbwait += (t2-t1);
}
printf("\nNb Get Latency=%lf Nb Wait=%lf\n",1e6*tnbget/LOOP,1e6*tnbwait/LOOP);fflush(stdout);
}
else
sleep(1);
MP_BARRIER();
ARMCI_Finalize();
MP_FINALIZE();
return 0;
}
int main(int argc, char *argv[])
{
int i, j;
int ch;
extern char *optarg;
int edge;
int size;
/* ARMCI */
void **ptr;
double **ptr_loc;
MP_INIT(argc,argv);
MP_PROCS(&nproc);
MP_MYID(&me);
while ((ch = getopt(argc, argv, "n:b:p:h")) != -1) {
switch(ch) {
case 'n': n = atoi(optarg); break;
case 'b': block_size = atoi(optarg); break;
case 'p': nproc = atoi(optarg); break;
case 'h': {
printf("Usage: LU, or \n");
printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC\n");
MP_BARRIER();
MP_FINALIZE();
exit(0);
}
}
}
if(me == 0) {
printf("\n Blocked Dense LU Factorization\n");
printf(" %d by %d Matrix\n", n, n);
printf(" %d Processors\n", nproc);
printf(" %d by %d Element Blocks\n", block_size, block_size);
printf("\n");
}
/* num_rows = (int) sqrt((double) nproc); */
/* for (;;) { */
/* num_cols = nproc/num_rows; */
/* if (num_rows*num_cols == nproc) */
/* break; */
/* num_rows--; */
/* } */
nblocks = n/block_size;
if (block_size * nblocks != n) {
nblocks++;
}
nnodes = nproc / 4;
if((nnodes * 4) != nproc) {
num_cols = nproc - nnodes * 4;
nnodes++;
num_rows = 1;
}
else {
num_cols = 2;
num_rows = 2;
}
num = (nblocks * nblocks)/nnodes;
if((num * nnodes) != (nblocks * nblocks))
num++;
#ifdef DEBUG
if(me == 0)
for (i=0;i<nblocks;i++) {
for (j=0;j<nblocks;j++)
printf("%d ", block_owner(i, j));
printf("\n");
}
MP_BARRIER();
MP_FINALIZE();
exit(0);
#endif
edge = n%block_size;
if (edge == 0) {
edge = block_size;
}
for (i=0;i<nblocks;i++) {
for (j=0;j<nblocks;j++) {
if(block_owner(i,j) == me) {
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
}
else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
}
else {
size = block_size*block_size;
}
proc_bytes += size*sizeof(double);
}
}
}
/* initialize ARMCI */
ARMCI_Init();
ptr = (void **)malloc(nproc * sizeof(void *));
ARMCI_Malloc(ptr, proc_bytes);
a = (double **)malloc(nblocks*nblocks*sizeof(double *));
if (a == NULL) {
fprintf(stderr, "Could not malloc memory for a\n");
exit(-1);
}
ptr_loc = (double **)malloc(nproc*sizeof(double *));
for(i=0; i<nproc; i++) ptr_loc[i] = (double *)ptr[i];
for(i=0; i<nblocks;i ++) {
for(j=0; j<nblocks; j++) {
a[i+j*nblocks] = ptr_loc[block_owner(i, j)];
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
} else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
} else {
size = block_size*block_size;
}
ptr_loc[block_owner(i, j)] += size;
}
}
/* initialize the array */
init_array();
/* barrier to ensure all initialization is done */
MP_BARRIER();
/* to remove cold-start misses, all processors touch their own data */
touch_array(block_size, me);
MP_BARRIER();
if(doprint) {
if(me == 0) {
printf("Matrix before LU decomposition\n");
print_array(me);
}
MP_BARRIER();
}
/* Starting the timer */
if(me == 0) start_timer();
lu(n, block_size, me);
MP_BARRIER();
/* Timer Stops here */
if(me == 0)
printf("\nRunning time = %lf milliseconds.\n\n", elapsed_time());
if(doprint) {
if(me == 0) {
printf("after LU\n");
print_array(me);
}
MP_BARRIER();
}
/* done */
ARMCI_Free(ptr[me]);
ARMCI_Finalize();
MP_FINALIZE();
return 0;
}
int main(int argc, char *argv[])
{
int ch;
extern char *optarg;
int i, j, r;
thread_t threads[MAX_TPP];
/* init MP */
MP_INIT(argc,argv);
MP_PROCS(&size);
MP_MYID(&rank);
while ((ch = getopt(argc, argv, "t:s:i:d:h")) != -1) {
switch(ch) {
