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);
}
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)
{
int status, me;
int max_arrays = 10;
double max_sz = 1e8, max_disk = 1e10, max_mem = 1e6;
int stack = 120000, heap = 3200000;
int numfiles, numioprocs;
int total, nproc;
MP_INIT(argc,argv);
GA_Initialize();
me = GA_Nodeid();
nproc = GA_Nnodes();
total = pow(SIZE,NDIM)*sizeof(double);
if (!GA_Uses_ma()) {
if (GA_Nodeid() == 0) {
printf("GA is not using MA\n");
}
stack = 100000;
heap = (int)(2.2*(float)(total));
}
if (MA_init(MT_F_DBL, stack, heap) ) {
if (DRA_Init(max_arrays, max_sz, max_disk, max_mem) != 0)
GA_Error("DRA_Init failed: ",0);
if (USER_CONFIG == 0) {
numfiles = -1;
numioprocs = -1;
} else if (USER_CONFIG == 1) {
numfiles = 1;
numioprocs = GA_Cluster_nnodes();
} else if (USER_CONFIG == 2) {
numfiles = GA_Cluster_nnodes();
numioprocs = GA_Cluster_nnodes();
} else {
numfiles = 1;
numioprocs = 1;
}
if (me==0) {
printf("Disk resident arrays configured as:\n");
printf(" Number of files: %d\n",numfiles);
printf(" Number of I/O processors: %d\n",numioprocs);
}
DRA_Set_default_config(numfiles,numioprocs);
if (me == 0) printf("\n");
if (me == 0) printf("TESTING PERFORMANCE OF DISK ARRAYS\n");
if (me == 0) printf("\n");
test_io_dbl();
status = DRA_Terminate();
GA_Terminate();
} else {
printf("MA_init failed\n");
}
if(me == 0) printf("all done ...\n");
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<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)
{
int size_dst = 15;
int g_a = 0;
int I_NEG_ONE = -1;
long L_NEG_ONE = -1;
long long LL_NEG_ONE = -1;
int FIVE = 5;
int TEN = 10;
int lo;
int hi;
int *ptr;
int i;
MP_INIT(argc,argv);
GA_INIT(argc,argv);
for (i=0; i<3; ++i) {
if (0 == i) {
g_a = NGA_Create(C_INT, 1, &size_dst, "dst", NULL);
GA_Fill(g_a, &I_NEG_ONE);
} else if (1 == i) {
g_a = NGA_Create(C_LONG, 1, &size_dst, "dst", NULL);
GA_Fill(g_a, &L_NEG_ONE);
} else if (2 == i) {
g_a = NGA_Create(C_LONGLONG, 1, &size_dst, "dst", NULL);
GA_Fill(g_a, &LL_NEG_ONE);
}
GA_Sync();
GA_Print(g_a);
NGA_Print_patch(g_a, &FIVE, &TEN, 0);
NGA_Print_patch(g_a, &FIVE, &TEN, 1);
NGA_Distribution(g_a, GA_Nodeid(), &lo, &hi);
NGA_Access(g_a, &lo, &hi, &ptr, NULL);
printf("[%d] (%d)=%d\n", GA_Nodeid(), lo, *ptr);
NGA_Release(g_a, &lo, &hi);
}
GA_Terminate();
MP_FINALIZE();
exit(EXIT_SUCCESS);
}
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);
}
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;
}
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();
}
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;
}
int main( int argc, char **argv ) {
int g_a, g_b, i, j, size, size_me;
int icnt, idx, jdx, ld;
int n=N, type=MT_C_INT, one;
int *values, *ptr;
int **indices;
int dims[2]={N,N};
int lo[2], hi[2];
int heap=3000000, stack=2000000;
int me, nproc;
int datatype, elements;
double *prealloc_mem;
MP_INIT(argc,argv);
#if 1
GA_INIT(argc,argv); /* initialize GA */
me=GA_Nodeid();
nproc=GA_Nnodes();
if(me==0) {
if(GA_Uses_fapi())GA_Error("Program runs with C array API only",1);
