static PetscErrorCode VecSetOption_MPI(Vec V,VecOption op,PetscBool flag)
{
Vec_MPI *v = (Vec_MPI*)V->data;
PetscErrorCode ierr;
PetscFunctionBegin;
switch (op) {
case VEC_IGNORE_OFF_PROC_ENTRIES: V->stash.donotstash = flag;
break;
case VEC_IGNORE_NEGATIVE_INDICES: V->stash.ignorenegidx = flag;
break;
case VEC_SUBSET_OFF_PROC_ENTRIES:
v->assembly_subset = flag; /* See the same logic in MatAssembly wrt MAT_SUBSET_OFF_PROC_ENTRIES */
if (!v->assembly_subset) { /* User indicates "do not reuse the communication pattern" */
ierr = VecAssemblyReset_MPI(V);CHKERRQ(ierr); /* Reset existing pattern to free memory */
v->first_assembly_done = PETSC_FALSE; /* Mark the first assembly is not done */
}
break;
}
PetscFunctionReturn(0);
}
static PetscErrorCode VecAssemblyEnd_MPI_BTS(Vec X)
{
Vec_MPI *x = (Vec_MPI*)X->data;
PetscInt bs = X->map->bs;
PetscMPIInt npending,*some_indices,r;
MPI_Status *some_statuses;
PetscScalar *xarray;
PetscErrorCode ierr;
VecAssemblyFrame *frame;
PetscFunctionBegin;
if (X->stash.donotstash) {
X->stash.insertmode = NOT_SET_VALUES;
X->bstash.insertmode = NOT_SET_VALUES;
PetscFunctionReturn(0);
}
ierr = VecGetArray(X,&xarray);CHKERRQ(ierr);
ierr = PetscSegBufferExtractInPlace(x->segrecvframe,&frame);CHKERRQ(ierr);
ierr = PetscMalloc2(4*x->nrecvranks,&some_indices,x->use_status?4*x->nrecvranks:0,&some_statuses);CHKERRQ(ierr);
for (r=0,npending=0; r<x->nrecvranks; r++) npending += frame[r].pendings + frame[r].pendingb;
while (npending>0) {
PetscMPIInt ndone,ii;
/* Filling MPI_Status fields requires some resources from the MPI library. We skip it on the first assembly, or
* when VEC_SUBSET_OFF_PROC_ENTRIES has not been set, because we could exchange exact sizes in the initial
* rendezvous. When the rendezvous is elided, however, we use MPI_Status to get actual message lengths, so that
* subsequent assembly can set a proper subset of the values. */
ierr = MPI_Waitsome(4*x->nrecvranks,x->recvreqs,&ndone,some_indices,x->use_status?some_statuses:MPI_STATUSES_IGNORE);CHKERRQ(ierr);
for (ii=0; ii<ndone; ii++) {
PetscInt i = some_indices[ii]/4,j,k;
InsertMode imode = (InsertMode)x->recvhdr[i].insertmode;
PetscInt *recvint;
PetscScalar *recvscalar;
PetscBool intmsg = (PetscBool)(some_indices[ii]%2 == 0);
PetscBool blockmsg = (PetscBool)((some_indices[ii]%4)/2 == 1);
npending--;
if (!blockmsg) { /* Scalar stash */
PetscMPIInt count;
if (--frame[i].pendings > 0) continue;
if (x->use_status) {
ierr = MPI_Get_count(&some_statuses[ii],intmsg ? MPIU_INT : MPIU_SCALAR,&count);CHKERRQ(ierr);
} else count = x->recvhdr[i].count;
for (j=0,recvint=frame[i].ints,recvscalar=frame[i].scalars; j<count; j++,recvint++) {
PetscInt loc = *recvint - X->map->rstart;
if (*recvint < X->map->rstart || X->map->rend <= *recvint) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Received vector entry %D out of local range [%D,%D)]",*recvint,X->map->rstart,X->map->rend);
switch (imode) {
case ADD_VALUES:
xarray[loc] += *recvscalar++;
break;
case INSERT_VALUES:
xarray[loc] = *recvscalar++;
break;
default: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Insert mode not supported 0x%x",imode);
}
}
} else { /* Block stash */
PetscMPIInt count;
if (--frame[i].pendingb > 0) continue;
if (x->use_status) {
ierr = MPI_Get_count(&some_statuses[ii],intmsg ? MPIU_INT : MPIU_SCALAR,&count);CHKERRQ(ierr);
if (!intmsg) count /= bs; /* Convert from number of scalars to number of blocks */
} else count = x->recvhdr[i].bcount;
for (j=0,recvint=frame[i].intb,recvscalar=frame[i].scalarb; j<count; j++,recvint++) {
PetscInt loc = (*recvint)*bs - X->map->rstart;
switch (imode) {
case ADD_VALUES:
for (k=loc; k<loc+bs; k++) xarray[k] += *recvscalar++;
break;
case INSERT_VALUES:
for (k=loc; k<loc+bs; k++) xarray[k] = *recvscalar++;
break;
default: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Insert mode not supported 0x%x",imode);
}
}
}
}
}
ierr = VecRestoreArray(X,&xarray);CHKERRQ(ierr);
ierr = MPI_Waitall(4*x->nsendranks,x->sendreqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
ierr = PetscFree2(some_indices,some_statuses);CHKERRQ(ierr);
if (x->assembly_subset) {
void *dummy; /* reset segbuffers */
ierr = PetscSegBufferExtractInPlace(x->segrecvint,&dummy);CHKERRQ(ierr);
ierr = PetscSegBufferExtractInPlace(x->segrecvscalar,&dummy);CHKERRQ(ierr);
} else {
ierr = VecAssemblyReset_MPI(X);CHKERRQ(ierr);
}
X->stash.insertmode = NOT_SET_VALUES;
X->bstash.insertmode = NOT_SET_VALUES;
ierr = VecStashScatterEnd_Private(&X->stash);CHKERRQ(ierr);
ierr = VecStashScatterEnd_Private(&X->bstash);CHKERRQ(ierr);
PetscFunctionReturn(0);
}