/*\ quick fix for inacurate timer
\*/
double Timer()
{
#define DELTA 0.000001
double t=MP_TIMER();
if(t<=_tt0 + DELTA) _tt0 += DELTA;
else _tt0 = t;
return _tt0;
}
void time_reduce(double *test,int len)
{
int i;
double t;
t = MP_TIMER();
for(i=0; i<LOOP; i++){
armci_msg_reduce(test, len, "+",ARMCI_DOUBLE);
}
t = MP_TIMER() -t;
t /= LOOP;
if(me==0){
printf("Time per reduce %lf len=%d doubles\n",t, len);
fflush(stdout);
}
}
void time_gop(double *test,int len)
{
int i;
double t;
t = MP_TIMER();
for(i=0; i<LOOP; i++){
armci_msg_dgop(test, len, "+");
}
t = MP_TIMER() -t;
t /= LOOP;
if(me==0){
printf("Time per gop %lf len=%d doubles\n",t, len);
fflush(stdout);
}
}
static int armci_profile_close_event(int event_type, int range, double *time,
char *name) {
int curr_event = gCURRENT_EVNT.event_type;
#ifdef DEBUG
if(SERVER_CONTEXT)
printf("\n%d(s):call profile close for %s isset is %d",armci_me,
gEventName[event_type],gCURRENT_EVNT.is_set);
else
printf("\n%d:call profile close for %s isset is %d",armci_me,
gEventName[event_type],gCURRENT_EVNT.is_set);
fflush(stdout);
#endif
if(gCURRENT_EVNT.is_set==1) { /* Yep, there is an event set. So close it.*/
/*Check if "profile stop" is called for corresponding "profile start"*/
if(event_type != curr_event) {
printf(
"%d: %s: ERROR:Profile started for %s, but stopped for %s\n",
armci_me,name,gEventName[curr_event],gEventName[event_type]);
fflush(stdout);
armci_die("Profile_stop is called a different event", armci_me);
}
*time = MP_TIMER() - gCURRENT_EVNT.start_time;
ARMCI_PROF[curr_event][range].time += *time;
gCURRENT_EVNT.is_set = 0; /* close the event */
return ARMCI_EVENT_CLOSED;
}
else { /* event overlapping */
gCURRENT_EVNT.is_set--;
if(gCURRENT_EVNT.is_set<=0) {
char *msg="Profile_stop is called before profile_start";
printf("%d: %s: ERROR: %s. Event Name = %s\n", armci_me,
name, msg, gEventName[curr_event]);
fflush(stdout);
armci_die(" profile_stop is called before profile_start", armci_me);
}
}
return ARMCI_EVENT_NOTCLOSED;
}
static int armci_profile_set_event(int event_type, int range) {
#ifdef DEBUG
if(SERVER_CONTEXT)
printf("\n%d(s):call profile set for %s isset is %d",armci_me,
gEventName[event_type],gCURRENT_EVNT.is_set);
else
printf("\n%d:call profile set for %s isset is %d",armci_me,
gEventName[event_type],gCURRENT_EVNT.is_set);
fflush(stdout);
#endif
if(gCURRENT_EVNT.is_set == 0) { /* set an event */
gCURRENT_EVNT.is_set = 1;
gCURRENT_EVNT.event_type = event_type;
gCURRENT_EVNT.range = range;
gCURRENT_EVNT.start_time = MP_TIMER();
return ARMCI_EVENT_SET;
}
else gCURRENT_EVNT.is_set++; /* event overlap */
return ARMCI_EVENT_NOTSET;
}
int
main(int argc, char **argv) {
Integer heap=9000000, stack=9000000;
int me, nproc;
DoublePrecision time;
MP_INIT(argc,argv);
GA_INIT(argc,argv); /* initialize GA */
nproc = GA_Nnodes();
me = GA_Nodeid();
if(me==0) printf("Using %d processes\n\n",nproc);
if (me==0) printf ("Matrix size is %d X %d\n",N,N);
#ifdef USE_REGULAR
if (me == 0) printf("\nUsing regular data distribution\n\n");
#endif
#ifdef USE_SIMPLE_CYCLIC