printf("\nUsing %ld processes\n",(long)nproc);
fflush(stdout);
}
heap /= nproc;
stack /= nproc;
if(! MA_init(MT_F_DBL, stack, heap))
GA_Error("MA_init failed",stack+heap); /* initialize memory allocator*/
/* Create a regular matrix. */
if(me==0)printf("\nCreating matrix A of size %d x %d\n",N,N);
g_a = NGA_Create(type, 2, dims, "A", NULL);
if(!g_a) GA_Error("create failed: A",n);
/* Fill matrix using scatter routines */
size = N*N;
if (size%nproc == 0) {
size_me = size/nproc;
} else {
i = size - size%nproc;
size_me = i/nproc;
if (me < size%nproc) size_me++;
}
/* Check that sizes are all okay */
i = size_me;
GA_Igop(&i,1,"+");
if (i != size) {
GA_Error("Sizes don't add up correctly: ",i);
} else if (me==0) {
printf("\nSizes add up correctly\n");
}
/* Allocate index and value arrays */
indices = (int**)malloc(size_me*sizeof(int*));
values = (int*)malloc(size_me*sizeof(int));
icnt = me;
for (i=0; i<size_me; i++) {
values[i] = icnt;
idx = icnt%N;
jdx = (icnt-idx)/N;
if (idx >= N || idx < 0) {
printf("p[%d] Bogus index i: %d\n",me,idx);
}
if (jdx >= N || jdx < 0) {
printf("p[%d] Bogus index j: %d\n",me,jdx);
}
indices[i] = (int*)malloc(2*sizeof(int));
(indices[i])[0] = idx;
(indices[i])[1] = jdx;
icnt += nproc;
}
/* Scatter values into g_a */
NGA_Scatter(g_a, values, indices, size_me);
GA_Sync();
/* Check to see if contents of g_a are correct */
NGA_Distribution( g_a, me, lo, hi );
NGA_Access(g_a, lo, hi, &ptr, &ld);
for (i=lo[0]; i<hi[0]; i++) {
idx = i-lo[0];
for (j=lo[1]; j<hi[1]; j++) {
jdx = j-lo[1];
if (ptr[idx*ld+jdx] != j*N+i) {
printf("p[%d] (Scatter) expected: %d actual: %d\n",me,j*N+i,ptr[idx*ld+jdx]);
}
}
}
if (me==0) printf("\nCompleted test of NGA_Scatter\n");
for (i=0; i<size_me; i++) {
values[i] = 0;
}
GA_Sync();
NGA_Gather(g_a, values, indices, size_me);
icnt = me;
for (i=0; i<size_me; i++) {
if (icnt != values[i]) {
printf("p[%d] (Gather) expected: %d actual: %d\n",me,icnt,values[i]);
}
icnt += nproc;
}
if (me==0) printf("\nCompleted test of NGA_Gather\n");
GA_Sync();
/* Scatter-accumulate values back into GA*/
one = 1;
NGA_Scatter_acc(g_a, values, indices, size_me, &one);
GA_Sync();
/* Check to see if contents of g_a are correct */
for (i=lo[0]; i<hi[0]; i++) {
idx = i-lo[0];
for (j=lo[1]; j<hi[1]; j++) {
jdx = j-lo[1];
if (ptr[idx*ld+jdx] != 2*(j*N+i)) {
printf("p[%d] (Scatter_acc) expected: %d actual: %d\n",me,2*(j*N+i),ptr[idx*ld+jdx]);
}
}
}
if (me==0) printf("\nCompleted test of NGA_Scatter_acc\n");
NGA_Release(g_a, lo, hi);
/* Test fixed buffer size */
NGA_Alloc_gatscat_buf(size_me);
/* Scatter-accumulate values back into GA*/
GA_Sync();
NGA_Scatter_acc(g_a, values, indices, size_me, &one);
GA_Sync();
/* Check to see if contents of g_a are correct */
for (i=lo[0]; i<hi[0]; i++) {
idx = i-lo[0];
for (j=lo[1]; j<hi[1]; j++) {
jdx = j-lo[1];
if (ptr[idx*ld+jdx] != 3*(j*N+i)) {
printf("p[%d] (Scatter_acc) expected: %d actual: %d\n",me,3*(j*N+i),ptr[idx*ld+jdx]);
}
}
}
if (me==0) printf("\nCompleted test of NGA_Scatter_acc using fixed buffers\n");
NGA_Release(g_a, lo, hi);
NGA_Free_gatscat_buf();
GA_Destroy(g_a);
