void DDI_ARMCI_Finalize() {
int code;
const DDI_Comm *comm = (const DDI_Comm *) Comm_find(DDI_COMM_WORLD);
#if defined DDI_ARMCI_FREE
code = ARMCI_Free((void*)(gv(armci_mem_addr)[comm->me]));
if (code > 0) fprintf(stderr,"ARMCI_Free(%p) failed: %i",gv(armci_mem_addr)[comm->me]);
code = ARMCI_Free((void*)(gv(armci_cnt_addr)[comm->me]));
if (code > 0) fprintf(stderr,"ARMCI_Free(%p) failed: %i",gv(armci_cnt_addr)[comm->me]);
#endif
code = ARMCI_Destroy_mutexes();
if (code > 0) fprintf(stderr,"ARMCI_Destory_mutexes failed: %i",code);
ARMCI_Finalize();
}
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)
{
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();
}
/** @see ddi_armci.h */
void DDI_ARMCI_Memory_finalize() {
int code;
code = ARMCI_Free(gv(dda_index));
if (code != 0) {
fprintf(DDI_STDERR, "%s: ARMCI_Free(%p) returned %i\n", DDI_Id(), gv(dda_index), code);
DDI_Error(DDI_ARMCI_MEMORY_FINALIZE_ERROR, DDI_ARMCI_MEMORY_FINALIZE_ERROR_MESSAGE);
}
code = ARMCI_Destroy_mutexes();
if (code != 0) {
fprintf(DDI_STDERR, "%s: ARMCI_Destroy_mutexes() returned %i\n", DDI_Id(), code);
DDI_Error(DDI_ARMCI_MEMORY_FINALIZE_ERROR, DDI_ARMCI_MEMORY_FINALIZE_ERROR_MESSAGE);
}
}
int main(int argc, char **argv)
{
int i, j, rank, nranks, msgsize, dest;
int xdim, ydim;
long bufsize;
double **buffer;
double t_start, t_stop, t_latency;
int count[2], src_stride, trg_stride, stride_level;
int provided;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
ARMCI_Init_args(&argc, &argv);
ARMCI_Barrier();
int me = armci_msg_me();
int node = armci_domain_my_id(ARMCI_DOMAIN_SMP);
printf("MPI_Rank: %d, \
armci_msg_nproc: %d \
armci_msg_me: %d, \
armci_domain_id: %d, \
armci_domain_same_id: %d,\
armci_domain_my_id: %d, \
armci_domain_count: %d, \
armci_domain_nprocs: %d, \
armci_domain_glob_proc_id: %d \n",
rank, armci_msg_nproc(), me, armci_domain_id(ARMCI_DOMAIN_SMP, me),
armci_domain_same_id(ARMCI_DOMAIN_SMP, me), armci_domain_my_id(ARMCI_DOMAIN_SMP),
armci_domain_count(ARMCI_DOMAIN_SMP), armci_domain_nprocs(ARMCI_DOMAIN_SMP, node),
armci_domain_glob_proc_id(ARMCI_DOMAIN_SMP, node, 0));
fflush(stdout);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Finalize();
MPI_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;
}
void destroy_array(void *ptr[])
{
ARMCI_Barrier();
assert(!ARMCI_Free(ptr[me]));
}
void destroy_array(void *ptr[])
{
armci_msg_barrier();
assert(!ARMCI_Free(ptr[me]));
}
int main(int argc, char* argv[])
{
int provided;
int i, rank, nranks, msgsize, target;
long bufsize;
int **counter;
int *complete;
int increment;
int counter_fetch;
int counters_received;
int t_start, t_stop, t_latency;
int expected;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
ARMCI_Init_args(&argc, &argv);
complete = (int *) malloc(sizeof(int) * COUNT);
counter = (int**) ARMCI_Malloc_local( nranks * sizeof(int*) );
ARMCI_Malloc((void *) counter[rank], sizeof(int));
if (rank == 0)
{
printf("ARMCI_RMW Test - in usec \n");
fflush(stdout);
}
target = 0;
for(i=0; i<COUNT; i++)
{
complete[i] = 0;
}
if(rank == target)
{
*(counter[rank]) = 0;
}
increment = 1;
counter_fetch = 0;
counters_received = 0;
MPI_Barrier(MPI_COMM_WORLD);
while(counter_fetch < COUNT)
{
ARMCI_Rmw(ARMCI_FETCH_AND_ADD,
(void *) &counter_fetch,
(void *) counter[target],
increment,
target);
/* s/1/rank/ means we will know who got the counter */