if (me == 0) printf("\nUsing simple block-cyclic data distribution\n\n");
#endif
#ifdef USE_SCALAPACK
if (me == 0) printf("\nUsing ScaLAPACK data distribution\n\n");
#endif
#ifdef USE_TILED
if (me == 0) printf("\nUsing tiled data distribution\n\n");
#endif
if(!MA_init((Integer)MT_F_DBL, stack/nproc, heap/nproc))
GA_Error("MA_init failed bytes= %d",stack+heap);
#ifdef PERMUTE
{
int i, *list = (int*)malloc(nproc*sizeof(int));
if(!list)GA_Error("malloc failed",nproc);
for(i=0; i<nproc;i++)list[i]=nproc-1-i;
GA_Register_proclist(list, nproc);
free(list);
}
#endif
if(GA_Uses_fapi())GA_Error("Program runs with C API only",1);
time = MP_TIMER();
do_work();
/* printf("%d: Total Time = %lf\n", me, MP_TIMER()-time);
printf("%d: GEMM Total Time = %lf\n", me, gTime);
*/
if(me==0)printf("\nSuccess\n\n");
GA_Terminate();
MP_FINALIZE();
return 0;
}
void test_aggregate(int dryrun) {
int i, j, rc, bytes, elems[2] = {MAXPROC, MAXELEMS};
double *ddst_put[MAXPROC];
double *ddst_get[MAXPROC];
double *dsrc[MAXPROC];
armci_hdl_t aggr_hdl_put[MAXPROC];
armci_hdl_t aggr_hdl_get[MAXPROC];
armci_hdl_t hdl_put[MAXELEMS];
armci_hdl_t hdl_get[MAXELEMS];
armci_giov_t darr;
void *src_ptr[MAX_REQUESTS], *dst_ptr[MAX_REQUESTS];
int start = 0, end = 0;
double start_time;
create_array((void**)ddst_put, sizeof(double),2, elems);
create_array((void**)ddst_get, sizeof(double),2, elems);
create_array((void**)dsrc, sizeof(double),1, &elems[1]);
for(i=0; i<elems[1]; i++) dsrc[me][i]=i*1.001*(me+1);
for(i=0; i<elems[0]*elems[1]; i++) {
ddst_put[me][i]=0.0;
ddst_get[me][i]=0.0;
}
MP_BARRIER();
/* only proc 0 does the work */
if(me == 0) {
if(!dryrun)printf("Transferring %d doubles (Not an array of %d doubles)\n", MAXELEMS, MAXELEMS);
/* initializing non-blocking handles */
for(i=0; i<elems[1]; i++) ARMCI_INIT_HANDLE(&hdl_put[i]);
for(i=0; i<elems[1]; i++) ARMCI_INIT_HANDLE(&hdl_get[i]);
/* aggregate handles */
for(i=0; i<nproc; i++) ARMCI_INIT_HANDLE(&aggr_hdl_put[i]);
for(i=0; i<nproc; i++) ARMCI_INIT_HANDLE(&aggr_hdl_get[i]);
for(i=0; i<nproc; i++) ARMCI_SET_AGGREGATE_HANDLE(&aggr_hdl_put[i]);
for(i=0; i<nproc; i++) ARMCI_SET_AGGREGATE_HANDLE(&aggr_hdl_get[i]);
bytes = sizeof(double);
/* **************** PUT **************** */
/* register put */
start_time=MP_TIMER();
start = 0; end = elems[1];
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
ARMCI_NbPutValueDouble(dsrc[me][j], &ddst_put[i][me*elems[1]+j], i,
&hdl_put[j]);
}
for(j=start; j<end; j++) ARMCI_Wait(&hdl_put[j]);
}
if(!dryrun)printf("%d: Value Put time = %.2es\n", me, MP_TIMER()-start_time);
/* vector put */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
src_ptr[j] = (void *)&dsrc[me][j];
dst_ptr[j] = (void *)&ddst_put[i][me*elems[1]+j];
}
darr.src_ptr_array = src_ptr;
darr.dst_ptr_array = dst_ptr;
darr.bytes = sizeof(double);
darr.ptr_array_len = elems[1];
if((rc=ARMCI_NbPutV(&darr, 1, i, &hdl_put[i])))
ARMCI_Error("armci_nbputv failed\n",rc);
}
for(i=1; i<nproc; i++) ARMCI_Wait(&hdl_put[i]);
if(!dryrun)printf("%d: Vector Put time = %.2es\n", me, MP_TIMER()-start_time);
/* regular put */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
