| 1 |
// @HEADER |
| 2 |
// ***************************************************************************** |
| 3 |
// MueLu: A package for multigrid based preconditioning |
| 4 |
// |
| 5 |
// Copyright 2012 NTESS and the MueLu contributors. |
| 6 |
// SPDX-License-Identifier: BSD-3-Clause |
| 7 |
// ***************************************************************************** |
| 8 |
// @HEADER |
| 9 |
|
| 10 |
#ifndef MUELU_CONSTRAINT_DEF_HPP |
| 11 |
#define MUELU_CONSTRAINT_DEF_HPP |
| 12 |
|
| 13 |
#include <Teuchos_BLAS.hpp> |
| 14 |
#include <Teuchos_LAPACK.hpp> |
| 15 |
#include <Teuchos_SerialDenseVector.hpp> |
| 16 |
#include <Teuchos_SerialDenseMatrix.hpp> |
| 17 |
#include <Teuchos_SerialDenseHelpers.hpp> |
| 18 |
|
| 19 |
#include <Xpetra_ImportFactory.hpp> |
| 20 |
#include <Xpetra_Map.hpp> |
| 21 |
#include <Xpetra_Matrix.hpp> |
| 22 |
#include <Xpetra_MultiVectorFactory.hpp> |
| 23 |
#include <Xpetra_MultiVector.hpp> |
| 24 |
#include <Xpetra_CrsGraph.hpp> |
| 25 |
|
| 26 |
#include "MueLu_Exceptions.hpp" |
| 27 |
|
| 28 |
#include "MueLu_Constraint_decl.hpp" |
| 29 |
|
| 30 |
namespace MueLu { |
| 31 |
|
| 32 |
template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node> |
| 33 |
void Constraint<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Setup(const MultiVector& /* B */, const MultiVector& Bc, RCP<const CrsGraph> Ppattern) { |
| 34 |
const size_t NSDim = Bc.getNumVectors(); |
| 35 |
|
| 36 |
Ppattern_ = Ppattern; |
| 37 |
|
| 38 |
size_t numRows = Ppattern_->getLocalNumRows(); |
| 39 |
XXtInv_.resize(numRows); |
| 40 |
|
| 41 |
RCP<const Import> importer = Ppattern_->getImporter(); |
| 42 |
|
| 43 |
X_ = MultiVectorFactory::Build(Ppattern_->getColMap(), NSDim); |
| 44 |
if (!importer.is_null()) |
| 45 |
X_->doImport(Bc, *importer, Xpetra::INSERT); |
| 46 |
else |
| 47 |
*X_ = Bc; |
| 48 |
|
| 49 |
std::vector<const SC*> Xval(NSDim); |
| 50 |
for (size_t j = 0; j < NSDim; j++) |
| 51 |
Xval[j] = X_->getData(j).get(); |
| 52 |
|
| 53 |
SC zero = Teuchos::ScalarTraits<SC>::zero(); |
| 54 |
SC one = Teuchos::ScalarTraits<SC>::one(); |
| 55 |
|
| 56 |
Teuchos::BLAS<LO, SC> blas; |
| 57 |
Teuchos::LAPACK<LO, SC> lapack; |
| 58 |
LO lwork = 3 * NSDim; |
| 59 |
ArrayRCP<LO> IPIV(NSDim); |
| 60 |
ArrayRCP<SC> WORK(lwork); |
| 61 |
|
| 62 |
for (size_t i = 0; i < numRows; i++) { |
| 63 |
Teuchos::ArrayView<const LO> indices; |
| 64 |
Ppattern_->getLocalRowView(i, indices); |
| 65 |
|
| 66 |
size_t nnz = indices.size(); |
| 67 |
|
| 68 |
XXtInv_[i] = Teuchos::SerialDenseMatrix<LO, SC>(NSDim, NSDim, false /*zeroOut*/); |
| 69 |
Teuchos::SerialDenseMatrix<LO, SC>& XXtInv = XXtInv_[i]; |
| 70 |
|
| 71 |
if (NSDim == 1) { |
