template<typename SparseMatrixType> void sparse_product(const SparseMatrixType& ref)
{
const int rows = ref.rows();
const int cols = ref.cols();
typedef typename SparseMatrixType::Scalar Scalar;
enum { Flags = SparseMatrixType::Flags };
double density = std::max(8./(rows*cols), 0.01);
typedef Matrix<Scalar,Dynamic,Dynamic> DenseMatrix;
typedef Matrix<Scalar,Dynamic,1> DenseVector;
// test matrix-matrix product
{
DenseMatrix refMat2 = DenseMatrix::Zero(rows, rows);
DenseMatrix refMat3 = DenseMatrix::Zero(rows, rows);
DenseMatrix refMat4 = DenseMatrix::Zero(rows, rows);
DenseMatrix dm4 = DenseMatrix::Zero(rows, rows);
SparseMatrixType m2(rows, rows);
SparseMatrixType m3(rows, rows);
SparseMatrixType m4(rows, rows);
initSparse<Scalar>(density, refMat2, m2);
initSparse<Scalar>(density, refMat3, m3);
initSparse<Scalar>(density, refMat4, m4);
VERIFY_IS_APPROX(m4=m2*m3, refMat4=refMat2*refMat3);
VERIFY_IS_APPROX(m4=m2.transpose()*m3, refMat4=refMat2.transpose()*refMat3);
VERIFY_IS_APPROX(m4=m2.transpose()*m3.transpose(), refMat4=refMat2.transpose()*refMat3.transpose());
VERIFY_IS_APPROX(m4=m2*m3.transpose(), refMat4=refMat2*refMat3.transpose());
// sparse * dense
VERIFY_IS_APPROX(dm4=m2*refMat3, refMat4=refMat2*refMat3);
VERIFY_IS_APPROX(dm4=m2*refMat3.transpose(), refMat4=refMat2*refMat3.transpose());
VERIFY_IS_APPROX(dm4=m2.transpose()*refMat3, refMat4=refMat2.transpose()*refMat3);
VERIFY_IS_APPROX(dm4=m2.transpose()*refMat3.transpose(), refMat4=refMat2.transpose()*refMat3.transpose());
// dense * sparse
VERIFY_IS_APPROX(dm4=refMat2*m3, refMat4=refMat2*refMat3);
VERIFY_IS_APPROX(dm4=refMat2*m3.transpose(), refMat4=refMat2*refMat3.transpose());
VERIFY_IS_APPROX(dm4=refMat2.transpose()*m3, refMat4=refMat2.transpose()*refMat3);
VERIFY_IS_APPROX(dm4=refMat2.transpose()*m3.transpose(), refMat4=refMat2.transpose()*refMat3.transpose());
VERIFY_IS_APPROX(m3=m3*m3, refMat3=refMat3*refMat3);
}
// test matrix - diagonal product
if(false) // it compiles, but the precision is terrible. probably doesn't matter in this branch....