if((rc=ARMCI_NbPut(&dsrc[me][j], &ddst_put[i][me*elems[1]+j], bytes,
i, &hdl_put[j])))
ARMCI_Error("armci_nbput failed\n",rc);
}
for(j=start; j<end; j++) ARMCI_Wait(&hdl_put[j]);
}
if(!dryrun)printf("%d: Regular Put time = %.2es\n", me, MP_TIMER()-start_time);
/* aggregate put */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
if((rc=ARMCI_NbPut(&dsrc[me][j], &ddst_put[i][me*elems[1]+j], bytes,
i, &aggr_hdl_put[i])))
ARMCI_Error("armci_nbput failed\n",rc);
}
}
for(i=1; i<nproc; i++) ARMCI_Wait(&aggr_hdl_put[i]);
if(!dryrun)printf("%d: Aggregate Put time = %.2es\n\n", me, MP_TIMER()-start_time);
/* **************** GET **************** */
/* vector get */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
src_ptr[j] = (void *)&dsrc[i][j];
dst_ptr[j] = (void *)&ddst_get[me][i*elems[1]+j];
}
darr.src_ptr_array = src_ptr;
darr.dst_ptr_array = dst_ptr;
darr.bytes = sizeof(double);
darr.ptr_array_len = elems[1];
if((rc=ARMCI_NbGetV(&darr, 1, i, &hdl_get[i])))
ARMCI_Error("armci_nbgetv failed\n",rc);
ARMCI_Wait(&hdl_get[i]);
}
if(!dryrun)printf("%d: Vector Get time = %.2es\n", me, MP_TIMER()-start_time);
/* regular get */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
if((rc=ARMCI_NbGet(&dsrc[i][j], &ddst_get[me][i*elems[1]+j], bytes,
i, &hdl_get[j])))
ARMCI_Error("armci_nbget failed\n",rc);
}
for(j=start; j<end; j++) ARMCI_Wait(&hdl_get[j]);
}
if(!dryrun)printf("%d: Regular Get time = %.2es\n", me, MP_TIMER()-start_time);
/* aggregate get */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
ARMCI_NbGet(&dsrc[i][j], &ddst_get[me][i*elems[1]+j], bytes,
i, &aggr_hdl_get[i]);
}
}
for(i=1; i<nproc; i++) ARMCI_Wait(&aggr_hdl_get[i]);
if(!dryrun)printf("%d: Aggregate Get time = %.2es\n", me, MP_TIMER()-start_time);
}
MP_BARRIER();
ARMCI_AllFence();
MP_BARRIER();
/* Verify */
if(!(me==0))
for(j=0; j<elems[1]; j++) {
if( ARMCI_ABS(ddst_put[me][j]-j*1.001) > 0.1) {
ARMCI_Error("aggregate put failed...1", 0);
}
}
MP_BARRIER();
if(!dryrun)if(me==0) printf("\n aggregate put ..O.K.\n"); fflush(stdout);
if(me==0) {
for(i=1; i<nproc; i++) {
for(j=0; j<elems[1]; j++) {
if( ARMCI_ABS(ddst_get[me][i*elems[1]+j]-j*1.001*(i+1)) > 0.1) {
ARMCI_Error("aggregate get failed...1", 0);
}
}
}
}
MP_BARRIER();
if(!dryrun)if(me==0) printf(" aggregate get ..O.K.\n"); fflush(stdout);
ARMCI_AllFence();
MP_BARRIER();
if(!dryrun)if(me==0){printf("O.K.\n"); fflush(stdout);}
destroy_array((void **)ddst_put);
destroy_array((void **)ddst_get);
destroy_array((void **)dsrc);
}
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;
}
void test_io_dbl()
{
int n, ndim = NDIM;
double err, tt0, tt1, mbytes;
int g_a, g_b, d_a;
int i, itmp, j, req, loop;
int glo[MAXDIM],ghi[MAXDIM];
dra_size_t dlo[MAXDIM],dhi[MAXDIM];
dra_size_t ddims[MAXDIM],reqdims[MAXDIM];
dra_size_t m;
int index[MAXDIM], dims[MAXDIM];
int me, nproc, isize;
double *ptr;
double plus, minus;
int ld[MAXDIM], chunk[MAXDIM];
char filename[80];
FILE *fd;
n = SIZE;
m = ((dra_size_t)NFACTOR)*((dra_size_t)SIZE);
loop = 1;
for (i=0; i<ndim; i++) loop *= NFACTOR;
req = -1;
nproc = GA_Nnodes();
