int main(int argc, char **argv)
{
int k,i;
double **myptrs[10];
double t0,t1,tget=0,tnbget=0,tput=0,tnbput=0,tnbwait=0,t2=0;
#if PORTALS
ARMCI_NetInit();
#endif
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD,&me);
MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
ARMCI_Init();
ARMCI_Init();
for(k=0;k<10;k++){
myptrs[k] = (double **)malloc(sizeof(double *)*nprocs);
ARMCI_Malloc((void **)myptrs[k],400000*LOOP*sizeof(double));
for(i=0;i<LOOP;i++)myptrs[k][me][i]=me+0.414;
MPI_Barrier(MPI_COMM_WORLD);
for(i=0;i<LOOP;i++){
ARMCI_Get(myptrs[k][(me+1)%nprocs]+i,myptrs[k][me]+i,sizeof(double),(me+1)%nprocs);
/*if(myptrs[k][me][i]!=0.414+(me+1)%nprocs)ARMCI_Error("errr",myptrs[k][me][i]);*/
}
t0=t1=tget=tnbget=tput=tnbput=tnbwait=t2=0;
t0 = MPI_Wtime();
for(i=0;i<LOOP;i++){
ARMCI_Get(myptrs[k][(me+1)%nprocs]+i,myptrs[k][me]+i,sizeof(double),(me+1)%nprocs);
}
t1 = MPI_Wtime();
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 = MPI_Wtime();
ARMCI_NbGet(myptrs[k][(me+1)%nprocs]+i,myptrs[k][me]+i,sizeof(double),(me+1)%nprocs,&nbh);
t1 = MPI_Wtime();
ARMCI_Wait(&nbh);
t2 = MPI_Wtime();
tnbget+=(t1-t0);
tnbwait+=(t2-t1);
}
printf("\nNb Get Latency=%lf Nb Wait=%lf\n",1e6*tnbget/LOOP,1e6*tnbwait/LOOP);fflush(stdout);
MPI_Barrier(MPI_COMM_WORLD);
}
for(k=0;k<10;k++)ARMCI_Free(myptrs[k][me]);
MPI_Barrier(MPI_COMM_WORLD);
ARMCI_Finalize();
ARMCI_Finalize();
MPI_Finalize();
}
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)
{
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 ndim;
armci_msg_init(&argc, &argv);
nproc = armci_msg_nproc();
me = armci_msg_me();
ARMCI_Init();
armci_msg_barrier();
if(me==0){
printf("\nTesting armci_notify\n");
fflush(stdout);
sleep(1);
}
armci_msg_barrier();
for(ndim=1; ndim<=MAXDIMS; ndim++) test_notify(ndim);
armci_msg_barrier();
ARMCI_Finalize();
armci_msg_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) {
/* int heap=300000, stack=300000; */
int me, nprocs;
/* Step1: Initialize Message Passing library */
armci_msg_init(&argc, &argv);
/* Step2: Initialize ARMCI */
ARMCI_Init();
/* Step3: Initialize Memory Allocator (MA) */
/*bjp
if(! MA_init(C_DBL, stack, heap) ) ARMCI_Error("MA_init failed",stack+heap);
*/
me = armci_msg_me();
nprocs = armci_msg_nproc();
if(me==0) {
printf("\nUsing %d processes\n\n", nprocs); fflush(stdout);
}
TRANSPOSE1D();
if(me==0)printf("\nTerminating ..\n");
ARMCI_Finalize();
armci_msg_finalize();
return(0);
}
int main(int argc, char ** argv) {
int rank, nproc, val, i;
void **base_ptrs;
MPI_Init(&argc, &argv);
ARMCI_Init();
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
if (rank == 0) printf("Starting ARMCI mutex read-modify-write test with %d processes\n", nproc);
base_ptrs = malloc(nproc*sizeof(void*));
ARMCI_Create_mutexes(rank == 0 ? 1 : 0);
ARMCI_Malloc(base_ptrs, (rank == 0) ? sizeof(int) : 0); // Proc 0 has a shared int
if (rank == 0) {
val = 0;
ARMCI_Put(&val, base_ptrs[0], sizeof(int), 0);
}
ARMCI_Barrier();
for (i = 0; i < NITER; i++) {
ARMCI_Lock(0, 0);
ARMCI_Get(base_ptrs[0], &val, sizeof(int), 0);
val += ADDIN;
ARMCI_Put(&val, base_ptrs[0], sizeof(int), 0);