{
DenseMatrix refM2 = DenseMatrix::Zero(rows, rows);
DenseMatrix refM3 = DenseMatrix::Zero(rows, rows);
DiagonalMatrix<DenseVector> d1(DenseVector::Random(rows));
SparseMatrixType m2(rows, rows);
SparseMatrixType m3(rows, rows);
initSparse<Scalar>(density, refM2, m2);
initSparse<Scalar>(density, refM3, m3);
VERIFY_IS_APPROX(m3=m2*d1, refM3=refM2*d1);
VERIFY_IS_APPROX(m3=m2.transpose()*d1, refM3=refM2.transpose()*d1);
VERIFY_IS_APPROX(m3=d1*m2, refM3=d1*refM2);
VERIFY_IS_APPROX(m3=d1*m2.transpose(), refM3=d1 * refM2.transpose());
}
// test self adjoint products
{
DenseMatrix b = DenseMatrix::Random(rows, rows);
DenseMatrix x = DenseMatrix::Random(rows, rows);
DenseMatrix refX = DenseMatrix::Random(rows, rows);
DenseMatrix refUp = DenseMatrix::Zero(rows, rows);
DenseMatrix refLo = DenseMatrix::Zero(rows, rows);
DenseMatrix refS = DenseMatrix::Zero(rows, rows);
SparseMatrixType mUp(rows, rows);
SparseMatrixType mLo(rows, rows);
SparseMatrixType mS(rows, rows);
do {
initSparse<Scalar>(density, refUp, mUp, ForceRealDiag|/*ForceNonZeroDiag|*/MakeUpperTriangular);
} while (refUp.isZero());
refLo = refUp.transpose().conjugate();
mLo = mUp.transpose().conjugate();
refS = refUp + refLo;
refS.diagonal() *= 0.5;
mS = mUp + mLo;
for (int k=0; k<mS.outerSize(); ++k)
for (typename SparseMatrixType::InnerIterator it(mS,k); it; ++it)
if (it.index() == k)
it.valueRef() *= 0.5;
VERIFY_IS_APPROX(refS.adjoint(), refS);
VERIFY_IS_APPROX(mS.transpose().conjugate(), mS);
VERIFY_IS_APPROX(mS, refS);
VERIFY_IS_APPROX(x=mS*b, refX=refS*b);
VERIFY_IS_APPROX(x=mUp.template marked<UpperTriangular|SelfAdjoint>()*b, refX=refS*b);
VERIFY_IS_APPROX(x=mLo.template marked<LowerTriangular|SelfAdjoint>()*b, refX=refS*b);
VERIFY_IS_APPROX(x=mS.template marked<SelfAdjoint>()*b, refX=refS*b);
}
}
template<typename SparseMatrixType> void sparse_product()
{
typedef typename SparseMatrixType::Index Index;
Index n = 100;
const Index rows = internal::random<int>(1,n);
const Index cols = internal::random<int>(1,n);
const Index depth = internal::random<int>(1,n);
typedef typename SparseMatrixType::Scalar Scalar;
enum { Flags = SparseMatrixType::Flags };
double density = (std::max)(8./(rows*cols), 0.1);
typedef Matrix<Scalar,Dynamic,Dynamic> DenseMatrix;
typedef Matrix<Scalar,Dynamic,1> DenseVector;
Scalar s1 = internal::random<Scalar>();
Scalar s2 = internal::random<Scalar>();
// test matrix-matrix product
{
DenseMatrix refMat2 = DenseMatrix::Zero(rows, depth);
DenseMatrix refMat2t = DenseMatrix::Zero(depth, rows);
DenseMatrix refMat3 = DenseMatrix::Zero(depth, cols);
DenseMatrix refMat3t = DenseMatrix::Zero(cols, depth);
DenseMatrix refMat4 = DenseMatrix::Zero(rows, cols);
DenseMatrix refMat4t = DenseMatrix::Zero(cols, rows);
DenseMatrix refMat5 = DenseMatrix::Random(depth, cols);
DenseMatrix refMat6 = DenseMatrix::Random(rows, rows);
DenseMatrix dm4 = DenseMatrix::Zero(rows, rows);