me = GA_Nodeid();
if (me == 0) {
printf("Creating temporary global arrays %d",n);
for (i=1; i<ndim; i++) {
printf(" x %d",n);
}
printf("\n");
}
if (me == 0) fflush(stdout);
GA_Sync();
for (i=0; i<ndim; i++) {
dims[i] = n;
chunk[i] = 1;
}
g_a = NGA_Create(MT_DBL, ndim, dims, "a", chunk);
if (!g_a) GA_Error("NGA_Create failed: a", 0);
g_b = NGA_Create(MT_DBL, ndim, dims, "b", chunk);
if (!g_b) GA_Error("NGA_Create failed: b", 0);
if (me == 0) printf("done\n");
if (me == 0) fflush(stdout);
/* initialize g_a, g_b with random values
... use ga_access to avoid allocating local buffers for ga_put */
GA_Sync();
NGA_Distribution(g_a, me, glo, ghi);
NGA_Access(g_a, glo, ghi, &ptr, ld);
isize = 1;
for (i=0; i<ndim; i++) isize *= (ghi[i]-glo[i]+1);
fill_random(ptr, isize);
GA_Sync();
GA_Zero(g_b);
/*.......................................................................*/
if (me == 0) {
printf("Creating Disk array %ld",m);
for (i=1; i<ndim; i++) {
printf(" x %ld",m);
}
printf("\n");
}
if (me == 0) fflush(stdout);
for (i=0; i<ndim; i++) {
ddims[i] = m;
reqdims[i] = (dra_size_t)n;
}
GA_Sync();
strcpy(filename,FNAME);
if (! (fd = fopen(filename, "w"))) {
strcpy(filename,FNAME_ALT);
if (! (fd = fopen(filename, "w"))) {
GA_Error("open failed",0);
}
}
fclose(fd);
if (NDRA_Create(MT_DBL, ndim, ddims, "A", filename, DRA_RW,
reqdims, &d_a) != 0) {
GA_Error("NDRA_Create failed(d_a): ",0);
}
if (me == 0) printf("testing write\n");
fflush(stdout);
tt1 = 0.0;
for (i=0; i<loop; i++) {
itmp=i;
for (j=0; j<ndim; j++) {
index[j] = itmp%NFACTOR;
itmp = (itmp - index[j])/NFACTOR;
}
for (j=0; j<ndim; j++) {
glo[j] = 0;
ghi[j] = SIZE - 1;
dlo[j] = ((dra_size_t)index[j])*((dra_size_t)SIZE);
dhi[j] = (((dra_size_t)index[j])+(dra_size_t)1)
* ((dra_size_t)SIZE) - (dra_size_t)1;
}
tt0 = MP_TIMER();
if (NDRA_Write_section(FALSE, g_a, glo, ghi,
d_a, dlo, dhi, &req) != 0) {
GA_Error("ndra_write_section failed:",0);
}
if (DRA_Wait(req) != 0) {
GA_Error("DRA_Wait failed(d_a): ",req);
}
tt1 += (MP_TIMER() - tt0);
}
GA_Dgop(&tt1,1,"+");
tt1 = tt1/((double)nproc);
mbytes = 1.e-6 * (double)(pow(m,ndim)*sizeof(double));
if (me == 0) {
printf("%11.2f MB time = %11.2f rate = %11.3f MB/s\n",
mbytes,tt1,mbytes/tt1);
}
if (DRA_Close(d_a) != 0) {
GA_Error("DRA_Close failed(d_a): ",d_a);
}
if (me == 0) printf("\n");
if (me == 0) printf("disk array closed\n");
if (me == 0) fflush(stdout);
/*..........................................................*/
if (me == 0) printf("\n");
if (me == 0) printf("opening disk array\n");
if (DRA_Open(filename, DRA_R, &d_a) != 0) {
GA_Error("DRA_Open failed",0);
}
if (me == 0) printf("testing read\n");
/* printf("testing read on proc %d\n",me); */
if (me == 0) fflush(stdout);
tt1 = 0.0;
for (i=0; i<loop; i++) {
itmp=i;
for (j=0; j<ndim; j++) {
index[j] = itmp%NFACTOR;
itmp = (itmp - index[j])/NFACTOR;
}
for (j=0; j<ndim; j++) {
glo[j] = 0;
ghi[j] = SIZE - 1;
dlo[j] = ((dra_size_t)index[j])*((dra_size_t)SIZE);
dhi[j] = (((dra_size_t)index[j])+(dra_size_t)1)
* ((dra_size_t)SIZE) - (dra_size_t)1;
}
tt0 = MP_TIMER();
if (NDRA_Read_section(FALSE, g_b, glo, ghi,