| 72 |
SC d = zero; |
| 73 |
for (size_t j = 0; j < nnz; j++) |
| 74 |
d += Xval[0][indices[j]] * Xval[0][indices[j]]; |
| 75 |
XXtInv(0, 0) = one / d; |
| 76 |
|
| 77 |
} else { |
| 78 |
Teuchos::SerialDenseMatrix<LO, SC> locX(NSDim, nnz, false /*zeroOut*/); |
| 79 |
for (size_t j = 0; j < nnz; j++) |
| 80 |
for (size_t k = 0; k < NSDim; k++) |
| 81 |
locX(k, j) = Xval[k][indices[j]]; |
| 82 |
|
| 83 |
// XXtInv_ = (locX*locX^T)^{-1} |
| 84 |
blas.GEMM(Teuchos::NO_TRANS, Teuchos::CONJ_TRANS, NSDim, NSDim, nnz, |
| 85 |
one, locX.values(), locX.stride(), |
| 86 |
locX.values(), locX.stride(), |
| 87 |
zero, XXtInv.values(), XXtInv.stride()); |
| 88 |
|
| 89 |
LO info; |
| 90 |
// Compute LU factorization using partial pivoting with row exchanges |
| 91 |
lapack.GETRF(NSDim, NSDim, XXtInv.values(), XXtInv.stride(), IPIV.get(), &info); |
| 92 |
// Use the computed factorization to compute the inverse |
| 93 |
lapack.GETRI(NSDim, XXtInv.values(), XXtInv.stride(), IPIV.get(), WORK.get(), lwork, &info); |
| 94 |
} |
| 95 |
} |
| 96 |
} |
| 97 |
|
| 98 |
//! \note We assume that the graph of Projected is the same as Ppattern_ |
| 99 |
template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node> |
| 100 |
void Constraint<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Apply(const Matrix& P, Matrix& Projected) const { |
| 101 |
// We check only row maps. Column may be different. |
| 102 |
TEUCHOS_TEST_FOR_EXCEPTION(!P.getRowMap()->isSameAs(*Projected.getRowMap()), Exceptions::Incompatible, |
| 103 |
"Row maps are incompatible"); |
| 104 |
const size_t NSDim = X_->getNumVectors(); |
| 105 |
const size_t numRows = P.getLocalNumRows(); |
| 106 |
|
| 107 |
const Map& colMap = *P.getColMap(); |
| 108 |
const Map& PColMap = *Projected.getColMap(); |
| 109 |
|
| 110 |
Projected.resumeFill(); |
| 111 |
|
| 112 |
Teuchos::ArrayView<const LO> indices, pindices; |
| 113 |
Teuchos::ArrayView<const SC> values, pvalues; |
| 114 |
Teuchos::Array<SC> valuesAll(colMap.getLocalNumElements()), newValues; |
| 115 |
|
| 116 |
LO invalid = Teuchos::OrdinalTraits<LO>::invalid(); |
| 117 |
LO oneLO = Teuchos::OrdinalTraits<LO>::one(); |
| 118 |
SC zero = Teuchos::ScalarTraits<SC>::zero(); |
| 119 |
SC one = Teuchos::ScalarTraits<SC>::one(); |
| 120 |
|
| 121 |
std::vector<const SC*> Xval(NSDim); |
| 122 |
for (size_t j = 0; j < NSDim; j++) |
| 123 |
Xval[j] = X_->getData(j).get(); |
| 124 |
|
| 125 |
for (size_t i = 0; i < numRows; i++) { |
| 126 |
P.getLocalRowView(i, indices, values); |
| 127 |
Projected.getLocalRowView(i, pindices, pvalues); |
| 128 |
|
| 129 |
size_t nnz = indices.size(); // number of nonzeros in the supplied matrix |
| 130 |