// DenseVector dv1 = DenseVector::Random(rows);
SparseMatrixType m2 (rows, depth);
SparseMatrixType m2t(depth, rows);
SparseMatrixType m3 (depth, cols);
SparseMatrixType m3t(cols, depth);
SparseMatrixType m4 (rows, cols);
SparseMatrixType m4t(cols, rows);
SparseMatrixType m6(rows, rows);
initSparse(density, refMat2, m2);
initSparse(density, refMat2t, m2t);
initSparse(density, refMat3, m3);
initSparse(density, refMat3t, m3t);
initSparse(density, refMat4, m4);
initSparse(density, refMat4t, m4t);
initSparse(density, refMat6, m6);
// int c = internal::random<int>(0,depth-1);
// sparse * sparse
VERIFY_IS_APPROX(m4=m2*m3, refMat4=refMat2*refMat3);
VERIFY_IS_APPROX(m4=m2t.transpose()*m3, refMat4=refMat2t.transpose()*refMat3);
VERIFY_IS_APPROX(m4=m2t.transpose()*m3t.transpose(), refMat4=refMat2t.transpose()*refMat3t.transpose());
VERIFY_IS_APPROX(m4=m2*m3t.transpose(), refMat4=refMat2*refMat3t.transpose());
VERIFY_IS_APPROX(m4 = m2*m3/s1, refMat4 = refMat2*refMat3/s1);
VERIFY_IS_APPROX(m4 = m2*m3*s1, refMat4 = refMat2*refMat3*s1);
VERIFY_IS_APPROX(m4 = s2*m2*m3*s1, refMat4 = s2*refMat2*refMat3*s1);
VERIFY_IS_APPROX(m4=(m2*m3).pruned(0), refMat4=refMat2*refMat3);
VERIFY_IS_APPROX(m4=(m2t.transpose()*m3).pruned(0), refMat4=refMat2t.transpose()*refMat3);
VERIFY_IS_APPROX(m4=(m2t.transpose()*m3t.transpose()).pruned(0), refMat4=refMat2t.transpose()*refMat3t.transpose());
VERIFY_IS_APPROX(m4=(m2*m3t.transpose()).pruned(0), refMat4=refMat2*refMat3t.transpose());
// test aliasing
m4 = m2; refMat4 = refMat2;
VERIFY_IS_APPROX(m4=m4*m3, refMat4=refMat4*refMat3);
// sparse * dense
VERIFY_IS_APPROX(dm4=m2*refMat3, refMat4=refMat2*refMat3);
VERIFY_IS_APPROX(dm4=m2*refMat3t.transpose(), refMat4=refMat2*refMat3t.transpose());
VERIFY_IS_APPROX(dm4=m2t.transpose()*refMat3, refMat4=refMat2t.transpose()*refMat3);
VERIFY_IS_APPROX(dm4=m2t.transpose()*refMat3t.transpose(), refMat4=refMat2t.transpose()*refMat3t.transpose());
VERIFY_IS_APPROX(dm4=m2*(refMat3+refMat3), refMat4=refMat2*(refMat3+refMat3));
VERIFY_IS_APPROX(dm4=m2t.transpose()*(refMat3+refMat5)*0.5, refMat4=refMat2t.transpose()*(refMat3+refMat5)*0.5);
// dense * sparse
VERIFY_IS_APPROX(dm4=refMat2*m3, refMat4=refMat2*refMat3);
VERIFY_IS_APPROX(dm4=refMat2*m3t.transpose(), refMat4=refMat2*refMat3t.transpose());
VERIFY_IS_APPROX(dm4=refMat2t.transpose()*m3, refMat4=refMat2t.transpose()*refMat3);
VERIFY_IS_APPROX(dm4=refMat2t.transpose()*m3t.transpose(), refMat4=refMat2t.transpose()*refMat3t.transpose());
// sparse * dense and dense * sparse outer product
test_outer<SparseMatrixType,DenseMatrix>::run(m2,m4,refMat2,refMat4);
VERIFY_IS_APPROX(m6=m6*m6, refMat6=refMat6*refMat6);
}
// test matrix - diagonal product
{
DenseMatrix refM2 = DenseMatrix::Zero(rows, cols);
DenseMatrix refM3 = DenseMatrix::Zero(rows, cols);
DenseMatrix d3 = DenseMatrix::Zero(rows, cols);