d_a, dlo, dhi, &req) != 0) {
GA_Error("ndra_read_section failed:",0);
}
if (DRA_Wait(req) != 0) {
GA_Error("DRA_Wait failed(d_a): ",req);
}
tt1 += (MP_TIMER() - tt0);
plus = 1.0;
minus = -1.0;
GA_Add(&plus, g_a, &minus, g_b, g_b);
err = GA_Ddot(g_b, g_b);
if (err != 0) {
if (me == 0) {
printf("BTW, we have error = %f on loop value %d\n",
err,i);
}
GA_Error(" bye",0);
}
}
GA_Dgop(&tt1,1,"+");
tt1 = tt1/((double)nproc);
if (me == 0) {
printf("%11.2f MB time = %11.2f rate = %11.3f MB/s\n",
mbytes,tt1,mbytes/tt1);
}
if (DRA_Delete(d_a) != 0) GA_Error("DRA_Delete failed",0);
/*.......................................................................*/
GA_Destroy(g_a);
GA_Destroy(g_b);
}
/*
* test ga_dgemm
* Note: - change nummax for large arrays
* - turn off "dgemm_verify" for large arrays due to memory
* limitations, as dgemm_verify=1 for large arrays produces
* segfault, dumps core,or any crap.
*/
int main(int argc, char **argv)
{
int num_m;
int num_n;
int num_k;
int i;
int ii;
double *h0;
int g_c;
int g_b;
int g_a;
double a;
double t1;
double mf;
double avg_t[ntrans];
double avg_mf[ntrans];
int itime;
int ntimes;
int nums_m[/*howmany*/] = {512,1024};
int nums_n[/*howmany*/] = {512,1024};
int nums_k[/*howmany*/] = {512,1024};
char transa[/*ntrans*/] = "ntnt";
char transb[/*ntrans*/] = "nntt";
char ta;
char tb;
double *tmpa;
double *tmpb;
double *tmpc;
int ndim;
int dims[2];
#ifdef BLOCK_CYCLIC
int block_size[2];
#endif
#if defined(USE_ELEMENTAL)
// initialize Elemental (which will initialize MPI)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &me );
ElMPICommSize( MPI_COMM_WORLD, &nproc );
// instantiate el::global array
ElGlobalArraysConstruct_d( &eldga );
// initialize global arrays
ElGlobalArraysInitialize_d( eldga );
#else
MP_INIT(argc,argv);
if (!MA_init(MT_DBL,1,20000000)) {
GA_Error("failed: ma_init(MT_DBL,1,20000000)",10);
}
GA_INIT(argc,argv);
me = GA_Nodeid();
#endif
h0 = (double*)malloc(sizeof(double) * nummax*nummax);
tmpa = (double*)malloc(sizeof(double) * nummax*nummax);
tmpb = (double*)malloc(sizeof(double) * nummax*nummax);
tmpc = (double*)malloc(sizeof(double) * nummax*nummax);
ii = 0;
for (i=0; i<nummax*nummax; i++) {
ii = ii + 1;
if (ii > nummax) {
ii = 0;
}
h0[i] = ii;
}
/* Compute times assuming 500 mflops and 5 second target time */
/* ntimes = max(3.0d0,5.0d0/(4.0d-9*num**3)); */
ntimes = 5;
for (ii=0; ii<howmany; ii++) {
num_m = nums_m[ii];
num_n = nums_n[ii];
num_k = nums_k[ii];
a = 0.5/(num_m*num_n);
if (num_m > nummax || num_n > nummax || num_k > nummax) {
GA_Error("Insufficient memory: check nummax", 1);
}
#ifndef BLOCK_CYCLIC
ndim = 2;
/*
dims[0] = num_m;
dims[1] = num_n;
*/
dims[1] = num_m;
dims[0] = num_n;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_c", NULL, &g_c );
#else
if (!((g_c = NGA_Create(MT_DBL,ndim,dims,"g_c",NULL)))) {
GA_Error("failed: create g_c",20);
}
#endif
/*
dims[0] = num_k;
dims[1] = num_n;
*/
dims[1] = num_k;
dims[0] = num_n;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_b", NULL, &g_b );
#else
if (!((g_b = NGA_Create(MT_DBL,ndim,dims,"g_b",NULL)))) {