ARMCI_Unlock(0, 0);
}
printf(" + %3d done\n", rank);
fflush(NULL);
ARMCI_Barrier();
if (rank == 0) {
ARMCI_Get(base_ptrs[0], &val, sizeof(int), 0);
if (val == ADDIN*nproc*NITER)
printf("Test complete: PASS.\n");
else
printf("Test complete: FAIL. Got %d, expected %d.\n", val, ADDIN*nproc*NITER);
}
ARMCI_Free(base_ptrs[rank]);
ARMCI_Destroy_mutexes();
free(base_ptrs);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
int main(int argc, char ** argv) {
MPI_Init(&argc, &argv);
ARMCI_Init();
ARMCI_Get(NULL, NULL, 1, 0);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
int main(int argc, char** argv)
{
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
MPI_Comm_rank(MPI_COMM_WORLD, &me);
if(me==0)printf("Testing IPCs (%d MPI processes)\n\n",nproc);
ARMCI_Init();
test();
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
int main(int argc, char ** argv) {
int rank, nproc, i;
int *buf;
MPI_Init(&argc, &argv);
ARMCI_Init();
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
if (rank == 0) printf("Starting ARMCI GOP test with %d processes\n", nproc);
buf = malloc(DATA_SZ*sizeof(int));
if (rank == 0) printf(" - Testing ABSMIN\n");
for (i = 0; i < DATA_SZ; i++)
buf[i] = (rank+1) * ((i % 2) ? -1 : 1);
armci_msg_igop(buf, DATA_SZ, "absmin");
for (i = 0; i < DATA_SZ; i++)
if (buf[i] != 1) {
printf("Err: buf[%d] = %d expected 1\n", i, buf[i]);
ARMCI_Error("Fail", 1);
}
if (rank == 0) printf(" - Testing ABSMAX\n");
for (i = 0; i < DATA_SZ; i++)
buf[i] = (rank+1) * ((i % 2) ? -1 : 1);
armci_msg_igop(buf, DATA_SZ, "absmax");
for (i = 0; i < DATA_SZ; i++)
if (buf[i] != nproc) {
printf("Err: buf[%d] = %d expected %d\n", i, buf[i], nproc);
ARMCI_Error("Fail", 1);
}
free(buf);
if (rank == 0) printf("Pass.\n");
ARMCI_Finalize();
MPI_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[]) {
void *baseAddress[MAX_PROCESSORS];
char *local;
int thisImage;
int iter = 100, size;
double startTime, endTime;
int i;
// initialize
ARMCI_Init();
ARMCI_Myid(&thisImage);
// allocate data (collective operation)
ARMCI_Malloc(baseAddress, MAX_BUF_SIZE*sizeof(char));
local = (char *)ARMCI_Malloc_local(MAX_BUF_SIZE*sizeof(char));
ARMCI_Barrier();
ARMCI_Migrate();
if (thisImage == 0) {
for(size = 1; size <= MAX_BUF_SIZE; size = size<<1){
startTime = CkWallTimer();
for(i = 0; i < iter; i++){
ARMCI_Put(local, baseAddress[1], size, 1);
}
ARMCI_Fence(1);
endTime = CkWallTimer();
printf("%d: %f us\n", size, (endTime-startTime)*1000);
}
ARMCI_Barrier();
} else if (thisImage == 1) {
ARMCI_Barrier();
}
ARMCI_Free(baseAddress[thisImage]);
ARMCI_Free_local(local);
// finalize
ARMCI_Finalize();
return 0;
}
int main(int argc, char ** argv) {
int rank, nproc, test_iter;
void ***base_ptrs;
MPI_Init(&argc, &argv);
ARMCI_Init();
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
if (rank == 0) printf("Starting ARMCI memory allocation test with %d processes\n", nproc);
base_ptrs = malloc(sizeof(void**)*NUM_ITERATIONS);
// Perform a pile of allocations
for (test_iter = 0; test_iter < NUM_ITERATIONS; test_iter++) {
if (rank == 0) printf(" + allocation %d\n", test_iter);