if(me==0)printf("\nSuccess\n");
GA_Terminate();
#endif
MP_FINALIZE();
return 0;
}
int main(int argc, char **argv)
{
/* initialize GA */
#if defined(USE_ELEMENTAL)
// initialize Elemental (which will initialize MPI)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &me );
ElMPICommSize( MPI_COMM_WORLD, &nproc );
// instantiate el::global array
ElGlobalArraysConstruct_d( &eldga );
// initialize global arrays
ElGlobalArraysInitialize_d( eldga );
#else
MP_INIT(argc,argv);
GA_Initialize_args(&argc, &argv);
me = GA_Nodeid();
nproc = GA_Nnodes();
#endif
if (nproc < 2) {
if (me == 0) {
fprintf(stderr, "USAGE: 2 <= processes - got %d\n", nproc);
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysTerminate_d( eldga );
// call el::global arrays destructor
ElGlobalArraysDestruct_d( eldga );
ElFinalize();
#else
GA_Terminate();
MP_FINALIZE();
#endif
exit(0);
}
if (!me) {
printf("\n Performance of Basic Blocking Communication Operations\n");
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysSync_d( eldga );
#else
GA_Sync();
#endif
/* 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();
#if 0
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------------ Now we test the same data transfer for correctness ----------\n");
fflush(stdout);
}
#endif
#if defined(USE_ELEMENTAL)
ElGlobalArraysTerminate_d( eldga );
// call el::global arrays destructor
ElGlobalArraysDestruct_d( eldga );
ElFinalize();
#else
GA_Terminate();
MP_FINALIZE();
#endif
return(0);
}
int
main(int argc, char **argv) {
Integer heap=9000000, stack=9000000;
int me, nproc;
DoublePrecision time;
MP_INIT(argc,argv);
GA_INIT(argc,argv); /* initialize GA */
nproc = GA_Nnodes();
me = GA_Nodeid();
if(me==0) printf("Using %d processes\n\n",nproc);
if (me==0) printf ("Matrix size is %d X %d\n",N,N);
#ifdef USE_REGULAR
if (me == 0) printf("\nUsing regular data distribution\n\n");
#endif
#ifdef USE_SIMPLE_CYCLIC
if (me == 0) printf("\nUsing simple block-cyclic data distribution\n\n");
#endif
#ifdef USE_SCALAPACK
if (me == 0) printf("\nUsing ScaLAPACK data distribution\n\n");
#endif
#ifdef USE_TILED
if (me == 0) printf("\nUsing tiled data distribution\n\n");
#endif
if(!MA_init((Integer)MT_F_DBL, stack/nproc, heap/nproc))
GA_Error("MA_init failed bytes= %d",stack+heap);
#ifdef PERMUTE
{
int i, *list = (int*)malloc(nproc*sizeof(int));
if(!list)GA_Error("malloc failed",nproc);
for(i=0; i<nproc;i++)list[i]=nproc-1-i;
GA_Register_proclist(list, nproc);
free(list);
}
#endif
if(GA_Uses_fapi())GA_Error("Program runs with C API only",1);
time = MP_TIMER();
do_work();
/* printf("%d: Total Time = %lf\n", me, MP_TIMER()-time);
printf("%d: GEMM Total Time = %lf\n", me, gTime);
*/
if(me==0)printf("\nSuccess\n\n");
GA_Terminate();
MP_FINALIZE();
return 0;
}
int main(int argc, char **argv) {
int me;
int g_a;
int status;
int i,j;
int dims[] = {n,n};
int proc_group[PROC_LIST_SIZE],proclist[PROC_LIST_SIZE],inode;
int sbuf[1],rbuf[1];
MPI_Comm comm;
MP_INIT(argc,argv);
GA_Initialize();
me = GA_Nodeid();
status = MA_init(MT_DBL, 100000, 100000);
if (!status) GA_Error("ma_init failed",-1);