if (counter_fetch < COUNT) complete[counter_fetch] = rank;
counters_received++;
}
MPI_Allreduce(MPI_IN_PLACE,complete,COUNT,MPI_INT,MPI_SUM,MPI_COMM_WORLD);
for(i=0; i<COUNT; i++)
{
if (complete[i] == 0)
{
printf("[%d] The RMW update failed at index: %d \n", rank, i);
fflush(stdout);
exit(-1);
}
}
printf("[%d] The RMW update completed successfully \n", rank);
fflush(stdout);
MPI_Barrier(MPI_COMM_WORLD);
if (0==rank)
{
printf("Checking for fairness...\n", rank);
fflush(stdout);
for(i=0; i<COUNT; i++)
{
printf("counter value %d was received by process %d\n", i, complete[i]);
}
fflush(stdout);
}
MPI_Barrier(MPI_COMM_WORLD);
printf("process %d received %d counters\n", rank, counters_received);
fflush(stdout);
ARMCI_Free(counter[rank]);
ARMCI_Free_local(counter);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
void test_2D()
{
int i;
int src, dst;
int ierr;
double *buf;
void *ptr[MAXPROC], *get_ptr[MAXPROC];
/* find who I am and the dst process */
src = me;
#ifdef MALLOC_LOC
if(me == 0) {
buf = (double *)ARMCI_Malloc_local(SIZE * SIZE * sizeof(double));
assert(buf != NULL);
}
#else
if(me == 0) {
buf = (double *)malloc(SIZE * SIZE * sizeof(double));
assert(buf != NULL);
}
#endif
ierr = ARMCI_Malloc(ptr, (SIZE * SIZE * sizeof(double)));
assert(ierr == 0); assert(ptr[me]);
ierr = ARMCI_Malloc(get_ptr, (SIZE * SIZE * sizeof(double)));
assert(ierr == 0); assert(get_ptr[me]);
/* ARMCI - initialize the data window */
fill_array(ptr[me], SIZE*SIZE, me);
fill_array(get_ptr[me], SIZE*SIZE, me);
MP_BARRIER();
/* only the proc 0 doest the work */
/* print the title */
if(me == 0) {
if(!CHECK_RESULT){
printf(" section get put");
printf(" acc\n");
printf("bytes loop sec MB/s sec MB/s");
printf(" sec MB/s\n");
printf("------- ------ -------- -------- -------- --------");
printf(" -------- --------\n");
fflush(stdout);
}
for(i=0; i<CHUNK_NUM; i++) {
int loop;
int bytes = chunk[i] * chunk[i] * sizeof(double);
double t_get = 0, t_put = 0, t_acc = 0;
double latency_get, latency_put, latency_acc;
double bandwidth_get, bandwidth_put, bandwidth_acc;
loop = SIZE / chunk[i];
if(loop<2)loop=2;
for(dst=1; dst<nproc; dst++) {
/* strided get */
fill_array(buf, SIZE*SIZE, me*10);
t_get += time_get((double *)(get_ptr[dst]), (double *)buf,
chunk[i], loop, dst, 1);
/* strided put */
fill_array(buf, SIZE*SIZE, me*10);
t_put += time_put((double *)buf, (double *)(ptr[dst]),
chunk[i], loop, dst, 1);
/* strided acc */
fill_array(buf, SIZE*SIZE, me*10);
t_acc += time_acc((double *)buf, (double *)(ptr[dst]),
chunk[i], loop, dst, 1);
}
latency_get = t_get/(nproc - 1);
latency_put = t_put/(nproc - 1);
latency_acc = t_acc/(nproc - 1);
bandwidth_get = (bytes * (nproc - 1) * 1e-6)/t_get;
bandwidth_put = (bytes * (nproc - 1) * 1e-6)/t_put;
bandwidth_acc = (bytes * (nproc - 1) * 1e-6)/t_acc;
/* print */
if(!CHECK_RESULT)printf("%d\t%d\t%.2e %.2e %.2e %.2e %.2e %.2e\n",
bytes, loop, latency_get, bandwidth_get,
latency_put, bandwidth_put, latency_acc, bandwidth_acc);
}
}
else sleep(3);
ARMCI_AllFence();
MP_BARRIER();
/* cleanup */
ARMCI_Free(get_ptr[me]);
ARMCI_Free(ptr[me]);
#ifdef MALLOC_LOC
if(me == 0) ARMCI_Free_local(buf);
#else
if(me == 0) free(buf);
#endif
}
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 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, j, rank, nranks, peer;
size_t xdim, ydim;
unsigned long bufsize;
double **buffer, *src_buf;