base_ptrs[test_iter] = malloc(sizeof(void*)*nproc);
ARMCI_Malloc((void**)base_ptrs[test_iter], (test_iter % 4 == 0) ? 0 : DATA_SZ);
}
ARMCI_Barrier();
// Free all allocations
for (test_iter = 0; test_iter < NUM_ITERATIONS; test_iter++) {
if (rank == 0) printf(" + free %d\n", test_iter);
ARMCI_Free(((void**)base_ptrs[test_iter])[rank]);
free(base_ptrs[test_iter]);
}
free(base_ptrs);
if (rank == 0) printf("Test complete: PASS.\n");
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
int main(int argc, char* argv[])
{
armci_msg_init(&argc, &argv);
nproc = armci_msg_nproc();
me = armci_msg_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("\n Performing Sparse Matrix-Vector Multiplication ...\n\n");
fflush(stdout);
}
test_sparse();
ARMCI_AllFence();
armci_msg_barrier();
if(me==0) {
printf("\nSuccess!!\n");
fflush(stdout);
}
sleep(2);
armci_msg_barrier();
ARMCI_Finalize();
armci_msg_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, l;
int ch;
extern char *optarg;
int edge;
int size;
int lu_arg[MAX_THREADS][3];
/* ARMCI */
void **ptr;
double **ptr_loc;
THREAD_LOCK_INIT(mutex);
armci_msg_init(&argc,&argv);
nproc = armci_msg_nproc();
me = armci_msg_me();
while ((ch = getopt(argc, argv, "n:b:p:t:d:h")) != -1) {
switch(ch) {
case 'n': n = atoi(optarg); break;
case 'b': block_size = atoi(optarg); break;
case 'p': nproc = atoi(optarg); break;
case 't': th_per_p = atoi(optarg); break;
case 'd': d = atoi(optarg); break;
case 'h': {
printf("Usage: LU, or \n");
printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC -tTH_PER_P\n");
armci_msg_barrier();
armci_msg_finalize();
exit(0);
}
}
}
if(th_per_p>MAX_THREADS) {
th_per_p=MAX_THREADS;
if(me==0)printf("Warning: cannot run more than %d threads, adjust MAX_THREADS",MAX_THREADS);
}
if (d) {
fprintf(stderr, "%d: %d\n", me, getpid());
sleep(d);
}
nthreads = th_per_p * nproc;
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 thread(s) per processor, %d threads total\n", th_per_p, nthreads);
printf(" %d by %d Element Blocks\n", block_size, block_size);
printf("\n");
}
num_rows = (int) sqrt((double) nthreads);
for (;;) {
num_cols = nthreads/num_rows;
if (num_rows*num_cols == nthreads)
break;
num_rows--;
}
nblocks = n/block_size;
if (block_size * nblocks != n) {
nblocks++;
}
num = (nblocks * nblocks)/nthreads;
if((num * nthreads) != (nblocks * nblocks))
num++;
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");
}
armci_msg_barrier();
/* armci_msg_finalize(); */
/* exit(0); */
#endif
for (l = 0; l < th_per_p; l++) {
me_th[l] = me * th_per_p + l;
for (i=0;i<nblocks;i++) {
for (j=0;j<nblocks;j++) {
if(block_owner(i,j) == me_th[l]) {
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;
}
thread_doubles[l] += size;
}
}
}
proc_bytes += thread_doubles[l] * 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(nthreads*sizeof(double *));
for (i = 0; i < nproc; i++) {
ptr_loc[i * th_per_p] = (double *)ptr[i];
for (j = 1; j < th_per_p; j++)
ptr_loc[i * th_per_p + j] = ptr_loc[i * th_per_p + j - 1] + thread_doubles[j - 1];
}
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;
}
}
#if 0
