inline int DDI_ARMCI_Acc_proc(DDI_Patch *patch, void *scale, void *buf, int proc) {
int handle = patch->handle;
int nops = 1;
DDA_ARMCI_Index *armci_index = gv(dda_armci_index)[handle];
int trows,tcols,nrows,ncols;
size_t offset;
char *dst,*src = (char*)buf;
int src_stride_arr[2],dst_stride_arr[2],count[2];
int stride_levels = 1;
int armci_proc;
trows = gv(dda_index)[handle].nrows;
tcols = gv(pcmap)[handle][proc+1] - gv(pcmap)[handle][proc];
nrows = patch->ihi - patch->ilo + 1;
ncols = patch->jhi - patch->jlo + 1;
offset = (patch->jlo - gv(pcmap)[handle][proc])*trows + patch->ilo;
offset *= sizeof(double);
DDI_ARMCI_Acquire(armci_index,handle,proc,DDI_WRITE_ACCESS,(void**)&dst,&armci_proc);
dst += offset;
if (nrows == trows) {
src_stride_arr[0] = sizeof(double)*nrows;
dst_stride_arr[0] = sizeof(double)*trows;
count[0] = patch->size;
stride_levels = 0;
#if defined DDI_ARMCI_IMPLICIT_NBACC
// Apparantely ARMCI_Acc is not always present
//ARMCI_Acc(ARMCI_ACC_DBL, 1.0, (void*)src, (void*)dst, subp[i].size, armci_proc, NULL);
ARMCI_NbAccS(ARMCI_ACC_DBL, scale,
(void*)src, src_stride_arr,
(void*)dst, dst_stride_arr,
count, stride_levels, armci_proc, NULL);
#else
// Apparantely ARMCI_Acc is not always present
//ARMCI_Acc(ARMCI_ACC_DBL, 1.0, (void*)src, (void*)dst, subp[i].size, armci_proc);
ARMCI_AccS(ARMCI_ACC_DBL, scale,
(void*)src, src_stride_arr,
(void*)dst, dst_stride_arr,
count, stride_levels, armci_proc);
#endif
}
else {
// i dimensions
src_stride_arr[0] = sizeof(double)*nrows;
dst_stride_arr[0] = sizeof(double)*trows;
// j dimensions
src_stride_arr[1] = src_stride_arr[0]*ncols;
dst_stride_arr[1] = dst_stride_arr[0]*tcols;
// block size count, first dimension must be in bytes
count[0] = sizeof(double)*nrows;
count[1] = ncols;
stride_levels = 1;
#if defined DDI_ARMCI_IMPLICIT_NBACC
ARMCI_NbAccS(ARMCI_ACC_DBL, scale,
(void*)src, src_stride_arr,
(void*)dst, dst_stride_arr,
count, stride_levels, armci_proc, NULL);
#else
ARMCI_AccS(ARMCI_ACC_DBL, scale,
(void*)src, src_stride_arr,
(void*)dst, dst_stride_arr,
count, stride_levels, armci_proc);
#endif
}
DDI_ARMCI_Release(armci_index,handle,proc,DDI_WRITE_ACCESS);
return nops;
}
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);
}