double t_start=0.0, t_stop;
int count[2], src_stride, trg_stride, stride_level;
double scaling;
int provided;
MPI_Init_thread(&argc, &argv, MPI_THREAD_SINGLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
if (nranks < 2) {
printf("%s: Must be run with at least 2 processes\n", argv[0]);
MPI_Abort(MPI_COMM_WORLD, 1);
}
ARMCI_Init_args(&argc, &argv);
buffer = (double **) malloc(sizeof(double *) * nranks);
bufsize = MAX_XDIM * MAX_YDIM * sizeof(double);
ARMCI_Malloc((void **) buffer, bufsize);
src_buf = ARMCI_Malloc_local(bufsize);
if (rank == 0)
{
printf("ARMCI_AccS Latency - local and remote completions - in usec \n");
printf("%30s %22s %22s\n",
"Dimensions(array of double)",
"Local Completion",
"Remote completion");
fflush(stdout);
}
ARMCI_Access_begin(buffer[rank]);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
*(src_buf + i) = 1.0 + rank;
}
ARMCI_Access_end(buffer[rank]);
scaling = 2.0;
src_stride = MAX_YDIM * sizeof(double);
trg_stride = MAX_YDIM * sizeof(double);
stride_level = 1;
ARMCI_Barrier();
for (xdim = 1; xdim <= MAX_XDIM; xdim *= 2)
{
count[1] = xdim;
for (ydim = 1; ydim <= MAX_YDIM; ydim *= 2)
{
count[0] = ydim * sizeof(double);
if (rank == 0)
{
peer = 1;
for (i = 0; i < ITERATIONS + SKIP; i++)
{
if (i == SKIP) t_start = MPI_Wtime();
ARMCI_AccS(ARMCI_ACC_DBL,
(void *) &scaling,
/* (void *) buffer[rank] */ src_buf,
&src_stride,
(void *) buffer[peer],
&trg_stride,
count,
stride_level,
1);
}
t_stop = MPI_Wtime();
ARMCI_Fence(1);
char temp[10];
sprintf(temp, "%dX%d", (int) xdim, (int) ydim);
printf("%30s %20.2f ", temp, ((t_stop - t_start) * 1000000)
/ ITERATIONS);
fflush(stdout);
ARMCI_Barrier();
ARMCI_Barrier();
for (i = 0; i < ITERATIONS + SKIP; i++)
{
if (i == SKIP) t_start = MPI_Wtime();
ARMCI_AccS(ARMCI_ACC_DBL,
(void *) &scaling,
/* (void *) buffer[rank] */ src_buf,
&src_stride,
(void *) buffer[peer],
&trg_stride,
count,
stride_level,
1);
ARMCI_Fence(1);
}
t_stop = MPI_Wtime();
printf("%20.2f \n", ((t_stop - t_start) * 1000000) / ITERATIONS);
fflush(stdout);
ARMCI_Barrier();
ARMCI_Barrier();
}
else
{
peer = 0;
ARMCI_Barrier();
if (rank == 1)
{
ARMCI_Access_begin(buffer[rank]);
for (i = 0; i < xdim; i++)
{
for (j = 0; j < ydim; j++)
{
if (*(buffer[rank] + i * MAX_XDIM + j) != ((1.0 + rank)
+ scaling * (1.0 + peer) * (ITERATIONS + SKIP)))
{
printf("Data validation failed at X: %d Y: %d Expected : %f Actual : %f \n",
i,
j,
((1.0 + rank) + scaling * (1.0 + peer)),
*(buffer[rank] + i * MAX_YDIM + j));
fflush(stdout);
ARMCI_Error("Bailing out", 1);
}
}
}
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_Access_end(buffer[rank]);
}
ARMCI_Barrier();
ARMCI_Barrier();
if (rank == 1)
{
ARMCI_Access_begin(buffer[rank]);
for (i = 0; i < xdim; i++)
{
for (j = 0; j < ydim; j++)
{
if (*(buffer[rank] + i * MAX_XDIM + j) != ((1.0 + rank)
+ scaling * (1.0 + peer) * (ITERATIONS + SKIP)))
{
printf("Data validation failed at X: %d Y: %d Expected : %f Actual : %f \n",
i,
j,
((1.0 + rank) + scaling * (1.0 + peer)),
*(buffer[rank] + i * MAX_YDIM + j));
fflush(stdout);
ARMCI_Error("Bailing out", 1);
}
}
}
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_Access_end(buffer[rank]);
}
ARMCI_Barrier();
}
}
}
ARMCI_Barrier();
ARMCI_Free((void *) buffer[rank]);
ARMCI_Free_local(src_buf);
free(buffer);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
int main(int argc, char *argv[])
{