DiagonalMatrix<Scalar,Dynamic> d1(DenseVector::Random(cols));
DiagonalMatrix<Scalar,Dynamic> d2(DenseVector::Random(rows));
SparseMatrixType m2(rows, cols);
SparseMatrixType m3(rows, cols);
initSparse<Scalar>(density, refM2, m2);
initSparse<Scalar>(density, refM3, m3);
VERIFY_IS_APPROX(m3=m2*d1, refM3=refM2*d1);
VERIFY_IS_APPROX(m3=m2.transpose()*d2, refM3=refM2.transpose()*d2);
VERIFY_IS_APPROX(m3=d2*m2, refM3=d2*refM2);
VERIFY_IS_APPROX(m3=d1*m2.transpose(), refM3=d1*refM2.transpose());
// evaluate to a dense matrix to check the .row() and .col() iterator functions
VERIFY_IS_APPROX(d3=m2*d1, refM3=refM2*d1);
VERIFY_IS_APPROX(d3=m2.transpose()*d2, refM3=refM2.transpose()*d2);
VERIFY_IS_APPROX(d3=d2*m2, refM3=d2*refM2);
VERIFY_IS_APPROX(d3=d1*m2.transpose(), refM3=d1*refM2.transpose());
}
// test self adjoint products
{
DenseMatrix b = DenseMatrix::Random(rows, rows);
DenseMatrix x = DenseMatrix::Random(rows, rows);
DenseMatrix refX = DenseMatrix::Random(rows, rows);
DenseMatrix refUp = DenseMatrix::Zero(rows, rows);
DenseMatrix refLo = DenseMatrix::Zero(rows, rows);
DenseMatrix refS = DenseMatrix::Zero(rows, rows);
SparseMatrixType mUp(rows, rows);
SparseMatrixType mLo(rows, rows);
SparseMatrixType mS(rows, rows);
do {
initSparse<Scalar>(density, refUp, mUp, ForceRealDiag|/*ForceNonZeroDiag|*/MakeUpperTriangular);
} while (refUp.isZero());
refLo = refUp.adjoint();
mLo = mUp.adjoint();
refS = refUp + refLo;
refS.diagonal() *= 0.5;
mS = mUp + mLo;
// TODO be able to address the diagonal....
for (int k=0; k<mS.outerSize(); ++k)
for (typename SparseMatrixType::InnerIterator it(mS,k); it; ++it)
if (it.index() == k)
it.valueRef() *= 0.5;
VERIFY_IS_APPROX(refS.adjoint(), refS);
VERIFY_IS_APPROX(mS.adjoint(), mS);
VERIFY_IS_APPROX(mS, refS);
VERIFY_IS_APPROX(x=mS*b, refX=refS*b);
VERIFY_IS_APPROX(x=mUp.template selfadjointView<Upper>()*b, refX=refS*b);
VERIFY_IS_APPROX(x=mLo.template selfadjointView<Lower>()*b, refX=refS*b);
VERIFY_IS_APPROX(x=mS.template selfadjointView<Upper|Lower>()*b, refX=refS*b);
}
}
template<typename Scalar> void sparse_llt(int rows, int cols)
{
double density = std::max(8./(rows*cols), 0.01);
typedef Matrix<Scalar,Dynamic,Dynamic> DenseMatrix;
typedef Matrix<Scalar,Dynamic,1> DenseVector;
// TODO fix the issue with complex (see SparseLLT::solveInPlace)
SparseMatrix<Scalar> m2(rows, cols);
DenseMatrix refMat2(rows, cols);
DenseVector b = DenseVector::Random(cols);
DenseVector ref_x(cols), x(cols);
DenseMatrix B = DenseMatrix::Random(rows,cols);
DenseMatrix ref_X(rows,cols), X(rows,cols);
initSparse<Scalar>(density, refMat2, m2, ForceNonZeroDiag|MakeLowerTriangular, 0, 0);
for(int i=0; i<rows; ++i)
m2.coeffRef(i,i) = refMat2(i,i) = internal::abs(internal::real(refMat2(i,i)));
ref_x = refMat2.template selfadjointView<Lower>().llt().solve(b);
if (!NumTraits<Scalar>::IsComplex)
{
x = b;
SparseLLT<SparseMatrix<Scalar> > (m2).solveInPlace(x);