for(i=0; i<nblocks*nblocks;i ++) printf("%d: a[%d]=%p\n", me, i, a[i]);
fflush(stdout);
#endif
/* initialize the array */
init_array();
/* barrier to ensure all initialization is done */
armci_msg_barrier();
/* to remove cold-start misses, all processors touch their own data */
/* for (l = 0; l < th_per_p; l++) touch_array(block_size, me_th[l]); */
armci_msg_barrier();
if(doprint) {
if(me == 0) {
printf("Matrix before LU decomposition\n");
print_array(me);
}
armci_msg_barrier();
}
#if 1
for (i = 0; i < nblocks; i++)
for (j = 0; j < nblocks; j++)
print_block_dbg(a[i + j * nblocks], "proc %d, a[%d, %d]:\n", me, i, j);
#endif
TH_INIT(nproc,th_per_p);
/* Starting the timer */
if(me == 0) start_timer();
for (l = 0; l < th_per_p; l++) {
lu_arg[l][0] = n;
lu_arg[l][1] = block_size;
lu_arg[l][2] = l;
THREAD_CREATE(threads + l, lu, lu_arg[l]);
}
for (l = 0; l < th_per_p; l++) THREAD_JOIN(threads[l], NULL);
armci_msg_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);
}
armci_msg_barrier();
}
/* done */
ARMCI_Free(ptr[me]);
ARMCI_Finalize();
armci_msg_finalize();
THREAD_LOCK_DESTROY(mutex);
}
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, j, rank, nranks, peer, bufsize, errors, total_errors;
double **buf_bvec, **src_bvec, *src_buf;
int count[2], src_stride, trg_stride, stride_level;
double scaling, time;
MPI_Init(&argc, &argv);
ARMCI_Init();
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
buf_bvec = (double **) malloc(sizeof(double *) * nranks);
src_bvec = (double **) malloc(sizeof(double *) * nranks);
bufsize = XDIM * YDIM * sizeof(double);
ARMCI_Malloc((void **) buf_bvec, bufsize);
ARMCI_Malloc((void **) src_bvec, bufsize);
src_buf = src_bvec[rank];
if (rank == 0)
printf("ARMCI Strided DLA Accumulate Test:\n");
ARMCI_Access_begin(buf_bvec[rank]);
ARMCI_Access_begin(src_buf);
for (i = 0; i < XDIM*YDIM; i++) {
*(buf_bvec[rank] + i) = 1.0 + rank;
*(src_buf + i) = 1.0 + rank;
}
ARMCI_Access_end(src_buf);
ARMCI_Access_end(buf_bvec[rank]);
scaling = 2.0;
src_stride = XDIM * sizeof(double);
trg_stride = XDIM * sizeof(double);
stride_level = 1;
count[1] = YDIM;
count[0] = XDIM * sizeof(double);
ARMCI_Barrier();
time = MPI_Wtime();
peer = (rank+1) % nranks;
for (i = 0; i < ITERATIONS; i++) {
ARMCI_AccS(ARMCI_ACC_DBL,
(void *) &scaling,
src_buf,
&src_stride,
(void *) buf_bvec[peer],
&trg_stride,
count,
stride_level,
peer);
}
ARMCI_Barrier();
time = MPI_Wtime() - time;
if (rank == 0) printf("Time: %f sec\n", time);
ARMCI_Access_begin(buf_bvec[rank]);
for (i = errors = 0; i < XDIM; i++) {
for (j = 0; j < YDIM; j++) {
const double actual = *(buf_bvec[rank] + i + j*XDIM);
const double expected = (1.0 + rank) + scaling * (1.0 + ((rank+nranks-1)%nranks)) * (ITERATIONS);
if (actual - expected > 1e-10) {
printf("%d: Data validation failed at [%d, %d] expected=%f actual=%f\n",
rank, j, i, expected, actual);
errors++;
fflush(stdout);
}
}
}
ARMCI_Access_end(buf_bvec[rank]);
MPI_Allreduce(&errors, &total_errors, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
ARMCI_Free((void *) buf_bvec[rank]);
ARMCI_Free((void *) src_bvec[rank]);
free(buf_bvec);
free(src_bvec);