GA_Error("failed: create g_b",30);
}
#endif
/*
dims[0] = num_m;
dims[1] = num_k;
*/
dims[1] = num_m;
dims[0] = num_k;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_a", NULL, &g_a );
#else
if (!((g_a = NGA_Create(MT_DBL,ndim,dims,"g_a",NULL)))) {
GA_Error("failed: create g_a",40);
}
#endif
#else
ndim = 2;
block_size[0] = 128;
block_size[1] = 128;
dims[0] = num_m;
dims[1] = num_n;
g_c = GA_Create_handle();
GA_Set_data(g_c,ndim,dims,MT_DBL);
GA_Set_array_name(g_c,"g_c");
GA_Set_block_cyclic(g_c,block_size);
if (!GA_Allocate(g_c)) {
GA_Error("failed: create g_c",40);
}
dims[0] = num_k;
dims[1] = num_n;
g_b = GA_Create_handle();
GA_Set_data(g_b,ndim,dims,MT_DBL);
GA_Set_array_name(g_b,"g_b");
GA_Set_block_cyclic(g_b,block_size);
if (!ga_allocate(g_b)) {
GA_Error("failed: create g_b",40);
}
dims[0] = num_m;
dims[1] = num_k;
g_a = GA_Create_handle();
GA_Set_data(g_a,ndim,dims,MT_DBL);
GA_Set_array_name(g_a,"g_a");
GA_Set_block_cyclic(g_a,block_size);
if (!ga_allocate(g_a)) {
GA_Error('failed: create g_a',40);
}
#endif
/* Initialize matrices A and B */
if (me == 0) {
load_ga(g_a, h0, num_m, num_k);
load_ga(g_b, h0, num_k, num_n);
}
#if defined(USE_ELEMENTAL)
double zero = 0.0;
ElGlobalArraysFill_d( eldga, g_c, &zero );
ElGlobalArraysSync_d( eldga );
#else
GA_Zero(g_c);
GA_Sync();
#endif
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("\nMatrix Multiplication on C = A[%ld,%ld]xB[%ld,%ld]\n",
(long)num_m, (long)num_k, (long)num_k, (long)num_n);
fflush(stdout);
}
for (i=0; i<ntrans; i++) {
avg_t[i] = 0.0;
avg_mf[i] = 0.0;
}
for (itime=0; itime<ntimes; itime++) {
for (i=0; i<ntrans; i++) {
#if defined(USE_ELEMENTAL)
ElGlobalArraysSync_d( eldga );
#else
GA_Sync();
#endif
ta = transa[i];
tb = transb[i];
t1 = MP_TIMER();
#if defined(USE_ELEMENTAL)
ElGlobalArraysDgemm_d( eldga, ta, tb, num_m, num_n, num_k, 1.0, g_a, g_b, 0.0, g_c );
#else
GA_Dgemm(ta,tb,num_m,num_n,num_k,1.0, g_a, g_b, 0.0, g_c);
#endif
t1 = MP_TIMER() - t1;
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
#if defined(USE_ELEMENTAL)
mf = 2e0*num_m*num_n*num_k/t1*1e-6/nproc;
#else
mf = 2e0*num_m*num_n*num_k/t1*1e-6/GA_Nnodes();
#endif
avg_t[i] = avg_t[i]+t1;
avg_mf[i] = avg_mf[i] + mf;
printf("%15s%2d: %12.4f seconds %12.1f mflops/proc %c %c\n",
"Run#", itime, t1, mf, ta, tb);
fflush(stdout);
if (dgemm_verify && itime == 0) {
/* recall the C API swaps the matrix order */
/* we swap it here for the Fortran-based verify */
verify_ga_dgemm(tb, ta, num_n, num_m, num_k, 1.0,
g_b, g_a, 0.0, g_c, tmpb, tmpa, tmpc);
}
}
}
}
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("\n");
for (i=0; i<ntrans; i++) {
printf("%17s: %12.4f seconds %12.1f mflops/proc %c %c\n",
"Average", avg_t[i]/ntimes, avg_mf[i]/ntimes,
transa[i], transb[i]);
}
if(dgemm_verify) {
printf("All GA_Dgemms are verified...O.K.\n");
}
fflush(stdout);
}
/*
GA_Print(g_a);
GA_Print(g_b);
GA_Print(g_c);
*/
#if defined(USE_ELEMENTAL)
ElGlobalArraysDestroy_d( eldga, g_a );
ElGlobalArraysDestroy_d( eldga, g_b );
ElGlobalArraysDestroy_d( eldga, g_c );
#else
GA_Destroy(g_c);
GA_Destroy(g_b);
GA_Destroy(g_a);
#endif
}
/* ???