VERIFY(ref_x.isApprox(x,test_precision<Scalar>()) && "LLT: default");
}
#ifdef EIGEN_CHOLMOD_SUPPORT
// legacy API
{
// Cholmod, as configured in CholmodSupport.h, only supports self-adjoint matrices
SparseMatrix<Scalar> m3 = m2.adjoint()*m2;
DenseMatrix refMat3 = refMat2.adjoint()*refMat2;
ref_x = refMat3.template selfadjointView<Lower>().llt().solve(b);
x = b;
SparseLLT<SparseMatrix<Scalar>, Cholmod>(m3).solveInPlace(x);
VERIFY((m3*x).isApprox(b,test_precision<Scalar>()) && "LLT legacy: cholmod solveInPlace");
x = SparseLLT<SparseMatrix<Scalar>, Cholmod>(m3).solve(b);
VERIFY(ref_x.isApprox(x,test_precision<Scalar>()) && "LLT legacy: cholmod solve");
}
// new API
{
// Cholmod, as configured in CholmodSupport.h, only supports self-adjoint matrices
SparseMatrix<Scalar> m3 = m2 * m2.adjoint(), m3_lo(rows,rows), m3_up(rows,rows);
DenseMatrix refMat3 = refMat2 * refMat2.adjoint();
m3_lo.template selfadjointView<Lower>().rankUpdate(m2,0);
m3_up.template selfadjointView<Upper>().rankUpdate(m2,0);
// with a single vector as the rhs
ref_x = refMat3.template selfadjointView<Lower>().llt().solve(b);
x = CholmodDecomposition<SparseMatrix<Scalar>, Lower>(m3).solve(b);
VERIFY(ref_x.isApprox(x,test_precision<Scalar>()) && "LLT: cholmod solve, single dense rhs");
x = CholmodDecomposition<SparseMatrix<Scalar>, Upper>(m3).solve(b);
VERIFY(ref_x.isApprox(x,test_precision<Scalar>()) && "LLT: cholmod solve, single dense rhs");
x = CholmodDecomposition<SparseMatrix<Scalar>, Lower>(m3_lo).solve(b);
VERIFY(ref_x.isApprox(x,test_precision<Scalar>()) && "LLT: cholmod solve, single dense rhs");
x = CholmodDecomposition<SparseMatrix<Scalar>, Upper>(m3_up).solve(b);
VERIFY(ref_x.isApprox(x,test_precision<Scalar>()) && "LLT: cholmod solve, single dense rhs");
// with multiple rhs
ref_X = refMat3.template selfadjointView<Lower>().llt().solve(B);
#ifndef EIGEN_DEFAULT_TO_ROW_MAJOR
// TODO make sure the API is properly documented about this fact
X = CholmodDecomposition<SparseMatrix<Scalar>, Lower>(m3).solve(B);
VERIFY(ref_X.isApprox(X,test_precision<Scalar>()) && "LLT: cholmod solve, multiple dense rhs");
X = CholmodDecomposition<SparseMatrix<Scalar>, Upper>(m3).solve(B);
VERIFY(ref_X.isApprox(X,test_precision<Scalar>()) && "LLT: cholmod solve, multiple dense rhs");
#endif
// with a sparse rhs
SparseMatrix<Scalar> spB(rows,cols), spX(rows,cols);
B.diagonal().array() += 1;
spB = B.sparseView(0.5,1);
ref_X = refMat3.template selfadjointView<Lower>().llt().solve(DenseMatrix(spB));
spX = CholmodDecomposition<SparseMatrix<Scalar>, Lower>(m3).solve(spB);
VERIFY(ref_X.isApprox(spX.toDense(),test_precision<Scalar>()) && "LLT: cholmod solve, multiple sparse rhs");
spX = CholmodDecomposition<SparseMatrix<Scalar>, Upper>(m3).solve(spB);
VERIFY(ref_X.isApprox(spX.toDense(),test_precision<Scalar>()) && "LLT: cholmod solve, multiple sparse rhs");
}
#endif
}