ARMCI_Finalize();
MPI_Finalize();
if (total_errors == 0) {
if (rank == 0) printf("Success.\n");
return 0;
} else {
if (rank == 0) printf("Fail.\n");
return 1;
}
}
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);
}
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, rank, nranks, peer, bufsize, errors;
double **buffer, *src_buf;
int count[2], src_stride, trg_stride, stride_level;
MPI_Init(&argc, &argv);
ARMCI_Init();
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
buffer = (double **) malloc(sizeof(double *) * nranks);
bufsize = XDIM * YDIM * sizeof(double);
ARMCI_Malloc((void **) buffer, bufsize);
src_buf = ARMCI_Malloc_local(bufsize);
if (rank == 0)
printf("ARMCI Strided Put Test:\n");
src_stride = XDIM * sizeof(double);
trg_stride = XDIM * sizeof(double);
stride_level = 1;
count[1] = YDIM;
count[0] = XDIM * sizeof(double);
ARMCI_Barrier();
peer = (rank+1) % nranks;
for (i = 0; i < ITERATIONS; i++) {
for (j = 0; j < XDIM*YDIM; j++) {
*(src_buf + j) = rank + i;
}
ARMCI_PutS(
src_buf,
&src_stride,
(void *) buffer[peer],
&trg_stride,
count,
stride_level,
peer);
}
ARMCI_Barrier();
ARMCI_Access_begin(buffer[rank]);
for (i = errors = 0; i < XDIM; i++) {
for (j = 0; j < YDIM; j++) {
const double actual = *(buffer[rank] + i + j*XDIM);
const double expected = (1.0 + rank) + (1.0 + ((rank+nranks-1)%nranks)) + (ITERATIONS);
if (actual - expected > 1e-10) {
printf("%d: Data validation failed at [%d, %d] expected=%f actual=%f\n",
rank, j, i, expected, actual);
errors++;
fflush(stdout);
}
}
}
ARMCI_Access_end(buffer[rank]);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Free_local(src_buf);
free(buffer);
ARMCI_Finalize();
MPI_Finalize();
if (errors == 0) {
printf("%d: Success\n", rank);
return 0;
} else {
printf("%d: Fail\n", rank);
return 1;
}
}
int main(int argc, char ** argv) {
int rank, nproc, i, test_iter;
int *my_data, *buf;
void **base_ptrs;
MPI_Init(&argc, &argv);
ARMCI_Init();
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
if (rank == 0) printf("Starting ARMCI test with %d processes\n", nproc);
buf = malloc(DATA_SZ);
base_ptrs = malloc(sizeof(void*)*nproc);
for (test_iter = 0; test_iter < NUM_ITERATIONS; test_iter++) {
if (rank == 0) printf(" + iteration %d\n", test_iter);
/*** Allocate the shared array ***/
ARMCI_Malloc(base_ptrs, DATA_SZ);
my_data = base_ptrs[rank];
/*** Get from our right neighbor and verify correct data ***/
ARMCI_Access_begin(my_data);
for (i = 0; i < DATA_NELTS; i++) my_data[i] = rank*test_iter;
ARMCI_Access_end(my_data);
ARMCI_Barrier(); // Wait for all updates to data to complete
ARMCI_Get(base_ptrs[(rank+1) % nproc], buf, DATA_SZ, (rank+1) % nproc);
for (i = 0; i < DATA_NELTS; i++) {
if (buf[i] != ((rank+1) % nproc)*test_iter) {
printf("%d: GET expected %d, got %d\n", rank, (rank+1) % nproc, buf[i]);
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
ARMCI_Barrier(); // Wait for all gets to complete
/*** Put to our left neighbor and verify correct data ***/
for (i = 0; i < DATA_NELTS; i++) buf[i] = rank*test_iter;
ARMCI_Put(buf, base_ptrs[(rank+nproc-1) % nproc], DATA_SZ, (rank+nproc-1) % nproc);