format(a15, i2, ': ', e12.4, ' seconds ',f12.1,
. ' mflops/proc ', 3a2)
*/
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("All tests successful\n");
}
free(h0);
free(tmpa);
free(tmpb);
free(tmpc);
#if defined(USE_ELEMENTAL)
// call el::global arrays destructor
ElGlobalArraysTerminate_d( eldga );
ElGlobalArraysDestruct_d( eldga );
ElFinalize();
#else
GA_Terminate();
MP_FINALIZE();
#endif
return 0;
}
/*
* Verify for correctness. Process 0 computes BLAS dgemm
* locally. For larger arrays, disbale this test as memory
* might not be sufficient
*/
void verify_ga_dgemm(char xt1, char xt2, int num_m, int num_n, int num_k,
double alpha, int g_a, int g_b, double beta, int g_c,
double *tmpa, double *tmpb, double *tmpc)
{
int i,j,type,ndim,dims[2],lo[2],hi[2];
double abs_value;
for (i=0; i<num_n; i++) {
for (j=0; j<num_m; j++) {
tmpc[j+i*num_m] = -1.0;
tmpa[j+i*num_m] = -2.0;
}
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_a, lo, hi, tmpa, &dims[1] );
#else
NGA_Get(g_a, lo, hi, tmpa, &dims[1]);
#endif
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_b, lo, hi, tmpb, &dims[1] );
#else
NGA_Get(g_b, lo, hi, tmpb, &dims[1]);
#endif
/* compute dgemm sequentially */
#if defined(USE_ELEMENTAL)
cblas_dgemm ( CblasRowMajor, ( xt1 == 'n'? CblasNoTrans: CblasTrans ),
( xt2 == 'n'? CblasNoTrans: CblasTrans ),
num_m /* M */, num_n /* N */, num_k /* K */,
alpha, tmpa, num_m, /* lda */
tmpb, num_k, /* ldb */ beta,
tmpc, num_m /* ldc */);
#else
xb_dgemm(&xt1, &xt2, &num_m, &num_n, &num_k,
&alpha, tmpa, &num_m,
tmpb, &num_k, &beta,
tmpc, &num_m);
#endif
/* after computing c locally, verify it with the values in g_c */
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_c, lo, hi, tmpa, &dims[1] );
#else
NGA_Get(g_c, lo, hi, tmpa, &dims[1]);
#endif
for (i=0; i<num_n; i++) {
for (j=0; j<num_m; j++) {
abs_value = fabs(tmpc[j+i*num_m]-tmpa[j+i*num_m]);
if(abs_value > 1.0 || abs_value < -1.0) {
printf("Values are = %f %f\n",
tmpc[j+i*num_m], tmpa[j+i*num_m]);
printf("Values are = %f %f\n",
fabs(tmpc[j+i*num_m]-tmpa[j*i*num_m]), abs_value);
fflush(stdout);
GA_Error("verify ga_dgemm failed", 1);
}
}
}
}
/**
* called by process '0' (or your master process )
*/
void load_ga(int handle, double *f, int dim1, int dim2)
{
int lo[2], hi[2];
if (dim1 < 0 || dim2 < 0) {
return;
}
lo[0] = 0;
lo[1] = 0;
hi[0] = dim1-1;
hi[1] = dim2-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysPut_d( eldga, handle, lo, hi, f, &dim1 );
#else
NGA_Put(handle, lo, hi, f, &dim1);
#endif
}
void test_perf_nb(int dry_run) {
int i, j, loop, rc, bytes, elems[2] = {MAXPROC, MAXELEMS};
int stride, k=0, ntimes;
double stime, t1, t2, t3, t4, t5, t6, t7, t8, t9;
double *dsrc[MAXPROC], scale=1.0;
armci_hdl_t hdl_get, hdl_put, hdl_acc;
create_array((void**)ddst, sizeof(double),2, elems);
create_array((void**)dsrc, sizeof(double),1, &elems[1]);
if(!dry_run)if(me == 0) {
printf("\n\t\t\tRemote 1-D Array Section\n");
printf("section get nbget wait put nbput ");
printf(" wait acc nbacc wait\n");
printf("------- -------- -------- -------- -------- --------");
printf(" -------- -------- -------- --------\n");
fflush(stdout);
}
for(loop=1; loop<=MAXELEMS; loop*=2, k++) {
elems[1] = loop;
ntimes = (int)sqrt((double)(MAXELEMS/elems[1]));
if(ntimes <1) ntimes=1;
/* -------------------------- SETUP --------------------------- */
/*initializing non-blocking handles,time,src & dst buffers*/