size_t i, rank, nranks, msgsize, dest;
size_t iterations, max_msgsize;
int bufsize;
double **buffer;
double t_start, t_stop, t_total, d_total;
double expected, bandwidth;
int provided;
armci_hdl_t handle;
max_msgsize = MAX_MSGSIZE;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
ARMCI_Init_args(&argc, &argv);
bufsize = max_msgsize * ITERATIONS_LARGE;
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_INIT_HANDLE(&handle);
ARMCI_SET_AGGREGATE_HANDLE(&handle);
ARMCI_Barrier();
if (rank == 0)
{
printf("ARMCI_Put Bandwidth in MBPS \n");
printf("%20s %22s \n", "Message Size", "Bandwidth");
fflush(stdout);
dest = 1;
expected = 1 + dest;
for (msgsize = sizeof(double); msgsize <= max_msgsize; msgsize *= 2)
{
if (msgsize <= 16 * 1024) iterations = ITERATIONS_VERYSMALL;
else if (msgsize <= 64 * 1024) iterations = ITERATIONS_SMALL;
else if (msgsize <= 512 * 1024) iterations = ITERATIONS_MEDIUM;
else iterations = ITERATIONS_LARGE;
t_start = MPI_Wtime();
for (i = 0; i < iterations; i++)
{
ARMCI_NbPut((void *) ((size_t) buffer[dest] + (size_t)(i
* msgsize)), (void *) ((size_t) buffer[rank]
+ (size_t)(i * msgsize)), msgsize, dest, &handle);
}
ARMCI_Wait(&handle);
t_stop = MPI_Wtime();
d_total = (iterations * msgsize) / (1024 * 1024);
t_total = t_stop - t_start;
bandwidth = d_total / t_total;
printf("%20d %20.4lf \n", msgsize, bandwidth);
fflush(stdout);
ARMCI_Fence(dest);
}
}
ARMCI_Barrier();
ARMCI_UNSET_AGGREGATE_HANDLE(&handle);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
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;
}
}
void TRANSPOSE1D() {
int dims[1];
int nelem, i, ierr, min, max, cmin, cmax, lmin, lmax, pmin, pmax;
int src_offset, dst_offset, length;
int *buf, *map;
void *src_ptr, *dst_ptr;
void **a_ptr, **b_ptr;
int *a, *b;
/* Find local processor ID and number of processors */
int me, nprocs;
me = armci_msg_me();
nprocs = armci_msg_nproc();
/* Allocate pointers to data on all processors */
a_ptr = (void**)malloc(nprocs*sizeof(int*));
b_ptr = (void**)malloc(nprocs*sizeof(int*));
map = (int*)malloc(nprocs*sizeof(int));
/* Configure array dimensions. Force an unequal data distribution */
dims[0] = nprocs*TOTALELEMS + nprocs/2;
if (me == 0) printf("Size of array: %d\n\n",dims[0]);
/* Find first (zero-based) index of chunk owned by each processor and
store it in map array */
for (i=0; i<nprocs; i++) {
map[i] = (int)(((double)i)*(((double)dims[0])/((double)nprocs)));
}
/* Figure out what size my portion of array is */
if (me<nprocs-1) {
nelem = map[me+1]-map[me];
} else {
nelem = dims[0]-map[me];
}
/* Allocate memory for array A */
ierr = ARMCI_Malloc(a_ptr, nelem*sizeof(int));
assert(ierr == 0);
assert(a_ptr[me]);
/* Allocate memory for array B */
ierr = ARMCI_Malloc(b_ptr, nelem*sizeof(int));
assert(ierr == 0);
assert(b_ptr[me]);
/* initialize data in array A and zero data in array B */
a = (int*)a_ptr[me];
b = (int*)b_ptr[me];
for (i=0; i<nelem; i++) {
a[i] = i + map[me] + 1;
b[i] = 0;
}
/* Synchronize all processors to guarantee that everyone has data
before proceeding to the next step. */
armci_msg_barrier();
/* Create local buffer for performing inversion */
buf = (int*)malloc(nelem*sizeof(int));
/* Copy inverted data into local buffer */
a = (int*)a_ptr[me];
for (i=0; i<nelem; i++) {
buf[i] = a[nelem-i-1];
}
/* Find out which blocks of array B inverted block should be copied to.