status = MA_set_auto_verify(1);
status = MA_set_hard_fail(1);
status = MA_set_error_print(1);
inode = GA_Cluster_nodeid();
if (me == 0) {
printf("there are %d nodes, node 0 has %d procs\n",
GA_Cluster_nnodes(), GA_Cluster_nprocs(0));
fflush(stdout);
}
GA_Sync();
for (i=0; i<GA_Cluster_nnodes(); ++i) {
for (j=0; j<GA_Cluster_nprocs(i); ++j) {
proclist[j]=GA_Cluster_procid(i,j);
}
proc_group[i]=GA_Pgroup_create(proclist,GA_Cluster_nprocs(i));
}
GA_Sync();
for (i=0; i<GA_Cluster_nnodes(); ++i) {
if (i == inode) {
printf("%d joining group %d\n", me, proc_group[inode]);
GA_Pgroup_set_default(proc_group[inode]);
g_a = NGA_Create(C_DBL, 2, dims, "a", NULL);
if (!g_a) GA_Error("NGA_Create failed",-1);
printf("%d Created array of group %d as proc no. %d\n",
me, proc_group[inode], GA_Nodeid());
GA_Print_distribution(g_a);
comm = GA_MPI_Comm_pgroup_default();
if (comm != MPI_COMM_NULL) {
sbuf[0] = GA_Nodeid();
status = MPI_Allreduce(sbuf, rbuf, 1, MPI_INT, MPI_MAX, comm);
printf("%d max nodeid is %d\n", me, rbuf[0]);
if ((rbuf[0]+1) != GA_Cluster_nprocs(i)) {
GA_Error("MPI_Allreduce failed",1);
}
}
else {
printf("MPI_Comm was null!\n");
}
GA_Pgroup_set_default(GA_Pgroup_get_world());
}
GA_Sync();
}
GA_Terminate();
MP_FINALIZE();
return 0;
}
// note: Sayan: brings down memory requirement to about 268 MB
int main(int argc, char **argv)
{
int me, nproc, g_a = -1, i, j;
#if defined(USE_ELEMENTAL)
int ndim=2, dims[2]= {N1,N2};
#else
int ndim=2, type=MT_F_DBL, dims[2]= {N1,N2};
#endif
double *buf;
int lo[2], hi[2], ld[1];
double alpha = 1.0;
#if defined(USE_ELEMENTAL)
// initialize Elemental (which will initialize MPI)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &me );
ElMPICommSize( MPI_COMM_WORLD, &nproc );
ElGlobalArrays_d eldga;
// instantiate el::global array
ElGlobalArraysConstruct_d( &eldga );
// initialize global arrays
ElGlobalArraysInitialize_d( eldga );
printf ("INITIALIZED elemental global array...\n");
#else
MP_INIT(argc,argv);
GA_Initialize_ltd(-1);
me=GA_Nodeid();
nproc=GA_Nnodes();
#endif
if(me==0) printf("Using %ld processes\n",(long)nproc);
if(me==0) printf("memory = %ld bytes\n",((long)N1)*((long)N2)*8);
#if defined(USE_ELEMENTAL)
// create and allocate a global array
printf ("ndim = %d\n", ndim);
printf ("dim[0] = %d and dim[1] = %d\n", dims[0], dims[1]);
ElGlobalArraysCreate_d( eldga, ndim, dims, "A", &g_a);
printf ("CREATED elemental global array...\n");
// print distribution
ElGlobalArraysPrint_d( eldga, g_a );
#else
g_a = NGA_Create(type, ndim, dims, "A", NULL);
GA_Zero(g_a); /* zero the matrix */
GA_Print_distribution(g_a);
#endif
if(me == 0) {
// buf = (double*)(malloc(N1*1024*sizeof(double)));
buf = (double*)(malloc(N1*128*sizeof(double)));
// for(j = 0; j < N1*1024; ++j) buf[j] = 1.0;
// for(i = 0; i < N2/1024; ++i) {
for(j = 0; j < N1*128; ++j) buf[j] = 1.0;
for(i = 0; i < N2/128; ++i) {
lo[0] = 0;
hi[0] = lo[0] + N1 -1;
/*
lo[1] = i*1024;
hi[1] = lo[1] + 1024 -1;
ld[0] = 1024;
*/
lo[1] = i*128;
hi[1] = lo[1] + 128 -1;
ld[0] = 128;
printf("NGA_Acc.%d: %d:%d %d:%d\n",i,lo[0],hi[0],lo[1],hi[1]);