ARMCI_Barrier(); // Wait for all updates to data to complete
ARMCI_Access_begin(my_data);
for (i = 0; i < DATA_NELTS; i++) {
if (my_data[i] != ((rank+1) % nproc)*test_iter) {
printf("%d: PUT expected %d, got %d\n", rank, (rank+1) % nproc, my_data[i]);
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
ARMCI_Access_end(my_data);
ARMCI_Barrier(); // Wait for all gets to complete
/*** Accumulate to our left neighbor and verify correct data ***/
for (i = 0; i < DATA_NELTS; i++) buf[i] = rank;
ARMCI_Access_begin(my_data);
for (i = 0; i < DATA_NELTS; i++) my_data[i] = rank;
ARMCI_Access_end(my_data);
ARMCI_Barrier();
int scale = test_iter;
ARMCI_Acc(ARMCI_ACC_INT, &scale, buf, base_ptrs[(rank+nproc-1) % nproc], DATA_SZ, (rank+nproc-1) % nproc);
ARMCI_Barrier(); // Wait for all updates to data to complete
ARMCI_Access_begin(my_data);
for (i = 0; i < DATA_NELTS; i++) {
if (my_data[i] != rank + ((rank+1) % nproc)*test_iter) {
printf("%d: ACC expected %d, got %d\n", rank, (rank+1) % nproc, my_data[i]);
//MPI_Abort(MPI_COMM_WORLD, 1);
}
}
ARMCI_Access_end(my_data);
ARMCI_Free(my_data);
}
free(buf);
free(base_ptrs);
if (rank == 0) printf("Test complete: PASS.\n");
ARMCI_Finalize();
MPI_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,peer,j;
cpu_set_t mycpuid,new_mask;
char str[CPU_SETSIZE];
int rrr;
char cid[8];
extern char * cpuset_to_cstr(cpu_set_t *mask, char *str);
extern int cstr_to_cpuset(cpu_set_t *mask, const char* str);
gpc_hdl_t nbh;
char rheader[100];
int hlen, rhlen, rhsize;
int rdsize;
int rem;
void *header=&rem;
int locval=0;
void *loc=&locval;
int right;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD,&me);
MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
if(nprocs<2){
printf("\ncan run only on >=2 procs\n");
MPI_Finalize();
exit(1);
}
right = (me+1)%nprocs;
hlen=sizeof(header);
bzero(rheader,100);
rhlen = hlen;
ARMCI_Init();
accloop=atoi(argv[1]);
rem=accloop;
myptrs = (char **)malloc(sizeof(char *)*nprocs);
ARMCI_Malloc((void **)myptrs,size);
MPI_Barrier(MPI_COMM_WORLD);
gpcwork_memcpy = ARMCI_Gpc_register(gpc_work_handler_memcpy);
gpcwork_ddot =ARMCI_Gpc_register(gpc_work_handler_ddot);
gpcwork_daxpy = ARMCI_Gpc_register(gpc_work_handler_daxpy);
gpcwork_dgemm = ARMCI_Gpc_register(gpc_work_handler_dgemm);
MPI_Barrier(MPI_COMM_WORLD);
MPI_Barrier(MPI_COMM_WORLD);
ARMCI_Gpc_init_handle(&nbh);
if(ARMCI_Gpc_exec(gpcwork_memcpy, right, &header, hlen, loc, sizeof(int),
rheader, rhlen,loc, sizeof(int), &nbh))
fprintf(stderr,"ARMCI_Gpc_exec failed\n");
{
int m,n,k;
char notr='n';
DoubleComplex ZERO;
usleep(100);
ZERO.real=0.;ZERO.imag=0.;
m=n=k=DGS;
t0=MPI_Wtime();
#ifdef DGEMM_WORK
for(j=0;j<4*15;j++){
c_alpha=c_alpha+j*rand();
dgemm_(¬r,¬r,&m,&n,&k,&c_alpha,c_dga,&m,c_dgb,&n,&ZERO,c_dgc,&k,1,1);
}
#elif IUNIT_WORK
for(j=0;j<2*LOOP*100;j++){
for(i=0;i<LOOP*100;i++){
tmpbuf1[i]=tmpbuf1[i]*1.1214+i/tmpbuf1[j/2];
}
}
#elif DAXPY_WORK
for(j=0;j<tmp_loop*80;j++){
alpha=alpha+j*rand();
daxpy_(&N,&alpha,tmpbuf1,&ONE,tmpbuf2,&ONE);
}
#endif
t1=MPI_Wtime()-t0;
printf("\n%d:Compute_During_memcpy %d %f\n",me,accloop,t1);
}
ARMCI_Gpc_wait(&nbh);