Start by finding min and max indices of data in array B*/
min = dims[0] - (map[me] + nelem);
max = dims[0] - map[me] - 1;
/* Locate processors containing the endpoints */
pmin = 0;
for (i=0; i<nprocs; i++) {
if (min >= map[i]) {
pmin = i;
} else {
break;
}
}
pmax = nprocs-1;
for (i=nprocs-2; i>=0; i--) {
if (max < map[i+1]) {
pmax = i;
} else {
break;
}
}
/* Loop over processors that will receive data and copy inverted data to
processors */
for (i=pmin; i<=pmax; i++) {
/* Find min and max indices owned by processor i */
lmin = map[i];
if (i<nprocs-1) {
lmax = map[i+1]-1;
} else {
lmax = dims[0]-1;
}
/* Find min and max indices that should be sent to processor i */
if (lmin > min) {
cmin = lmin;
} else {
cmin = min;
}
if (lmax < max) {
cmax = lmax;
} else {
cmax = max;
}
/* Find offsets on source and destination processors */
src_offset = cmin - min;
src_ptr = (void*)(buf + src_offset);
dst_offset = cmin - lmin;
dst_ptr = ((char*)b_ptr[i]) + sizeof(int)*dst_offset;
/* Find length of data (in bytes) to be sent to processor i */
length = sizeof(int)*(cmax-cmin+1);
/* Send data to processor */
ARMCI_Put(src_ptr, dst_ptr, length, i);
}
ARMCI_AllFence();
armci_msg_barrier();
free(buf);
VERIFY(b_ptr, dims, map);
free(map);
armci_msg_barrier();
ARMCI_Free(a_ptr[me]);
ARMCI_Free(b_ptr[me]);
free(a_ptr);
free(b_ptr);
}
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();
}
static void contig_test(size_t buffer_size, int op)
{
void **dst_ptr;
void **put_buf;
void **get_buf;
double *times;
dst_ptr = (void*)malloc(nproc * sizeof(void*));
put_buf = (void*)malloc(nproc * sizeof(void*));
get_buf = (void*)malloc(nproc * sizeof(void*));
times = (double*)malloc(nproc * sizeof(double));
ARMCI_Malloc(dst_ptr, buffer_size);
ARMCI_Malloc(put_buf, buffer_size);
ARMCI_Malloc(get_buf, buffer_size);
/* initialize what we're putting */
fill_array((double*)put_buf[me], buffer_size/sizeof(double), me);
size_t msg_size;
int dst = 1;
double scale = 1.0;
for (msg_size = 16; msg_size <= buffer_size; msg_size *= 2) {
int j;
int iter = msg_size > MEDIUM_MESSAGE_SIZE ? ITER_LARGE : ITER_SMALL;
double t_start, t_end;
if (0 == me) {
for (j= 0; j < iter + WARMUP; ++j) {
if (WARMUP == j) {
t_start = dclock();
}
switch (op) {
case PUT:
ARMCI_Put(put_buf[me], dst_ptr[dst], msg_size,
dst);
break;
case GET:
ARMCI_Get(dst_ptr[dst], get_buf[me], msg_size,
dst);
break;
case ACC:
ARMCI_Acc(ARMCI_ACC_DBL, &scale,
put_buf[me], dst_ptr[dst], msg_size,
dst);
break;
default:
ARMCI_Error("oops", 1);
}
}
}
/* calculate total time and average time */
t_end = dclock();
ARMCI_Barrier();
if (0 == me) {
printf("%8zu\t\t%6.2f\t\t%10.2f\n",
msg_size,
((t_end - t_start))/iter,
msg_size*iter/((t_end - t_start)));
}
}
ARMCI_Free(dst_ptr[me]);
ARMCI_Free(put_buf[me]);
ARMCI_Free(get_buf[me]);
free(dst_ptr);
free(put_buf);
free(get_buf);
free(times);
}
int main(int argc, char **argv)
{
int i, rank, nranks, msgsize, dest;
long bufsize;
double **buffer;
double t_start, t_stop, t_latency;
int provided;
ARMCI_Init_args(&argc, &argv);
rank = A1_Process_id(A1_GROUP_WORLD);
nranks = A1_Process_total(A1_GROUP_WORLD);
bufsize = MAX_MSG_SIZE * (ITERATIONS + SKIP);
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
A1_Barrier_group(A1_GROUP_WORLD);
if (rank == 0)
{
printf("ARMCI_Get Latency in usec \n");
printf("%20s %22s \n", "Message Size", "Latency");
fflush(stdout);
dest = 1;
for (msgsize = sizeof(double); msgsize <= MAX_MSG_SIZE; msgsize *= 2)