#if defined(USE_ELEMENTAL)
ElGlobalArraysAccumulate_d( eldga, g_a, lo, hi, buf, ld, &alpha );
// there is an explicit flush in NGA_Acc/Put, so when it returns, the buffer
// can be reused and data has reached the destination
#else
NGA_Init_fence();
NGA_Acc(g_a, lo, hi, buf, ld, &alpha);
NGA_Fence();
#endif
}
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysSync_d( eldga );
ElGlobalArraysDestroy_d( eldga, g_a );
ElGlobalArraysTerminate_d( eldga );
// call el::global arrays destructor
ElGlobalArraysDestruct_d( eldga );
ElFinalize();
#else
GA_Sync();
GA_Destroy(g_a);
GA_Terminate();
MP_FINALIZE();
#endif
return 0;
}
/*
* test ga_dgemm
* Note: - change nummax for large arrays
* - turn off "dgemm_verify" for large arrays due to memory
* limitations, as dgemm_verify=1 for large arrays produces
* segfault, dumps core,or any crap.
*/
int main(int argc, char **argv)
{
int num_m;
int num_n;
int num_k;
int i;
int ii;
double *h0;
int g_c;
int g_b;
int g_a;
double a;
double t1;
double mf;
double avg_t[ntrans];
double avg_mf[ntrans];
int itime;
int ntimes;
int nums_m[/*howmany*/] = {512,1024};
int nums_n[/*howmany*/] = {512,1024};
int nums_k[/*howmany*/] = {512,1024};
char transa[/*ntrans*/] = "ntnt";
char transb[/*ntrans*/] = "nntt";
char ta;
char tb;
double *tmpa;
double *tmpb;
double *tmpc;
int ndim;
int dims[2];
#ifdef BLOCK_CYCLIC
int block_size[2];
#endif
#if defined(USE_ELEMENTAL)
// initialize Elemental (which will initialize MPI)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &me );
ElMPICommSize( MPI_COMM_WORLD, &nproc );
// instantiate el::global array
ElGlobalArraysConstruct_d( &eldga );
// initialize global arrays
ElGlobalArraysInitialize_d( eldga );
#else
MP_INIT(argc,argv);
if (!MA_init(MT_DBL,1,20000000)) {
GA_Error("failed: ma_init(MT_DBL,1,20000000)",10);
}
GA_INIT(argc,argv);
me = GA_Nodeid();
#endif
h0 = (double*)malloc(sizeof(double) * nummax*nummax);
tmpa = (double*)malloc(sizeof(double) * nummax*nummax);
tmpb = (double*)malloc(sizeof(double) * nummax*nummax);
tmpc = (double*)malloc(sizeof(double) * nummax*nummax);
ii = 0;
for (i=0; i<nummax*nummax; i++) {
ii = ii + 1;
if (ii > nummax) {
ii = 0;
}
h0[i] = ii;
}
/* Compute times assuming 500 mflops and 5 second target time */
/* ntimes = max(3.0d0,5.0d0/(4.0d-9*num**3)); */
ntimes = 5;
for (ii=0; ii<howmany; ii++) {
num_m = nums_m[ii];
num_n = nums_n[ii];
num_k = nums_k[ii];
a = 0.5/(num_m*num_n);
if (num_m > nummax || num_n > nummax || num_k > nummax) {
GA_Error("Insufficient memory: check nummax", 1);
}
#ifndef BLOCK_CYCLIC
ndim = 2;
/*
dims[0] = num_m;
dims[1] = num_n;
*/
dims[1] = num_m;
dims[0] = num_n;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_c", NULL, &g_c );
#else
if (!((g_c = NGA_Create(MT_DBL,ndim,dims,"g_c",NULL)))) {
GA_Error("failed: create g_c",20);
}
#endif
/*
dims[0] = num_k;
dims[1] = num_n;
*/
dims[1] = num_k;
dims[0] = num_n;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_b", NULL, &g_b );
#else
if (!((g_b = NGA_Create(MT_DBL,ndim,dims,"g_b",NULL)))) {
GA_Error("failed: create g_b",30);
}
