void EpetraVector::applyOp(
const RTOpPack::RTOpT<double> &op,
const int num_vecs,
const VectorBase<double>*const vecs[],
const int num_targ_vecs,
VectorBase<double>*const targ_vecs[],
RTOpPack::ReductTarget *reduct_obj,
const OrdType first_ele_offset_in,
const OrdType sub_dim_in,
const OrdType global_offset_in
) const
{
#ifdef THYRA_SPMD_VECTOR_BASE_DUMP
Teuchos::RCP<Teuchos::FancyOStream>
out = Teuchos::VerboseObjectBase::getDefaultOStream();
Teuchos::OSTab tab(out);
if(show_dump) {
*out << "\nEntering SpmdVectorBase<double>::applyOp(...) ...\n";
*out
<< "\nop = " << typeName(op)
<< "\nnum_vecs = " << num_vecs
<< "\nnum_targ_vecs = " << num_targ_vecs
<< "\nreduct_obj = " << reduct_obj
<< "\nfirst_ele_offset_in = " << first_ele_offset_in
<< "\nsub_dim_in = " << sub_dim_in
<< "\nglobal_offset_in = " << global_offset_in
<< "\n"
;
}
#endif // THYRA_SPMD_VECTOR_BASE_DUMP
using Teuchos::dyn_cast;
using Teuchos::Workspace;
Teuchos::WorkspaceStore* wss = Teuchos::get_default_workspace_store().get();
#ifdef TEUCHOS_DEBUG
// ToDo: Validate input!
TEST_FOR_EXCEPTION(
in_applyOpImpl_, std::invalid_argument
,"SpmdVectorBase<>::applyOp(...): Error, this method is being entered recursively which is a "
"clear sign that one of the methods acquireDetachedView(...), releaseDetachedView(...) or commitDetachedView(...) "
"was not implemented properly!"
);
Thyra::apply_op_validate_input(
"SpmdVectorBase<>::applyOp(...)",*space()
,op,num_vecs,vecs,num_targ_vecs,targ_vecs,reduct_obj,first_ele_offset_in,sub_dim_in,global_offset_in
);
#endif
const Teuchos::RCP<const Teuchos::Comm<OrdType> >& comm
= epetraVecSpace_->getComm();
const SerialComm<OrdType>* serialComm =
dynamic_cast<const SerialComm<OrdType>* >(comm.get());
// Flag that we are in applyOp()
in_applyOpImpl_ = true;
// First see if this is a locally replicated vector in which case
// we treat this as a local operation only.
const bool locallyReplicated = ( serialComm != 0 && localSubDim_ == globalDim_ );
// Get the overlap in the current process with the input logical sub-vector
// from (first_ele_offset_in,sub_dim_in,global_offset_in)
OrdType overlap_first_local_ele_off = 0;
OrdType overlap_local_sub_dim = 0;
OrdType overlap_global_off = 0;
if(localSubDim_) {
RTOp_parallel_calc_overlap(
globalDim_, localSubDim_, localOffset_, first_ele_offset_in, sub_dim_in, global_offset_in
,&overlap_first_local_ele_off, &overlap_local_sub_dim, &overlap_global_off
);
}
const Range1D local_rng = (
overlap_first_local_ele_off>=0
? Range1D( localOffset_ + overlap_first_local_ele_off, localOffset_ + overlap_first_local_ele_off + overlap_local_sub_dim - 1 )
: Range1D::Invalid
);
#ifdef THYRA_SPMD_VECTOR_BASE_DUMP
if(show_dump) {
*out
<< "\noverlap_first_local_ele_off = " << overlap_first_local_ele_off
<< "\noverlap_local_sub_dim = " << overlap_local_sub_dim
<< "\noverlap_global_off = " << overlap_global_off
<< "\nlocal_rng = ["<<local_rng.lbound()<<","<<local_rng.ubound()<<"]"
<< "\n"
;
}
#endif // THYRA_SPMD_VECTOR_BASE_DUMP
// Create sub-vector views of all of the *participating* local data
Workspace<RTOpPack::ConstSubVectorView<double> > sub_vecs(wss,num_vecs);
Workspace<RTOpPack::SubVectorView<double> > sub_targ_vecs(wss,num_targ_vecs);
if( overlap_first_local_ele_off >= 0 ) {
{for(int k = 0; k < num_vecs; ++k ) {
vecs[k]->acquireDetachedView( local_rng, &sub_vecs[k] );
sub_vecs[k].setGlobalOffset( overlap_global_off );
}}
{for(int k = 0; k < num_targ_vecs; ++k ) {
targ_vecs[k]->acquireDetachedView( local_rng, &sub_targ_vecs[k] );
sub_targ_vecs[k].setGlobalOffset( overlap_global_off );
}}
}
// Apply the RTOp operator object (all processors must participate)
const bool validSubVecs = ( localSubDim_ > 0 && overlap_first_local_ele_off >= 0 );
RTOpPack::SPMD_apply_op(
locallyReplicated ? NULL : comm.get() // comm
,op // op
,num_vecs // num_vecs
,num_vecs && validSubVecs ? &sub_vecs[0] : NULL // sub_vecs
,num_targ_vecs // num_targ_vecs
,num_targ_vecs && validSubVecs ? &sub_targ_vecs[0] : NULL // targ_sub_vecs
,reduct_obj // reduct_obj
);
// Free and commit the local data
if( overlap_first_local_ele_off >= 0 ) {
{for(int k = 0; k < num_vecs; ++k ) {
sub_vecs[k].setGlobalOffset(local_rng.lbound());
vecs[k]->releaseDetachedView( &sub_vecs[k] );
}}
{for(int k = 0; k < num_targ_vecs; ++k ) {
sub_targ_vecs[k].setGlobalOffset(local_rng.lbound());
targ_vecs[k]->commitDetachedView( &sub_targ_vecs[k] );
}}
}
// Flag that we are leaving applyOp()
in_applyOpImpl_ = false;
#ifdef THYRA_SPMD_VECTOR_BASE_DUMP
if(show_dump) {
*out << "\nLeaving SpmdVectorBase<double>::applyOp(...) ...\n";
}
#endif // THYRA_SPMD_VECTOR_BASE_DUMP
}