MPI_Barrier(MPI_COMM_WORLD);
MPI_Barrier(MPI_COMM_WORLD);
ARMCI_Gpc_init_handle(&nbh);
if(ARMCI_Gpc_exec(gpcwork_ddot, right, &header, hlen, loc, sizeof(int),
rheader, rhlen,loc, sizeof(int), &nbh))
fprintf(stderr,"ARMCI_Gpc_exec failed\n");
{
int m,n,k;
char notr='n';
DoubleComplex ZERO;
usleep(100);
ZERO.real=0.;ZERO.imag=0.;
m=n=k=DGS;
t0=MPI_Wtime();
#ifdef DGEMM_WORK
for(j=0;j<4*15;j++){
c_alpha=c_alpha+j*rand();
dgemm_(¬r,¬r,&m,&n,&k,&c_alpha,c_dga,&m,c_dgb,&n,&ZERO,c_dgc,&k,1,1);
}
#elif IUNIT_WORK
for(j=0;j<2*LOOP*100;j++){
for(i=0;i<LOOP*100;i++){
tmpbuf1[i]=tmpbuf1[i]*1.1214+i/tmpbuf1[j/2];
}
}
#elif DAXPY_WORK
for(j=0;j<tmp_loop*80;j++){
alpha=alpha+j*rand();
daxpy_(&N,&alpha,tmpbuf1,&ONE,tmpbuf2,&ONE);
}
#endif
t1=MPI_Wtime()-t0;
printf("\n%d:Compute_During_Ddot %d %f\n",me,accloop,t1);
}
ARMCI_Gpc_wait(&nbh);
MPI_Barrier(MPI_COMM_WORLD);
MPI_Barrier(MPI_COMM_WORLD);
ARMCI_Gpc_init_handle(&nbh);
if(ARMCI_Gpc_exec(gpcwork_daxpy, right, &header, hlen, loc, sizeof(int),
rheader, rhlen,loc, sizeof(int), &nbh))
fprintf(stderr,"ARMCI_Gpc_exec failed\n");
{
int m,n,k;
char notr='n';
DoubleComplex ZERO;
usleep(100);
ZERO.real=0.;ZERO.imag=0.;
m=n=k=DGS;
t0=MPI_Wtime();
#ifdef DGEMM_WORK
for(j=0;j<4*15;j++){
c_alpha=c_alpha+j*rand();
dgemm_(¬r,¬r,&m,&n,&k,&c_alpha,c_dga,&m,c_dgb,&n,&ZERO,c_dgc,&k,1,1);
}
#elif IUNIT_WORK
for(j=0;j<2*LOOP*100;j++){
for(i=0;i<LOOP*100;i++){
tmpbuf1[i]=tmpbuf1[i]*1.1214+i/tmpbuf1[j/2];
}
}
#elif DAXPY_WORK
for(j=0;j<tmp_loop*80;j++){
alpha=alpha+j*rand();
daxpy_(&N,&alpha,tmpbuf1,&ONE,tmpbuf2,&ONE);
}
#endif
t1=MPI_Wtime()-t0;
printf("\n%d:Compute_During_Daxpy %d %f\n",me,accloop,t1);
}
ARMCI_Gpc_wait(&nbh);
MPI_Barrier(MPI_COMM_WORLD);
ARMCI_Gpc_init_handle(&nbh);
if(ARMCI_Gpc_exec(gpcwork_dgemm, right, &header, hlen, loc, sizeof(int),
rheader, rhlen,loc, sizeof(int), &nbh))
fprintf(stderr,"ARMCI_Gpc_exec failed\n");
{
int m,n,k;
char notr='n';
DoubleComplex ZERO;
usleep(100);
ZERO.real=0.;ZERO.imag=0.;
m=n=k=DGS;
t0=MPI_Wtime();
#ifdef DGEMM_WORK
for(j=0;j<4*15;j++){
c_alpha=c_alpha+j*rand();
dgemm_(¬r,¬r,&m,&n,&k,&c_alpha,c_dga,&m,c_dgb,&n,&ZERO,c_dgc,&k,1,1);
}
#elif IUNIT_WORK
for(j=0;j<2*LOOP*100;j++){
for(i=0;i<LOOP*100;i++){
tmpbuf1[i]=tmpbuf1[i]*1.1214+i/tmpbuf1[j/2];
}
}
#elif DAXPY_WORK
for(j=0;j<tmp_loop*80;j++){
alpha=alpha+j*rand();
daxpy_(&N,&alpha,tmpbuf1,&ONE,tmpbuf2,&ONE);
}
#endif
t1=MPI_Wtime()-t0;
printf("\n%d:Compute_During_Dgemm %d %f\n",me,accloop,t1);
}
ARMCI_Gpc_wait(&nbh);
MPI_Barrier(MPI_COMM_WORLD);
ARMCI_AllFence();
ARMCI_Finalize();
MPI_Finalize();
}
int main(int argc, char **argv) {
int me, nproc;
int i, *procs;
ARMCI_Group g_world, g_odd, g_even;
MPI_Init(&argc, &argv);
ARMCI_Init();
MPI_Comm_rank(MPI_COMM_WORLD, &me);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
procs = malloc(sizeof(int) * ( nproc/2 + (nproc % 2 ? 1 : 0 )));
if (me == 0) printf("ARMCI Group test starting on %d procs\n", nproc);
ARMCI_Group_get_world(&g_world);
if (me == 0) printf(" + Creating odd group\n");
for (i = 1; i < nproc; i += 2) {
procs[i/2] = i;
}
ARMCI_Group_create_child(i/2, procs, &g_odd, &g_world);