#endif
/*
dims[0] = num_m;
dims[1] = num_k;
*/
dims[1] = num_m;
dims[0] = num_k;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_a", NULL, &g_a );
#else
if (!((g_a = NGA_Create(MT_DBL,ndim,dims,"g_a",NULL)))) {
GA_Error("failed: create g_a",40);
}
#endif
#else
ndim = 2;
block_size[0] = 128;
block_size[1] = 128;
dims[0] = num_m;
dims[1] = num_n;
g_c = GA_Create_handle();
GA_Set_data(g_c,ndim,dims,MT_DBL);
GA_Set_array_name(g_c,"g_c");
GA_Set_block_cyclic(g_c,block_size);
if (!GA_Allocate(g_c)) {
GA_Error("failed: create g_c",40);
}
dims[0] = num_k;
dims[1] = num_n;
g_b = GA_Create_handle();
GA_Set_data(g_b,ndim,dims,MT_DBL);
GA_Set_array_name(g_b,"g_b");
GA_Set_block_cyclic(g_b,block_size);
if (!ga_allocate(g_b)) {
GA_Error("failed: create g_b",40);
}
dims[0] = num_m;
dims[1] = num_k;
g_a = GA_Create_handle();
GA_Set_data(g_a,ndim,dims,MT_DBL);
GA_Set_array_name(g_a,"g_a");
GA_Set_block_cyclic(g_a,block_size);
if (!ga_allocate(g_a)) {
GA_Error('failed: create g_a',40);
}
#endif
/* Initialize matrices A and B */
if (me == 0) {
load_ga(g_a, h0, num_m, num_k);
load_ga(g_b, h0, num_k, num_n);
}
#if defined(USE_ELEMENTAL)
double zero = 0.0;
ElGlobalArraysFill_d( eldga, g_c, &zero );
ElGlobalArraysSync_d( eldga );
#else
GA_Zero(g_c);
GA_Sync();
#endif
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("\nMatrix Multiplication on C = A[%ld,%ld]xB[%ld,%ld]\n",
(long)num_m, (long)num_k, (long)num_k, (long)num_n);
fflush(stdout);
}
for (i=0; i<ntrans; i++) {
avg_t[i] = 0.0;
avg_mf[i] = 0.0;
}
for (itime=0; itime<ntimes; itime++) {
for (i=0; i<ntrans; i++) {
#if defined(USE_ELEMENTAL)
ElGlobalArraysSync_d( eldga );
#else
GA_Sync();
#endif
ta = transa[i];
tb = transb[i];
t1 = MP_TIMER();
#if defined(USE_ELEMENTAL)
ElGlobalArraysDgemm_d( eldga, ta, tb, num_m, num_n, num_k, 1.0, g_a, g_b, 0.0, g_c );
#else
GA_Dgemm(ta,tb,num_m,num_n,num_k,1.0, g_a, g_b, 0.0, g_c);
#endif
t1 = MP_TIMER() - t1;
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
#if defined(USE_ELEMENTAL)
mf = 2e0*num_m*num_n*num_k/t1*1e-6/nproc;
#else
mf = 2e0*num_m*num_n*num_k/t1*1e-6/GA_Nnodes();
#endif
avg_t[i] = avg_t[i]+t1;
avg_mf[i] = avg_mf[i] + mf;
printf("%15s%2d: %12.4f seconds %12.1f mflops/proc %c %c\n",
"Run#", itime, t1, mf, ta, tb);
fflush(stdout);
if (dgemm_verify && itime == 0) {
/* recall the C API swaps the matrix order */
/* we swap it here for the Fortran-based verify */
verify_ga_dgemm(tb, ta, num_n, num_m, num_k, 1.0,
g_b, g_a, 0.0, g_c, tmpb, tmpa, tmpc);
}
}
}
}
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("\n");
for (i=0; i<ntrans; i++) {
printf("%17s: %12.4f seconds %12.1f mflops/proc %c %c\n",
"Average", avg_t[i]/ntimes, avg_mf[i]/ntimes,
transa[i], transb[i]);
}
if(dgemm_verify) {
printf("All GA_Dgemms are verified...O.K.\n");
}
fflush(stdout);
}
/*
GA_Print(g_a);
GA_Print(g_b);
GA_Print(g_c);
*/
#if defined(USE_ELEMENTAL)
ElGlobalArraysDestroy_d( eldga, g_a );
ElGlobalArraysDestroy_d( eldga, g_b );
ElGlobalArraysDestroy_d( eldga, g_c );
#else
GA_Destroy(g_c);
GA_Destroy(g_b);
GA_Destroy(g_a);
#endif
}
/* ???