case 't': /* # of threads */
tpp = atoi(optarg);
if (tpp < 1 || tpp > MAX_TPP) {
PRINTF0("\"%s\" is improper value for -t, should be a "
"number between 1 and %d(MAX_TPP)\n",
optarg, MAX_TPP);
usage();
}
break;
case 'i': /* # of iterations */
iters = atoi(optarg);
if (iters < 1) {
PRINTF0("\"%s\" is improper value for -t, should be a "
"number equal or larger than 1\n", optarg);
usage();
}
break;
case 's': /* # of elements in the array */
asize = atoi(optarg);
if (iters < 1) {
PRINTF0("\"%s\" is improper value for -s, should be a "
"number equal or larger than 1\n", optarg);
usage();
}
break;
case 'd': delay = atoi(optarg); break; /* delay before start */
case 'h': usage(); break; /* print usage info */
}
}
#ifdef NOTHREADS
tpp = 1;
PRINTF0("Warning: NOTHREADS debug symbol is set -- running w/o threads\n");
#endif
th_size = size * tpp;
PRINTF0("\nTest of multi-threaded capabilities:\n"
"%d threads per process (%d threads total),\n"
"%d array elements of size %d,\n"
"%d iteration(s)\n\n", tpp, th_size, asize, sizeof(atype_t), iters);
if (delay) {
printf("%d: %d\n", rank, getpid());
fflush(stdout);
sleep(delay);
MP_BARRIER();
}
TH_INIT(size,tpp);
for (i = 0; i < tpp; i++) th_rank[i] = rank * tpp + i;
#if defined(DEBUG) && defined(LOG2FILE)
for (i = 0; i < tpp; i++) {
fname[10] = '0' + th_rank[i] / 100;
fname[11] = '0' + th_rank[i] % 100 / 10;
fname[12] = '0' + th_rank[i] % 10;
dbg[i] = fopen(fname, "w");
}
#endif
for (i = 0; i < tpp; i++)
prndbg(i, "proc %d, thread %d(%d):\n", rank, i, th_rank[i]);
/* init ARMCI */
ARMCI_Init();
/* set global seed (to ensure same random sequence across procs) */
time_seed = (unsigned)time(NULL);
armci_msg_brdcst(&time_seed, sizeof(time_seed), 0);
srand(time_seed); rand();
prndbg(0, "seed = %u\n", time_seed);
/* random pairs */
pairs = calloc(th_size, sizeof(int));
for (i = 0; i < th_size; i++) pairs[i] = -1;
for (i = 0; i < th_size; i++) {
if (pairs[i] != -1) continue;
r = RND(0, th_size);
while (i == r || pairs[r] != -1 ) r = RND(0, th_size);
pairs[i] = r; pairs[r] = i;
}
for (i = 0, cbufl = 0; i < th_size; i++)
cbufl += sprintf(cbuf + cbufl, " %d->%d|%d->%d",
i, pairs[i], pairs[i], pairs[pairs[i]]);
prndbg(0, "random pairs:%s\n", cbuf);
/* random targets */
rnd_tgts = calloc(th_size, sizeof(int));
for (i = 0, cbufl = 0; i < th_size; i++) {
rnd_tgts[i] = RND(0, th_size);
if (rnd_tgts[i] == i) { i--; continue; }
cbufl += sprintf(cbuf + cbufl, " %d", rnd_tgts[i]);
}
prndbg(0, "random targets:%s\n", cbuf);
/* random one */
rnd_one = RND(0, th_size);
prndbg(0, "random one = %d\n", rnd_one);
assert(ptrs1 = calloc(th_size, sizeof(void *)));
assert(ptrs2 = calloc(th_size, sizeof(void *)));
#ifdef NOTHREADS
thread_main((void *)(long)0);
#else
for (i = 0; i < tpp; i++) THREAD_CREATE(threads + i, thread_main, (void *)(long)i);
for (i = 0; i < tpp; i++) THREAD_JOIN(threads[i], NULL);
#endif
MP_BARRIER();
PRINTF0("Tests Completed\n");
/* clean up */
#if defined(DEBUG) && defined(LOG2FILE)
for (i = 0; i < tpp; i++) fclose(dbg[i]);
#endif
ARMCI_Finalize();
TH_FINALIZE();
MP_FINALIZE();
return 0;
}
main(int argc, char *argv[])
{
int i, j;
int ch;
extern char *optarg;
int edge;
int size;
/* ARMCI */
void **ptr;
double **ptr_loc;
void **bufr_g, **bufc_g;
MP_INIT(arc,argv);
MP_PROCS(&nproc);
MP_MYID(&me);
while ((ch = getopt(argc, argv, "n:b:p:h")) != -1) {
switch(ch) {
case 'n': n = atoi(optarg); break;
case 'b': block_size = atoi(optarg); break;
case 'p': nproc = atoi(optarg); break;