if (me == 0) printf(" + Creating even group\n");
for (i = 0; i < nproc; i += 2) {
procs[i/2] = i;
}
ARMCI_Group_create_child(i/2, procs, &g_even, &g_world);
/***********************************************************************/
{
int grp_me, grp_nproc;
double t_abs_to_grp, t_grp_to_abs;
const int iter = 1000000;
if (me == 0) {
ARMCI_Group_rank(&g_even, &grp_me);
ARMCI_Group_size(&g_even, &grp_nproc);
t_abs_to_grp = MPI_Wtime();
for (i = 0; i < iter; i++)
ARMCII_Translate_absolute_to_group(&g_even, (grp_me+1) % grp_nproc);
t_abs_to_grp = MPI_Wtime() - t_abs_to_grp;
t_grp_to_abs = MPI_Wtime();
for (i = 0; i < iter; i++)
ARMCI_Absolute_id(&g_even, (grp_me+1) % grp_nproc);
t_grp_to_abs = MPI_Wtime() - t_grp_to_abs;
printf("t_abs_to_grp = %f us, t_grp_to_abs = %f us\n", t_abs_to_grp/iter * 1.0e6, t_grp_to_abs/iter * 1.0e6);
}
ARMCI_Barrier();
}
/***********************************************************************/
if (me == 0) printf(" + Freeing groups\n");
if (me % 2 > 0)
ARMCI_Group_free(&g_odd);
else
ARMCI_Group_free(&g_even);
free(procs);
ARMCI_Finalize();
MPI_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 me,nproc;
int status;
int rank;
/* initialization */
MPI_Init(&argc, &argv);
ARMCI_Init();
#ifdef HPC_PROFILING
HPM_Init();
#endif
MPI_Comm_rank(MPI_COMM_WORLD,&me);
MPI_Comm_size(MPI_COMM_WORLD,&nproc);
#ifdef DEBUG
if(me == 0){
printf("The result of MPI_Comm_size is %d\n",nproc);
fflush(stdout);
}
#endif
/* get the matrix parameters */
if (argc > 1){
rank = atoi(argv[1]);
} else {
rank = 8;
}
if (me == 0){
printf("Running matmul.x with rank = %d\n",rank);
fflush(stdout);
}
/* register remote pointers */
double** addr_A = (double **) ARMCI_Malloc_local(sizeof(double *) * nproc);
if (addr_A == NULL) ARMCI_Error("malloc A failed at line",0);
double** addr_B = (double **) ARMCI_Malloc_local(sizeof(double *) * nproc);
if (addr_B == NULL) ARMCI_Error("malloc B failed at line",0);
double** addr_C = (double **) ARMCI_Malloc_local(sizeof(double *) * nproc);
if (addr_C == NULL) ARMCI_Error("malloc C failed at line",0);
#ifdef DEBUG
if(me == 0) printf("ARMCI_Malloc A requests %lu bytes\n",rank*rank*sizeof(double));
fflush(stdout);
#endif
status = ARMCI_Malloc((void **) addr_A, rank*rank*sizeof(double));
if (status != 0) ARMCI_Error("ARMCI_Malloc A failed",status);
#ifdef DEBUG
if(me == 0) printf("ARMCI_Malloc B requests %lu bytes\n",rank*rank*sizeof(double));
fflush(stdout);
#endif
status = ARMCI_Malloc((void **) addr_B, rank*rank*sizeof(double));
if (status != 0) ARMCI_Error("ARMCI_Malloc B failed",status);
#ifdef DEBUG
if(me == 0) printf("ARMCI_Malloc C requests %lu bytes\n",rank*rank*sizeof(double));
fflush(stdout);
#endif
status = ARMCI_Malloc((void **) addr_C, rank*rank*sizeof(double));
if (status != 0) ARMCI_Error("ARMCI_Malloc C failed",status);
MPI_Barrier(MPI_COMM_WORLD);
/* free ARMCI pointers */
ARMCI_Free_local(addr_C);
ARMCI_Free_local(addr_B);
ARMCI_Free_local(addr_A);
#ifdef HPC_PROFILING
HPM_Print();
#endif
/* the end */
ARMCI_Finalize();
MPI_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;
}
