void Tikhonov
( Orientation orientation,
const Matrix<F>& A,
const Matrix<F>& B,
const Matrix<F>& G,
Matrix<F>& X,
TikhonovAlg alg )
{
DEBUG_CSE
const bool normal = ( orientation==NORMAL );
const Int m = ( normal ? A.Height() : A.Width() );
const Int n = ( normal ? A.Width() : A.Height() );
if( G.Width() != n )
LogicError("Tikhonov matrix was the wrong width");
if( orientation == TRANSPOSE && IsComplex<F>::value )
LogicError("Transpose version of complex Tikhonov not yet supported");
if( m >= n )
{
Matrix<F> Z;
if( alg == TIKHONOV_CHOLESKY )
{
if( orientation == NORMAL )
Herk( LOWER, ADJOINT, Base<F>(1), A, Z );
else
Herk( LOWER, NORMAL, Base<F>(1), A, Z );
Herk( LOWER, ADJOINT, Base<F>(1), G, Base<F>(1), Z );
Cholesky( LOWER, Z );
}
else
{
const Int mG = G.Height();
Zeros( Z, m+mG, n );
auto ZT = Z( IR(0,m), IR(0,n) );
auto ZB = Z( IR(m,m+mG), IR(0,n) );
if( orientation == NORMAL )
ZT = A;
else
Adjoint( A, ZT );
ZB = G;
qr::ExplicitTriang( Z );
}
if( orientation == NORMAL )
Gemm( ADJOINT, NORMAL, F(1), A, B, X );
else
Gemm( NORMAL, NORMAL, F(1), A, B, X );
cholesky::SolveAfter( LOWER, NORMAL, Z, X );
}
else
{
LogicError("This case not yet supported");
}
}
void Tikhonov
( Orientation orientation,
const SparseMatrix<F>& A,
const Matrix<F>& B,
const SparseMatrix<F>& G,
Matrix<F>& X,
const LeastSquaresCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
// Explicitly form W := op(A)
// ==========================
SparseMatrix<F> W;
if( orientation == NORMAL )
W = A;
else if( orientation == TRANSPOSE )
Transpose( A, W );
else
Adjoint( A, W );
const Int m = W.Height();
const Int n = W.Width();
const Int numRHS = B.Width();
// Embed into a higher-dimensional problem via appending regularization
// ====================================================================
SparseMatrix<F> WEmb;
if( m >= n )
VCat( W, G, WEmb );
else
HCat( W, G, WEmb );
Matrix<F> BEmb;
Zeros( BEmb, WEmb.Height(), numRHS );
if( m >= n )
{
auto BEmbT = BEmb( IR(0,m), IR(0,numRHS) );
BEmbT = B;
}
else
BEmb = B;
// Solve the higher-dimensional problem
// ====================================
Matrix<F> XEmb;
LeastSquares( NORMAL, WEmb, BEmb, XEmb, ctrl );
// Extract the solution
// ====================
if( m >= n )
X = XEmb;
else
X = XEmb( IR(0,n), IR(0,numRHS) );
}
void Skeleton
( const AbstractDistMatrix<F>& APre,
DistPermutation& PR,
DistPermutation& PC,
AbstractDistMatrix<F>& Z,
const QRCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
DistMatrixReadProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.GetLocked();
const Grid& g = A.Grid();
DistMatrix<F> AAdj(g);
Adjoint( A, AAdj );
// Find the row permutation
DistMatrix<F> B(AAdj);
DistMatrix<F,MD,STAR> householderScalars(g);
DistMatrix<Base<F>,MD,STAR> signature(g);
QR( B, householderScalars, signature, PR, ctrl );
const Int numSteps = householderScalars.Height();
B.Resize( B.Height(), numSteps );
// Form K' := (A pinv(AR))' = pinv(AR') A'
DistMatrix<F> KAdj(g);
qr::SolveAfter( NORMAL, B, householderScalars, signature, AAdj, KAdj );
// Form K := (K')'
DistMatrix<F> K(g);
Adjoint( KAdj, K );
// Find the column permutation (force the same number of steps)
B = A;
auto secondCtrl = ctrl;
secondCtrl.adaptive = false;
secondCtrl.boundRank = true;
secondCtrl.maxRank = numSteps;
QR( B, householderScalars, signature, PC, secondCtrl );