{
for (i = 0; i < ITERATIONS + SKIP; i++)
{
if (i == SKIP) t_start = A1_Time_seconds();
ARMCI_Get((void *) ((size_t) buffer[dest] + (size_t)(i
* msgsize)), (void *) ((size_t) buffer[rank]
+ (size_t)(i * msgsize)), msgsize, 1);
}
t_stop = A1_Time_seconds();
printf("%20d %20.2f \n", msgsize, ((t_stop - t_start) * 1000000)
/ ITERATIONS);
fflush(stdout);
for (i = 0; i < ((ITERATIONS + SKIP) * msgsize) / sizeof(double); i++)
{
if (*(buffer[rank] + i) != (1.0 + dest))
{
printf("Data validation failed At displacement : %d Expected : %f Actual : %f \n",
i,
(1.0 + dest),
*(buffer[rank] + i));
fflush(stdout);
return -1;
}
}
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
}
}
A1_Barrier_group(A1_GROUP_WORLD);
ARMCI_Free(buffer[rank]);
ARMCI_Finalize();
return 0;
}
int main(int argc, char *argv[])
{
int rank, nranks;
size_t i, msgsize, dest;
size_t iterations, max_msgsize;
int bufsize;
double **buffer;
double t_start, t_stop, t_total, d_total;
double expected, bandwidth;
int provided;
armci_hdl_t handle;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
max_msgsize = MAX_MSGSIZE;
ARMCI_Init_args(&argc, &argv);
bufsize = max_msgsize * ITERATIONS;
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_INIT_HANDLE(&handle);
ARMCI_SET_AGGREGATE_HANDLE(&handle);
ARMCI_Barrier();
if (rank == 0)
{
printf("ARMCI_Get Bandwidth in MBPS \n");
printf("%20s %22s \n", "Message Size", "Bandwidth");
fflush(stdout);
dest = 1;
expected = 1 + dest;
for (msgsize = sizeof(double); msgsize <= max_msgsize; msgsize *= 2)
{
iterations = bufsize/msgsize;
t_start = MPI_Wtime();
for (i = 0; i < iterations; i++)
{
ARMCI_NbGet((void *) ((size_t) buffer[dest] + (size_t)(i
* msgsize)), (void *) ((size_t) buffer[rank]
+ (size_t)(i * msgsize)), msgsize, dest, &handle);
}
ARMCI_Wait(&handle);
t_stop = MPI_Wtime();
d_total = (iterations * msgsize) / (1024 * 1024);
t_total = t_stop - t_start;
bandwidth = d_total / t_total;
printf("%20d %20.4lf \n", msgsize, bandwidth);
fflush(stdout);
#ifdef DATA_VALIDATION
{
for(j=0; j<((iterations*msgsize)/sizeof(double)); j++)
{
if(*(buffer[rank] + j) != expected)
{
printf("Data validation failed At displacement : %d Expected : %lf Actual : %lf \n",
j, expected, *(buffer[rank] + j));
fflush(stdout);
return -1;
}
}
for(j=0; j<bufsize/sizeof(double); j++)
{
*(buffer[rank] + j) = 1.0 + rank;
}
}
#endif
}
}
ARMCI_Barrier();
ARMCI_UNSET_AGGREGATE_HANDLE(&handle);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
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 i, j, rank, nranks;
int xdim, ydim;
long bufsize;
double **buffer;
double t_start=0.0, t_stop=0.0;
int count[2], src_stride, trg_stride, stride_level, peer;
double expected, actual;
int provided;
MPI_Init_thread(&argc, &argv, MPI_THREAD_SINGLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
if (nranks < 2) {
printf("%s: Must be run with at least 2 processes\n", argv[0]);
MPI_Abort(MPI_COMM_WORLD, 1);
}
ARMCI_Init_args(&argc, &argv);
bufsize = MAX_XDIM * MAX_YDIM * sizeof(double);
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for(i=0; i< bufsize/sizeof(double); i++) {
*(buffer[rank] + i) = 1.0 + rank;
}
if(rank == 0) {
printf("ARMCI_PutS Latency - local and remote completions - in usec \n");
printf("%30s %22s %22s\n", "Dimensions(array of doubles)", "Latency-LocalCompeltion", "Latency-RemoteCompletion");
fflush(stdout);
}
src_stride = MAX_YDIM*sizeof(double);
