void DDI_ARR_select_local(DDI_Patch *dAPatch, DDI_ARR_Element *element) {
DDA_Index *Index = gv(dda_index);
int dA = dAPatch->handle;
double *dALocal;
int op = dAPatch->oper;
int dALdaLocal, dAiLocal, dAjLocal, dAmLocal, dAnLocal;
/* A segment dimensions */
dALdaLocal = Index[dA].ihi - Index[dA].ilo + 1;
dAiLocal = dAPatch->ilo - Index[dA].ilo;
dAjLocal = dAPatch->jlo - Index[dA].jlo;
dAmLocal = dAPatch->ihi - dAPatch->ilo + 1;
dAnLocal = dAPatch->jhi - dAPatch->jlo + 1;
# if defined USE_SYSV
if(USING_DATA_SERVERS()) DDI_Fence_check(dA);
# endif
DDI_Acquire(Index, dA, DDI_READ_ACCESS, (void**)&dALocal);
switch(op) {
case DDI_ARR_MIN:
mmin(dALocal, dALdaLocal, dAiLocal, dAjLocal, dAmLocal, dAnLocal, &(element->alpha), element->index);
break;
case DDI_ARR_MAX:
mmax(dALocal, dALdaLocal, dAiLocal, dAjLocal, dAmLocal, dAnLocal, &(element->alpha), element->index);
break;
} /* switch */
/* adjust element indices to global values */
element->index[0] += Index[dA].ilo;
element->index[1] += Index[dA].jlo;
DDI_Release(Index, dA, DDI_READ_ACCESS);
}
/* -------------------------------------------------------------- *\
DDI_GetAcc_local(patch,buff)
=========================
[IN] patch - structure containing ilo, ihi, jlo, jhi, etc.
[IN] buff - Data segment to be operated on.
GetAccumulates the subpatch specified by patch and stored in buff
into the share-memory segment(s) of the local node.
\* -------------------------------------------------------------- */
void DDI_GetAcc_local(const DDI_Patch* patch,void *buff) {
/* --------------- *\
Local Variables
\* --------------- */
DDA_Index *Index = gv(dda_index);
int i,j,nrows,ncols,start_row,start_col;
size_t dda_offset,size;
double tmp,*dda,*dloc = (double *) buff;
int handle = patch->handle;
int ilo = patch->ilo;
int ihi = patch->ihi;
int jlo = patch->jlo;
int jhi = patch->jhi;
int trows = Index[handle].nrows;
# if FULL_SMP
int icpu,smpme,smpnp;
DDI_SMP_NProc(&smpnp,&smpme);
# endif
MAX_DEBUG((stdout,"%s: Entering DDI_GetAcc_local.\n",DDI_Id()))
/* ------------------------------------------------------------------ *\
For FULL SMP implementations, loop on the number of SMP processors
\* ------------------------------------------------------------------ */
# if FULL_SMP
for(icpu=0; icpu<smpnp; icpu++) {
Index = gv(smp_index)[icpu];
jlo = Index[handle].jlo;
jhi = Index[handle].jhi;
if(jlo > patch->jhi || jhi < patch->jlo) continue;
if(patch->jlo > jlo) jlo = patch->jlo;
if(jhi > patch->jhi) jhi = patch->jhi;
# endif
nrows = ihi - ilo + 1;
ncols = jhi - jlo + 1;
size = nrows*ncols;
start_row = ilo - Index[handle].ilo;
start_col = jlo - Index[handle].jlo;
/* ---------------------------------------------------------- *\
If the patch and the DD array have the same row dimensions
\* ---------------------------------------------------------- */
if(nrows == trows) {
dda_offset = start_col*nrows;
DDI_Acquire(Index,handle,DDI_WRITE_ACCESS,(void **) &dda);
dda += dda_offset;
for(i=0; i<size; i++) {
tmp = dda[i];
dda[i] += dloc[i];
dloc[i] = tmp;
}
DDI_Release(Index,handle,DDI_WRITE_ACCESS);
dloc += size;
} else {
/* ----------------------------------------------- *\
Otherwise, pack the local patch into the buffer
\* ----------------------------------------------- */
DDI_Acquire(Index,handle,DDI_WRITE_ACCESS,(void **) &dda);
dda_offset = start_col*trows;
dda += dda_offset;
dda += start_row;
size = nrows*sizeof(double);
for(i=0; i<ncols; i++) {
for(j=0; j<nrows; j++) {
tmp = dda[j];
dda[j] += dloc[j];
dloc[j] = tmp;
}
dloc += nrows;
dda += trows;
}
DDI_Release(Index,handle,DDI_WRITE_ACCESS);
}
# if FULL_SMP
} /* end for-loop on local cpus */
# endif
/* --------------------- *\
Shared-memory counter
\* --------------------- */
# if defined DDI_COUNTERS
gv(acc_profile).ncalls_shmem++;
gv(acc_profile).nbytes_shmem += patch->size;
# endif
MAX_DEBUG((stdout,"%s: Leaving DDI_GetAcc_local.\n",DDI_Id()))
}