// Form Z := pinv(AC) K = pinv(AC) (A pinv(AR))
B.Resize( B.Height(), numSteps );
qr::SolveAfter( NORMAL, B, householderScalars, signature, K, Z );
}
bool Matrix33 :: Invert ( Matrix33 & A, Matrix33 & Result )
{
double Determinate = GetDeterminate ( A );
if ( Determinate == 0.0 )
return false;
Adjoint ( A, Result );
Multiply ( Result, 1.0 / Determinate );
return true;
};
/**
* @brief Adjoint transform
*
* @param m To transform
* @return Transform
*/
inline
Matrix <T>
operator->* (const Matrix <T> & m) {
if (_wl_fam == ID)
return m;
else
{
Matrix <T> res (m);
Adjoint (m, res);
return res;
}
}
void Skeleton
( const Matrix<F>& A,
Permutation& PR,
Permutation& PC,
Matrix<F>& Z,
const QRCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
Matrix<F> AAdj;
Adjoint( A, AAdj );
// Find the row permutation
Matrix<F> B(AAdj);
Matrix<F> householderScalars;
Matrix<Base<F>> signature;
QR( B, householderScalars, signature, PR, ctrl );
const Int numSteps = householderScalars.Height();
B.Resize( B.Height(), numSteps );
// Form K' := (A pinv(AR))' = pinv(AR') A'
Matrix<F> KAdj;
qr::SolveAfter( NORMAL, B, householderScalars, signature, AAdj, KAdj );
// Form K := (K')'
Matrix<F> K;
Adjoint( KAdj, K );
// Find the column permutation (force the same number of steps)
B = A;
auto secondCtrl = ctrl;
secondCtrl.adaptive = false;
secondCtrl.boundRank = true;
secondCtrl.maxRank = numSteps;
QR( B, householderScalars, signature, PC, secondCtrl );
// Form Z := pinv(AC) K = pinv(AC) (A pinv(AR))
B.Resize( B.Height(), numSteps );
qr::SolveAfter( NORMAL, B, householderScalars, signature, K, Z );
}
Matrix Matrix::Reversed() const
{
Matrix result;
double determinant = Determinant();
if (determinant != 0)
{
result = Adjoint();
result.Transpose();
result *= (1 / determinant);
}
return result;
}
matrix4x4 Inverse(const matrix4x4 &m)
{
matrix4x4 retVal = Adjoint(m);
scalar_t det = Determinant(m);
assert(det);
for (int i = 0; i < 4; ++i)
{
for (int j = 0; j < 4; ++j)
{
retVal(i)[j] /= det;
}
}
return retVal;
}
void HermitianFromEVD
( UpperOrLower uplo,
AbstractDistMatrix<F>& APre,
const AbstractDistMatrix<Base<F>>& wPre,
const AbstractDistMatrix<F>& ZPre )
{
DEBUG_CSE
typedef Base<F> Real;
DistMatrixWriteProxy<F,F,MC,MR> AProx( APre );
DistMatrixReadProxy<Real,Real,VR,STAR> wProx( wPre );
DistMatrixReadProxy<F,F,MC,MR> ZProx( ZPre );
auto& A = AProx.Get();
auto& w = wProx.GetLocked();
auto& Z = ZProx.GetLocked();
const Grid& g = A.Grid();
DistMatrix<F,MC, STAR> Z1_MC_STAR(g);
DistMatrix<F,VR, STAR> Z1_VR_STAR(g);
DistMatrix<F,STAR,MR > Z1Adj_STAR_MR(g);
const Int m = Z.Height();
const Int n = Z.Width();
A.Resize( m, m );
if( uplo == LOWER )
MakeTrapezoidal( UPPER, A, 1 );
else
MakeTrapezoidal( LOWER, A, -1 );
const Int bsize = Blocksize();
for( Int k=0; k<n; k+=bsize )
{
const Int nb = Min(bsize,n-k);
auto Z1 = Z( ALL, IR(k,k+nb) );
auto w1 = w( IR(k,k+nb), ALL );
Z1_MC_STAR.AlignWith( A );
Z1_MC_STAR = Z1;
Z1_VR_STAR.AlignWith( A );
Z1_VR_STAR = Z1_MC_STAR;
DiagonalScale( RIGHT, NORMAL, w1, Z1_VR_STAR );
Z1Adj_STAR_MR.AlignWith( A );
Adjoint( Z1_VR_STAR, Z1Adj_STAR_MR );
LocalTrrk( uplo, F(1), Z1_MC_STAR, Z1Adj_STAR_MR, F(1), A );
}
}
Base<Field> ProductLanczosDecomp
( const DistSparseMatrix<Field>& A,
DistMultiVec<Field>& V,
AbstractDistMatrix<Base<Field>>& T,
DistMultiVec<Field>& v,