size_t pnnz = pindices.size(); // number of nonzeros in the constrained matrix |
| 131 |
|
| 132 |
newValues.resize(pnnz); |
| 133 |
|
| 134 |
// Step 1: fix stencil |
| 135 |
// Projected *must* already have the correct stencil |
| 136 |
|
| 137 |
// Step 2: copy correct stencil values |
| 138 |
// The algorithm is very similar to the one used in the calculation of |
| 139 |
// Frobenius dot product, see src/Transfers/Energy-Minimization/Solvers/MueLu_CGSolver_def.hpp |
| 140 |
|
| 141 |
// NOTE: using extra array allows us to skip the search among indices |
| 142 |
for (size_t j = 0; j < nnz; j++) |
| 143 |
valuesAll[indices[j]] = values[j]; |
| 144 |
for (size_t j = 0; j < pnnz; j++) { |
| 145 |
LO ind = colMap.getLocalElement(PColMap.getGlobalElement(pindices[j])); // FIXME: we could do that before the full loop just once |
| 146 |
if (ind != invalid) |
| 147 |
// index indices[j] is part of template, copy corresponding value |
| 148 |
newValues[j] = valuesAll[ind]; |
| 149 |
else |
| 150 |
newValues[j] = zero; |
| 151 |
} |
| 152 |
for (size_t j = 0; j < nnz; j++) |
| 153 |
valuesAll[indices[j]] = zero; |
| 154 |
|
| 155 |
// Step 3: project to the space |
| 156 |
Teuchos::SerialDenseMatrix<LO, SC>& XXtInv = XXtInv_[i]; |
| 157 |
|
| 158 |
Teuchos::SerialDenseMatrix<LO, SC> locX(NSDim, pnnz, false); |
| 159 |
for (size_t j = 0; j < pnnz; j++) |
| 160 |
for (size_t k = 0; k < NSDim; k++) |
| 161 |
locX(k, j) = Xval[k][pindices[j]]; |
| 162 |
|
| 163 |
Teuchos::SerialDenseVector<LO, SC> val(pnnz, false), val1(NSDim, false), val2(NSDim, false); |
| 164 |
for (size_t j = 0; j < pnnz; j++) |
| 165 |
val[j] = newValues[j]; |
| 166 |
|
| 167 |
Teuchos::BLAS<LO, SC> blas; |
| 168 |
// val1 = locX * val; |
| 169 |
blas.GEMV(Teuchos::NO_TRANS, NSDim, pnnz, |
| 170 |
one, locX.values(), locX.stride(), |
| 171 |
val.values(), oneLO, |
| 172 |
zero, val1.values(), oneLO); |
| 173 |
// val2 = XXtInv * val1 |
| 174 |
blas.GEMV(Teuchos::NO_TRANS, NSDim, NSDim, |
| 175 |
one, XXtInv.values(), XXtInv.stride(), |
| 176 |
val1.values(), oneLO, |
| 177 |
zero, val2.values(), oneLO); |
| 178 |
// val = X^T * val2 |
| 179 |
blas.GEMV(Teuchos::CONJ_TRANS, NSDim, pnnz, |
| 180 |
one, locX.values(), locX.stride(), |
| 181 |
val2.values(), oneLO, |
| 182 |
zero, val.values(), oneLO); |
| 183 |
|
| 184 |
for (size_t j = 0; j < pnnz; j++) |
| 185 |
newValues[j] -= val[j]; |
| 186 |
|
| 187 |
Projected.replaceLocalValues(i, pindices, newValues); |
| 188 |
} |
| 189 |
|
| 190 |
Projected.fillComplete(Projected.getDomainMap(), Projected.getRangeMap()); // FIXME: maps needed? |
| 191 |
} |
| 192 |
|
| 193 |
} // namespace MueLu |
| 194 |
|
| 195 |
#endif // ifndef MUELU_CONSTRAINT_DEF_HPP |