format(a15, i2, ': ', e12.4, ' seconds ',f12.1,
. ' mflops/proc ', 3a2)
*/
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("All tests successful\n");
}
free(h0);
free(tmpa);
free(tmpb);
free(tmpc);
#if defined(USE_ELEMENTAL)
// call el::global arrays destructor
ElGlobalArraysTerminate_d( eldga );
ElGlobalArraysDestruct_d( eldga );
ElFinalize();
#else
GA_Terminate();
MP_FINALIZE();
#endif
return 0;
}
/*
* Verify for correctness. Process 0 computes BLAS dgemm
* locally. For larger arrays, disbale this test as memory
* might not be sufficient
*/
void verify_ga_dgemm(char xt1, char xt2, int num_m, int num_n, int num_k,
double alpha, int g_a, int g_b, double beta, int g_c,
double *tmpa, double *tmpb, double *tmpc)
{
int i,j,type,ndim,dims[2],lo[2],hi[2];
double abs_value;
for (i=0; i<num_n; i++) {
for (j=0; j<num_m; j++) {
tmpc[j+i*num_m] = -1.0;
tmpa[j+i*num_m] = -2.0;
}
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_a, lo, hi, tmpa, &dims[1] );
#else
NGA_Get(g_a, lo, hi, tmpa, &dims[1]);
#endif
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_b, lo, hi, tmpb, &dims[1] );
#else
NGA_Get(g_b, lo, hi, tmpb, &dims[1]);
#endif
/* compute dgemm sequentially */
#if defined(USE_ELEMENTAL)
cblas_dgemm ( CblasRowMajor, ( xt1 == 'n'? CblasNoTrans: CblasTrans ),
( xt2 == 'n'? CblasNoTrans: CblasTrans ),
num_m /* M */, num_n /* N */, num_k /* K */,
alpha, tmpa, num_m, /* lda */
tmpb, num_k, /* ldb */ beta,
tmpc, num_m /* ldc */);
#else
xb_dgemm(&xt1, &xt2, &num_m, &num_n, &num_k,
&alpha, tmpa, &num_m,
tmpb, &num_k, &beta,
tmpc, &num_m);
#endif
/* after computing c locally, verify it with the values in g_c */
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_c, lo, hi, tmpa, &dims[1] );
#else
NGA_Get(g_c, lo, hi, tmpa, &dims[1]);
#endif
for (i=0; i<num_n; i++) {
for (j=0; j<num_m; j++) {
abs_value = fabs(tmpc[j+i*num_m]-tmpa[j+i*num_m]);
if(abs_value > 1.0 || abs_value < -1.0) {
printf("Values are = %f %f\n",
tmpc[j+i*num_m], tmpa[j+i*num_m]);
printf("Values are = %f %f\n",
fabs(tmpc[j+i*num_m]-tmpa[j*i*num_m]), abs_value);
fflush(stdout);
GA_Error("verify ga_dgemm failed", 1);
}
}
}
}
/**
* called by process '0' (or your master process )
*/
void load_ga(int handle, double *f, int dim1, int dim2)
{
int lo[2], hi[2];
if (dim1 < 0 || dim2 < 0) {
return;
}
lo[0] = 0;
lo[1] = 0;
hi[0] = dim1-1;
hi[1] = dim2-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysPut_d( eldga, handle, lo, hi, f, &dim1 );
#else
NGA_Put(handle, lo, hi, f, &dim1);
#endif
}
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 me;
int nproc;
int status;
int g_a;
int dims[NDIM];
int chunk[NDIM];
int pg_world;
size_t num = 10;