Int basisSize )
{
EL_DEBUG_CSE
const Int m = A.Height();
const Int n = A.Width();
mpi::Comm comm = A.Comm();
DistMultiVec<Field> S(comm);
// Cache the adjoint
// -----------------
DistSparseMatrix<Field> AAdj(comm);
Adjoint( A, AAdj );
if( m >= n )
{
auto applyA =
[&]( const DistMultiVec<Field>& X, DistMultiVec<Field>& Y )
{
Zeros( S, m, X.Width() );
Multiply( NORMAL, Field(1), A, X, Field(0), S );
Zeros( Y, n, X.Width() );
Multiply( NORMAL, Field(1), AAdj, S, Field(0), Y );
};
return LanczosDecomp( n, applyA, V, T, v, basisSize );
}
else
{
auto applyA =
[&]( const DistMultiVec<Field>& X, DistMultiVec<Field>& Y )
{
Zeros( S, n, X.Width() );
Multiply( NORMAL, Field(1), AAdj, X, Field(0), S );
Zeros( Y, m, X.Width() );
Multiply( NORMAL, Field(1), A, S, Field(0), Y );
};
return LanczosDecomp( m, applyA, V, T, v, basisSize );
}
}
inline void
SVD
( DistMatrix<F>& A,
DistMatrix<typename Base<F>::type,VR,STAR>& s,
DistMatrix<F>& V,
double heightRatio )
{
#ifndef RELEASE
PushCallStack("SVD");
if( heightRatio <= 1.0 )
throw std::logic_error("Nonsensical switchpoint for SVD");
#endif
typedef typename Base<F>::type Real;
// Check if we need to rescale the matrix, and do so if necessary
bool needRescaling;
Real scale;
svd::CheckScale( A, needRescaling, scale );
if( needRescaling )
Scale( scale, A );
// TODO: Switch between different algorithms. For instance, starting
// with a QR decomposition of tall-skinny matrices.
if( A.Height() >= A.Width() )
{
svd::SVDUpper( A, s, V, heightRatio );
}
else
{
// Lower bidiagonalization is not yet supported, so we instead play a
// trick to get the SVD of A.
Adjoint( A, V );
svd::SVDUpper( V, s, A, heightRatio );
}
// Rescale the singular values if necessary
if( needRescaling )
Scal( 1/scale, s );
#ifndef RELEASE
PopCallStack();
#endif
}
void ProductLanczos
( const SparseMatrix<Field>& A,
Matrix<Base<Field>>& T,
Int basisSize )
{
EL_DEBUG_CSE
const Int m = A.Height();
const Int n = A.Width();
Matrix<Field> S;
// Cache the adjoint
// -----------------
SparseMatrix<Field> AAdj;
Adjoint( A, AAdj );
if( m >= n )
{
auto applyA =
[&]( const Matrix<Field>& X, Matrix<Field>& Y )
{
Zeros( S, m, X.Width() );
Multiply( NORMAL, Field(1), A, X, Field(0), S );
Zeros( Y, n, X.Width() );
Multiply( NORMAL, Field(1), AAdj, S, Field(0), Y );
};
Lanczos<Field>( n, applyA, T, basisSize );
}
else
{
auto applyA =
[&]( const Matrix<Field>& X, Matrix<Field>& Y )
{
Zeros( S, n, X.Width() );
Multiply( NORMAL, Field(1), AAdj, X, Field(0), S );
Zeros( Y, m, X.Width() );
Multiply( NORMAL, Field(1), A, S, Field(0), Y );
};
Lanczos<Field>( m, applyA, T, basisSize );
}
}
inline void
SingularValues
( DistMatrix<F>& A,
DistMatrix<typename Base<F>::type,VR,STAR>& s,
double heightRatio )
{
#ifndef RELEASE
PushCallStack("SingularValues");
#endif
typedef typename Base<F>::type R;
// Check if we need to rescale the matrix, and do so if necessary
bool needRescaling;
R scale;
svd::CheckScale( A, needRescaling, scale );
if( needRescaling )
Scale( scale, A );
// TODO: Switch between different algorithms. For instance, starting
// with a QR decomposition of tall-skinny matrices.
if( A.Height() >= A.Width() )
{
svd::SingularValuesUpper( A, s, heightRatio );
}
else
{
// Lower bidiagonalization is not yet supported, so we instead play a
// trick to get the SVD of A.
