// Compute Y := alpha Op X + beta Y.
//
// We ignore the cases alpha != 1 and beta != 0 for simplicity.
void
apply (const MV& X,
MV& Y,
Teuchos::ETransp mode = Teuchos::NO_TRANS,
scalar_type alpha = Teuchos::ScalarTraits<scalar_type>::one (),
scalar_type beta = Teuchos::ScalarTraits<scalar_type>::zero ()) const
{
using Teuchos::RCP;
using Teuchos::rcp;
using std::cout;
using std::endl;
typedef Teuchos::ScalarTraits<scalar_type> STS;
RCP<const Teuchos::Comm<int> > comm = opMap_->getComm ();
const int myRank = comm->getRank ();
const int numProcs = comm->getSize ();
if (myRank == 0) {
cout << "MyOp::apply" << endl;
}
// We're writing the Operator subclass, so we are responsible for
// error handling. You can decide how much error checking you
// want to do. Just remember that checking things like Map
// sameness or compatibility are expensive.
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors () != Y.getNumVectors (), std::invalid_argument,
"X and Y do not have the same numbers of vectors (columns).");
// Let's make sure alpha is 1 and beta is 0...
// This will throw an exception if that is not the case.
TEUCHOS_TEST_FOR_EXCEPTION(
alpha != STS::one() || beta != STS::zero(), std::logic_error,
"MyOp::apply was given alpha != 1 or beta != 0. "
"These cases are not implemented.");
// Get the number of vectors (columns) in X (and Y).
const size_t numVecs = X.getNumVectors ();
// Make a temporary multivector for holding the redistributed
// data. You could also create this in the constructor and reuse
// it across different apply() calls, but you would need to be
// careful to reallocate if it has a different number of vectors
// than X. The number of vectors in X can vary across different
// apply() calls.
RCP<MV> redistData = rcp (new MV (redistMap_, numVecs));
// Redistribute the data.
// This will do all the necessary communication for you.
// All processes now own enough data to do the matvec.
redistData->doImport (X, *importer_, Tpetra::INSERT);
// Get the number of local rows in X, on the calling process.
const local_ordinal_type nlocRows =
static_cast<local_ordinal_type> (X.getLocalLength ());
// Perform the matvec with the data we now locally own.
//
// For each column...
for (size_t c = 0; c < numVecs; ++c) {
// Get a view of the desired column
Teuchos::ArrayRCP<scalar_type> colView = redistData->getDataNonConst (c);
local_ordinal_type offset;
// Y[0,c] = -colView[0] + 2*colView[1] - colView[2] (using local indices)
if (myRank > 0) {
Y.replaceLocalValue (0, c, -colView[0] + 2*colView[1] - colView[2]);
offset = 0;
}
// Y[0,c] = 2*colView[1] - colView[2] (using local indices)
else {
Y.replaceLocalValue (0, c, 2*colView[0] - colView[1]);
offset = 1;
}
// Y[r,c] = -colView[r-offset] + 2*colView[r+1-offset] - colView[r+2-offset]
for (local_ordinal_type r = 1; r < nlocRows - 1; ++r) {
const scalar_type newVal =
-colView[r-offset] + 2*colView[r+1-offset] - colView[r+2-offset];
Y.replaceLocalValue (r, c, newVal);
}
// Y[nlocRows-1,c] = -colView[nlocRows-1-offset] + 2*colView[nlocRows-offset]
// - colView[nlocRows+1-offset]
if (myRank < numProcs - 1) {
const scalar_type newVal =
-colView[nlocRows-1-offset] + 2*colView[nlocRows-offset]
- colView[nlocRows+1-offset];
Y.replaceLocalValue (nlocRows-1, c, newVal);
}
// Y[nlocRows-1,c] = -colView[nlocRows-1-offset] + 2*colView[nlocRows-offset]
else {
const scalar_type newVal =
-colView[nlocRows-1-offset] + 2*colView[nlocRows-offset];
Y.replaceLocalValue (nlocRows-1, c, newVal);
}
}
}
//
// Computes Y = alpha Op X + beta Y
// TraceMin will never use alpha ~= 1 or beta ~= 0,
// so we have ignored those options for simplicity.
//
void apply(const MV& X, MV& Y, Teuchos::ETransp mode=Teuchos::NO_TRANS, Scalar alpha=SCT::one(), Scalar beta=SCT::zero()) const
{
//
// Let's make sure alpha is 1 and beta is 0...
// This will throw an exception if that is not the case.
//
TEUCHOS_TEST_FOR_EXCEPTION(alpha != SCT::one() || beta != SCT::zero(),std::invalid_argument,
"MyOp::apply was given alpha != 1 or beta != 0. That's not supposed to happen.");
//
// Get the number of local rows
//
int nlocRows = X.getLocalLength();
//
// Get the number of vectors
//
int numVecs = X.getNumVectors();
//
// Make a multivector for holding the redistributed data
//
RCP<MV> redistData = rcp(new MV(redistMap_, numVecs));
//
// Redistribute the data.
// This will do all the necessary communication for you.
// All processes now own enough data to do the matvec.
//
redistData->doImport(X, *importer_, Tpetra::INSERT);
//
// Perform the matvec with the data we now locally own
//
// For each column...
for(int c=0; c<numVecs; c++)
{
// Get a view of the desired column
Teuchos::ArrayRCP<Scalar> colView = redistData->getDataNonConst(c);
int offset;
// Y[0,c] = -colView[0] + 2*colView[1] - colView[2] (using local indices)
if(myRank_ > 0)
{
Y.replaceLocalValue(0, c, -colView[0] + 2*colView[1] - colView[2]);
offset = 0;
}
// Y[0,c] = 2*colView[1] - colView[2] (using local indices)
else
{
Y.replaceLocalValue(0, c, 2*colView[0] - colView[1]);
offset = 1;
}
// Y[r,c] = -colView[r-offset] + 2*colView[r+1-offset] - colView[r+2-offset]
for(int r=1; r<nlocRows-1; r++)
{
Y.replaceLocalValue(r, c, -colView[r-offset] + 2*colView[r+1-offset] - colView[r+2-offset]);
}
// Y[nlocRows-1,c] = -colView[nlocRows-1-offset] + 2*colView[nlocRows-offset] - colView[nlocRows+1-offset]
if(myRank_ < numProcs_-1)
{
Y.replaceLocalValue(nlocRows-1, c, -colView[nlocRows-1-offset] + 2*colView[nlocRows-offset] - colView[nlocRows+1-offset]);
}
// Y[nlocRows-1,c] = -colView[nlocRows-1-offset] + 2*colView[nlocRows-offset]
else
{
Y.replaceLocalValue(nlocRows-1, c, -colView[nlocRows-1-offset] + 2*colView[nlocRows-offset]);
}
}
}