case 'h': {
printf("Usage: LU, or \n");
printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC\n");
MP_BARRIER();
MP_FINALIZE();
exit(0);
}
}
}
if(me == 0) {
printf("\nUsing pre-PUTing\n");
printf("\n Blocked Dense LU Factorization\n");
printf(" %d by %d Matrix\n", n, n);
printf(" %d Processors\n", nproc);
printf(" %d by %d Element Blocks\n", block_size, block_size);
printf("\n");
}
num_rows = (int) sqrt((double) nproc);
for (;;) {
num_cols = nproc/num_rows;
if (num_rows*num_cols == nproc)
break;
num_rows--;
}
nblocks = n/block_size;
if (block_size * nblocks != n) {
nblocks++;
}
edge = n%block_size;
if (edge == 0) {
edge = block_size;
}
#ifdef DEBUG
if(me == 0)
for (i=0;i<nblocks;i++) {
for (j=0;j<nblocks;j++)
printf("%d ", block_owner(i, j));
printf("\n");
}
MP_BARRIER();
MP_FINALIZE();
exit(0);
#endif
for (i=0;i<nblocks;i++) {
for (j=0;j<nblocks;j++) {
if(block_owner(i,j) == me) {
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
}
else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
}
else {
size = block_size*block_size;
}
proc_bytes += size*sizeof(double);
}
}
}
/* initialize ARMCI */
ARMCI_Init();
ptr = (void **)ARMCI_Malloc_local(nproc * sizeof(void *));
ARMCI_Malloc(ptr, proc_bytes);
a = (double **)ARMCI_Malloc_local(nblocks*nblocks*sizeof(double *));
if (a == NULL) {
fprintf(stderr, "Could not malloc memory for a\n");
exit(-1);
}
ptr_loc = (double **)ARMCI_Malloc_local(nproc*sizeof(double *));
for(i=0; i<nproc; i++) ptr_loc[i] = (double *)ptr[i];
for(i=0; i<nblocks;i ++) {
for(j=0; j<nblocks; j++) {
a[i+j*nblocks] = ptr_loc[block_owner(i, j)];
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
} else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
} else {
size = block_size*block_size;
}
ptr_loc[block_owner(i, j)] += size;
}
}
/* initialize the array */
init_array();
bufr = (double **)ARMCI_Malloc_local(nproc*nblocks * sizeof(double *));
bufc = (double **)ARMCI_Malloc_local(nproc*nblocks * sizeof(double *));
if (bufr == NULL || bufc == NULL)
printf("Could not ARMCI_Malloc_local() mem\n");
/* bufr points to all k-th row blocks */
/* save all block address in row-major order */
proc_bytes = nblocks*block_size*block_size * sizeof(double);
bufr_g = (void **)ARMCI_Malloc_local(nproc * sizeof(void *));
ARMCI_Malloc(bufr_g, proc_bytes);
for (i = 0; i < nproc; i++) {
bufr[i*nblocks] = (double *) bufr_g[i];
for (j = 1; j < nblocks; j++) {
bufr[i*nblocks + j] = bufr[i*nblocks + j-1] + block_size * block_size;
}
}
/* bufc points to all k-th column blocks */
bufc_g = (void **)ARMCI_Malloc_local(nproc * sizeof(void *));
ARMCI_Malloc(bufc_g, proc_bytes);
for (i = 0; i < nproc; i++) {
bufc[i*nblocks] = (double *) bufc_g[i];
for (j = 1; j < nblocks; j++) {
bufc[i*nblocks + j] = bufc[i*nblocks + j-1] + block_size * block_size;
}
}
/* barrier to ensure all initialization is done */
MP_BARRIER();
/* to remove cold-start misses, all processors touch their own data */
touch_array(block_size, me);
MP_BARRIER();
if(doprint) {
if(me == 0) {
printf("Matrix before LU decomposition\n");
print_array(me);
}
MP_BARRIER();
}
/* Starting the timer */
if(me == 0) start_timer();
lu(n, block_size, me);
MP_BARRIER();
/* Timer Stops here */
if(me == 0)
printf("\nRunning time = %lf milliseconds.\n\n", elapsed_time());
if(doprint) {
if(me == 0) {
printf("after LU\n");
print_array(me);
}
MP_BARRIER();
}
/* done */
ARMCI_Free(ptr[me]);
ARMCI_Free(bufc_g[me]);
ARMCI_Free(bufr_g[me]);
ARMCI_Finalize();
MP_FINALIZE();
}