trg_stride = MAX_YDIM*sizeof(double);
stride_level = 1;
ARMCI_Barrier();
for(xdim=1; xdim<=MAX_XDIM; xdim*=2) {
count[1] = xdim;
for(ydim=1; ydim<=MAX_YDIM; ydim*=2) {
count[0] = ydim*sizeof(double);
if(rank == 0)
{
peer = 1;
for(i=0; i<ITERATIONS+SKIP; i++) {
if(i == SKIP)
t_start = MPI_Wtime();
ARMCI_PutS((void *) buffer[rank], &src_stride, (void *) buffer[peer], &trg_stride, count, stride_level, peer);
}
t_stop = MPI_Wtime();
ARMCI_Fence(peer);
char temp[10];
sprintf(temp,"%dX%d", xdim, ydim);
printf("%30s %20.2f", temp, ((t_stop-t_start)*1000000)/ITERATIONS);
fflush(stdout);
ARMCI_Barrier();
ARMCI_Barrier();
for(i=0; i<ITERATIONS+SKIP; i++) {
if(i == SKIP)
t_start = MPI_Wtime();
ARMCI_PutS((void *) buffer[rank], &src_stride, (void *) buffer[peer], &trg_stride, count, stride_level, peer);
ARMCI_Fence(peer);
}
t_stop = MPI_Wtime();
printf("%20.2f \n", ((t_stop-t_start)*1000000)/ITERATIONS);
fflush(stdout);
ARMCI_Barrier();
ARMCI_Barrier();
}
else
{
peer = 0;
expected = (1.0 + (double) peer);
ARMCI_Barrier();
if (rank == 1)
{
for(i=0; i<xdim; i++)
{
for(j=0; j<ydim; j++)
{
actual = *(buffer[rank] + i*MAX_YDIM + j);
if(actual != expected)
{
printf("Data validation failed at X: %d Y: %d Expected : %f Actual : %f \n",
i, j, expected, actual);
fflush(stdout);
ARMCI_Error("Bailing out", 1);
}
}
}
}
for(i=0; i< bufsize/sizeof(double); i++) {
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_Barrier();
ARMCI_Barrier();
if (rank == 1)
{
for(i=0; i<xdim; i++)
{
for(j=0; j<ydim; j++)
{
actual = *(buffer[rank] + i*MAX_YDIM + j);
if(actual != expected)
{
printf("Data validation failed at X: %d Y: %d Expected : %f Actual : %f \n",
i, j, expected, actual);
fflush(stdout);
ARMCI_Error("Bailing out", 1);
}
}
}
for(i=0; i< bufsize/sizeof(double); i++) {
*(buffer[rank] + i) = 1.0 + rank;
}
}
ARMCI_Barrier();
}
}
}
ARMCI_Barrier();
ARMCI_Free((void *) buffer[rank]);
free(buffer);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
void destroy_array(void *ptr[])
{
MP_BARRIER();
assert(!ARMCI_Free(ptr[me]));
}
int main(int argc, char *argv[])
{
int i, j, rank, nranks, msgsize, dest;
int dim, iterations;
long bufsize;
double **buffer;
double t_start, t_stop, t_total, d_total, bw;
int count[2], src_stride, trg_stride, stride_level;
int provided;
armci_hdl_t handle;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
ARMCI_Init_args(&argc, &argv);
bufsize = MAX_DIM * MAX_DIM * sizeof(double);
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_INIT_HANDLE(&handle);
ARMCI_SET_AGGREGATE_HANDLE(&handle);
ARMCI_Barrier();
if (rank == 0)
{
printf("ARMCI_PutS Bandwidth in MBPS \n");
printf("%30s %22s \n", "Dimensions(array of doubles)", "Latency");
fflush(stdout);
dest = 1;
src_stride = MAX_DIM * sizeof(double);
trg_stride = MAX_DIM * sizeof(double);
stride_level = 1;
for (dim = 1; dim <= MAX_DIM; dim *= 2)
{
count[0] = dim*sizeof(double);
count[1] = dim;
iterations = 10*(MAX_DIM * MAX_DIM)/(dim * dim);
t_start = MPI_Wtime();
for (i = 0; i < iterations; i++)
{
ARMCI_NbPutS((void *) buffer[rank],
&src_stride,
(void *) buffer[dest],
&trg_stride,
count,
stride_level,
dest,
&handle);
}
ARMCI_Wait(&handle);
t_stop = MPI_Wtime();
ARMCI_Fence(1);
char temp[10];
sprintf(temp, "%dX%d", dim, dim);
t_total = t_stop - t_start;
d_total = (dim*dim*sizeof(double)*iterations)/(1024*1024);
bw = d_total/t_total;
printf("%30s %20.2f \n", temp, bw);
fflush(stdout);
}
}
ARMCI_Barrier();