ARMCI_INIT_HANDLE(&hdl_put);
ARMCI_INIT_HANDLE(&hdl_get);
ARMCI_INIT_HANDLE(&hdl_acc);
t1 = t2 = t3 = t4 = t5 = t6 = t7 = t8 = t9 = 0.0;
for(i=0; i<elems[1]; i++) dsrc[me][i]=i*1.001*(me+1);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
/* bytes transfered */
bytes = sizeof(double)*elems[1];
MP_BARRIER();
/* -------------------------- PUT/GET -------------------------- */
if(me == 0) {
for(i=1; i<nproc; i++) {
stime=MP_TIMER();
for(j=0; j<ntimes; j++)
if((rc=ARMCI_Put(&dsrc[me][0], &ddst[i][me*elems[1]], bytes,i)))
ARMCI_Error("armci_nbput failed\n",rc);
t1 += MP_TIMER()-stime;
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(PUT, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
if(me == 0) {
for(i=1; i<nproc; i++) {
stime=MP_TIMER();
for(j=0; j<ntimes; j++)
if((rc=ARMCI_Get(&dsrc[i][0], &ddst[me][i*elems[1]], bytes,i)))
ARMCI_Error("armci_nbget failed\n",rc);
t4 += MP_TIMER()-stime;
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(GET, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
/* ------------------------ nb PUT/GET ------------------------- */
if(me == 0) {
for(i=1; i<nproc; i++) {
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_NbPut(&dsrc[me][0], &ddst[i][me*elems[1]], bytes,
i, &hdl_put)))
ARMCI_Error("armci_nbput failed\n",rc);
t2 += MP_TIMER()-stime; stime=MP_TIMER();
ARMCI_Wait(&hdl_put);
t3 += MP_TIMER()-stime;
}
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(PUT, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
if(me == 0) {
for(i=1; i<nproc; i++) {
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_NbGet(&dsrc[i][0], &ddst[me][i*elems[1]], bytes,
i, &hdl_get)))
ARMCI_Error("armci_nbget failed\n",rc);
t5 += MP_TIMER()-stime; stime=MP_TIMER();
ARMCI_Wait(&hdl_get);
t6 += MP_TIMER()-stime;
}
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(GET, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
/* ------------------------ Accumulate ------------------------- */
for(i=0; i<elems[1]; i++) dsrc[me][i]=1.0; MP_BARRIER();
stride = elems[1]*sizeof(double); scale = 1.0;
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_AccS(ARMCI_ACC_DBL, &scale, &dsrc[me][0], &stride,
&ddst[0][0], &stride, &bytes, 0, 0)))
ARMCI_Error("armci_acc failed\n",rc);
t7 += MP_TIMER()-stime;
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(ACC, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
}
#if 1
/* See the note below why this part is disabled */
/* ---------------------- nb-Accumulate ------------------------ */
for(i=0; i<elems[1]; i++) dsrc[me][i]=1.0; MP_BARRIER();
stride = elems[1]*sizeof(double); scale = 1.0;
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_NbAccS(ARMCI_ACC_DBL, &scale, &dsrc[me][0], &stride,
&ddst[0][0], &stride, &bytes, 0, 0, &hdl_acc)))
ARMCI_Error("armci_nbacc failed\n",rc);
t8 += MP_TIMER()-stime; stime=MP_TIMER();
ARMCI_Wait(&hdl_acc);
t9 += MP_TIMER()-stime;
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(ACC, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
}
#endif
/* print timings */
if(!dry_run) if(me==0) printf("%d\t %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e\n",
bytes, t4/ntimes, t5/ntimes, t6/ntimes, t1/ntimes,
t2/ntimes, t3/ntimes, t7/ntimes, t8/ntimes, t9/ntimes);
}
ARMCI_AllFence();
MP_BARRIER();
if(!dry_run)if(me==0){printf("O.K.\n"); fflush(stdout);}
destroy_array((void **)ddst);
destroy_array((void **)dsrc);
}