double *p1 = NULL;
double *p2 = NULL;
size_t i;
int num_mutex;
int lo[1];
int hi[1];
int ld[1]={1};
MPI_Comm comm;
MP_INIT(argc,argv);
GA_INIT(argc,argv);
me = GA_Nodeid();
nproc = GA_Nnodes();
comm = GA_MPI_Comm_pgroup_default();
printf("%d: Hello world!\n",me);
if (me==0) printf("%d: GA_Initialize\n",me);
/*if (me==0) printf("%d: ARMCI_Init\n",me);*/
/*ARMCI_Init();*/
/*if (me==0) printf("%d: MA_Init\n",me);*/
/*MA_init(MT_DBL, 8*1024*1024, 2*1024*1024);*/
if (me==0) printf("%d: GA_Create_handle\n",me);
g_a = GA_Create_handle();
if (me==0) printf("%d: GA_Set_array_name\n",me);
GA_Set_array_name(g_a,"test array A");
dims[0] = 30;
if (me==0) printf("%d: GA_Set_data\n",me);
GA_Set_data(g_a,NDIM,dims,MT_DBL);
chunk[0] = -1;
if (me==0) printf("%d: GA_Set_chunk\n",me);
GA_Set_chunk(g_a,chunk);
if (me==0) printf("%d: GA_Pgroup_get_world\n",me);
pg_world = GA_Pgroup_get_world();
if (me==0) printf("%d: GA_Set_pgroup\n",me);
GA_Set_pgroup(g_a,pg_world);
if (me==0) printf("%d: GA_Allocate\n",me);
status = GA_Allocate(g_a);
if(0 == status) MPI_Abort(comm,100);
if (me==0) printf("%d: GA_Zero\n",me);
GA_Zero(g_a);
if (me==0) printf("%d: GA_Sync\n",me);
GA_Sync();
num = 10;
p1 = malloc(num*sizeof(double));
/*double* p1 = ARMCI_Malloc_local(num*sizeof(double));*/
if (p1==NULL) MPI_Abort(comm,1000);
p2 = malloc(num*sizeof(double));
/*double* p2 = ARMCI_Malloc_local(num*sizeof(double));*/
if (p2==NULL) MPI_Abort(comm,2000);
for ( i=0 ; i<num ; i++ ) p1[i] = 7.0;
for ( i=0 ; i<num ; i++ ) p2[i] = 3.0;
num_mutex = 17;
status = GA_Create_mutexes(num_mutex);
if (me==0) printf("%d: GA_Create_mutexes = %d\n",me,status);
/***************************************************************/
if (me==0) {
printf("%d: before GA_Lock\n",me);
GA_Lock(0);
lo[0] = 0;
hi[0] = num-1;
GA_Init_fence();
NGA_Put(g_a,lo,hi,p1,ld);
GA_Fence();
GA_Unlock(0);
printf("%d: after GA_Unlock\n",me);
}
GA_Print(g_a);
if (me==1) {
printf("%d: before GA_Lock\n",me);
GA_Lock(0);
lo[0] = 0;
hi[0] = num-1;
GA_Init_fence();
NGA_Get(g_a,lo,hi,p2,ld);
GA_Fence();
GA_Unlock(0);
printf("%d: after GA_Unlock\n",me);
for ( i=0 ; i<num ; i++ ) printf("p2[%2lu] = %20.10f\n",
(long unsigned)i,p2[i]);
}
/***************************************************************/
status = GA_Destroy_mutexes();
if (me==0) printf("%d: GA_Destroy_mutexes = %d\n",me,status);
/*ARMCI_Free(p2);*/
/*ARMCI_Free(p1);*/
free(p2);
free(p1);
if (me==0) printf("%d: GA_Destroy\n",me);
GA_Destroy(g_a);
/*if (me==0) printf("%d: ARMCI_Finalize\n",me);*/
/*ARMCI_Finalize();*/
if (me==0) printf("%d: GA_Terminate\n",me);
GA_Terminate();
if (me==0) printf("%d: MPI_Finalize\n",me);
MPI_Finalize();
return(0);
}