ARMCI_UNSET_AGGREGATE_HANDLE(&handle);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
main(int argc, char *argv[])
{
int i, j;
int ch;
extern char *optarg;
int edge;
int size;
int nloop=5;
double **ptr_loc;
MP_INIT(arc,argv);
MP_PROCS(&nproc);
MP_MYID(&me);
while ((ch = getopt(argc, argv, "n:b:p:h")) != -1) {
switch(ch) {
case 'n':
n = atoi(optarg);
break;
case 'b':
block_size = atoi(optarg);
break;
case 'p':
nproc = atoi(optarg);
break;
case 'h': {
printf("Usage: LU, or \n");
printf(" LU -nMATRIXSIZE -bBLOCKSIZE -pNPROC\n");
MP_BARRIER();
MP_FINALIZE();
exit(0);
}
}
}
if(me == 0) {
printf("\n Blocked Dense LU Factorization\n");
printf(" %d by %d Matrix\n", n, n);
printf(" %d Processors\n", nproc);
printf(" %d by %d Element Blocks\n", block_size, block_size);
printf("\n");
}
num_rows = (int) sqrt((double) nproc);
for (;;) {
num_cols = nproc/num_rows;
if (num_rows*num_cols == nproc)
break;
num_rows--;
}
nblocks = n/block_size;
if (block_size * nblocks != n) {
nblocks++;
}
edge = n%block_size;
if (edge == 0) {
edge = block_size;
}
#ifdef DEBUG
if(me == 0)
for (i=0; i<nblocks; i++) {
for (j=0; j<nblocks; j++)
printf("%d ", block_owner(i, j));
printf("\n");
}
MP_BARRIER();
MP_FINALIZE();
exit(0);
#endif
for (i=0; i<nblocks; i++) {
for (j=0; j<nblocks; j++) {
if(block_owner(i,j) == me) {
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
}
else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
}
else {
size = block_size*block_size;
}
proc_bytes += size*sizeof(double);
}
}
}
ptr = (void **)malloc(nproc * sizeof(void *));
#ifdef MPI2_ONESIDED
MPI_Alloc_mem(proc_bytes, MPI_INFO_NULL, &ptr[me]);
MPI_Win_create((void*)ptr[me], proc_bytes, 1, MPI_INFO_NULL,
MPI_COMM_WORLD, &win);
for(i=0; i<nproc; i++) ptr[i] = (double *)ptr[me];
MPI_Barrier(MPI_COMM_WORLD);
#else
/* initialize ARMCI */
ARMCI_Init();
ARMCI_Malloc(ptr, proc_bytes);
#endif
a = (double **)malloc(nblocks*nblocks*sizeof(double *));
if (a == NULL) {
fprintf(stderr, "Could not malloc memory for a\n");
exit(-1);
}
ptr_loc = (double **)malloc(nproc*sizeof(double *));
for(i=0; i<nproc; i++) ptr_loc[i] = (double *)ptr[i];
for(i=0; i<nblocks; i ++) {
for(j=0; j<nblocks; j++) {
a[i+j*nblocks] = ptr_loc[block_owner(i, j)];
if ((i == nblocks-1) && (j == nblocks-1)) {
size = edge*edge;
} else if ((i == nblocks-1) || (j == nblocks-1)) {
size = edge*block_size;
} else {
size = block_size*block_size;
}
ptr_loc[block_owner(i, j)] += size;
}
}
/* initialize the array */
init_array();
/* barrier to ensure all initialization is done */
MP_BARRIER();
/* to remove cold-start misses, all processors touch their own data */
touch_array(block_size, me);
MP_BARRIER();
if(doprint) {
if(me == 0) {
printf("Matrix before LU decomposition\n");
print_array(me);
}
MP_BARRIER();
}
lu(n, block_size, me); /* cold start */
/* Starting the timer */
MP_BARRIER();
if(me == 0) start_timer();
for(i=0; i<nloop; i++) lu(n, block_size, me);
MP_BARRIER();
/* Timer Stops here */
if(me == 0)
printf("\nRunning time = %lf milliseconds.\n\n", elapsed_time()/nloop);
printf("%d: (ngets=%d) Communication (get) time = %e milliseconds\n", me, get_cntr, comm_time*1000/nloop);
if(doprint) {
if(me == 0) {
printf("after LU\n");
print_array(me);
}
MP_BARRIER();
}
/* done */
#ifdef MPI2_ONESIDED
MPI_Win_free(&win);
MPI_Free_mem(ptr[me]);
#else
ARMCI_Free(ptr[me]);
ARMCI_Finalize();
#endif
MP_FINALIZE();
}
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);
}