DistMatrix<F> AAdj( A.Grid() );
Adjoint( A, AAdj );
svd::SingularValuesUpper( AAdj, s, heightRatio );
}
// Rescale the singular values if necessary
if( needRescaling )
Scal( 1/scale, s );
#ifndef RELEASE
PopCallStack();
#endif
}
inline int
Lyapunov( const Matrix<F>& A, const Matrix<F>& C, Matrix<F>& X )
{
#ifndef RELEASE
CallStackEntry cse("Lyapunov");
if( A.Height() != A.Width() )
LogicError("A must be square");
if( C.Height() != A.Height() || C.Width() != A.Height() )
LogicError("C must conform with A");
#endif
const Int m = A.Height();
Matrix<F> W, WTL, WTR,
WBL, WBR;
Zeros( W, 2*m, 2*m );
PartitionDownDiagonal
( W, WTL, WTR,
WBL, WBR, m );
WTL = A;
Adjoint( A, WBR ); Scale( F(-1), WBR );
WTR = C; Scale( F(-1), WTR );
return Sylvester( m, W, X );
}
inline void
QRSVD( Matrix<F>& A, Matrix<typename Base<F>::type>& s, Matrix<F>& V )
{
#ifndef RELEASE
PushCallStack("svd::QRSVD");
#endif
typedef typename Base<F>::type R;
const int m = A.Height();
const int n = A.Width();
const int k = std::min(m,n);
s.ResizeTo( k, 1 );
Matrix<F> U( m, k );
Matrix<F> VAdj( k, n );
lapack::QRSVD
( m, n, A.Buffer(), A.LDim(), s.Buffer(), U.Buffer(), U.LDim(),
VAdj.Buffer(), VAdj.LDim() );
A = U;
Adjoint( VAdj, V );
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmLLTCOld
( Orientation orientation,
UnitOrNonUnit diag,
T alpha,
const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLLTCOld");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( orientation == NORMAL )
throw std::logic_error("TrmmLLT expects a (Conjugate)Transpose option");
if( L.Height() != L.Width() || L.Height() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLTC: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,MC, STAR> L21_MC_STAR(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> D1AdjOrTrans_MR_STAR(g);
DistMatrix<T,MR, MC > D1AdjOrTrans_MR_MC(g);
DistMatrix<T,MC, MR > D1(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionDown
( X, XT,
XB, 0 );
while( XB.Height() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionDown
( XT, X0,
/**/ /**/
X1,
XB, X2 );
L21_MC_STAR.AlignWith( X2 );
D1AdjOrTrans_MR_STAR.AlignWith( X1 );
D1AdjOrTrans_MR_MC.AlignWith( X1 );
D1.AlignWith( X1 );
Zeros( X1.Width(), X1.Height(), D1AdjOrTrans_MR_STAR );
Zeros( X1.Height(), X1.Width(), D1 );
//--------------------------------------------------------------------//
X1_STAR_VR = X1;
L11_STAR_STAR = L11;
LocalTrmm
( LEFT, LOWER, orientation, diag, T(1), L11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
L21_MC_STAR = L21;
LocalGemm
( orientation, NORMAL,
T(1), X2, L21_MC_STAR, T(0), D1AdjOrTrans_MR_STAR );
D1AdjOrTrans_MR_MC.SumScatterFrom( D1AdjOrTrans_MR_STAR );
if( orientation == TRANSPOSE )
Transpose( D1AdjOrTrans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
else
Adjoint( D1AdjOrTrans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
Axpy( T(1), D1, X1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
D1AdjOrTrans_MR_MC.FreeAlignments();
D1AdjOrTrans_MR_STAR.FreeAlignments();
L21_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionDown
( XT, X0,
X1,
/**/ /**/
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void Ridge
( Orientation orientation,
const Matrix<Field>& A,
const Matrix<Field>& B,
Base<Field> gamma,
Matrix<Field>& X,
RidgeAlg alg )
{
EL_DEBUG_CSE
const bool normal = ( orientation==NORMAL );
const Int m = ( normal ? A.Height() : A.Width() );
const Int n = ( normal ? A.Width() : A.Height() );
if( orientation == TRANSPOSE && IsComplex<Field>::value )
LogicError("Transpose version of complex Ridge not yet supported");
if( m >= n )
{
Matrix<Field> Z;
if( alg == RIDGE_CHOLESKY )
{
if( orientation == NORMAL )
Herk( LOWER, ADJOINT, Base<Field>(1), A, Z );
else
Herk( LOWER, NORMAL, Base<Field>(1), A, Z );
ShiftDiagonal( Z, Field(gamma*gamma) );
Cholesky( LOWER, Z );
if( orientation == NORMAL )
Gemm( ADJOINT, NORMAL, Field(1), A, B, X );
else
Gemm( NORMAL, NORMAL, Field(1), A, B, X );
cholesky::SolveAfter( LOWER, NORMAL, Z, X );
}
else if( alg == RIDGE_QR )
{
Zeros( Z, m+n, n );
auto ZT = Z( IR(0,m), IR(0,n) );
auto ZB = Z( IR(m,m+n), IR(0,n) );
if( orientation == NORMAL )
ZT = A;
else
Adjoint( A, ZT );
FillDiagonal( ZB, Field(gamma) );
// NOTE: This QR factorization could exploit the upper-triangular
// structure of the diagonal matrix ZB
qr::ExplicitTriang( Z );
if( orientation == NORMAL )
Gemm( ADJOINT, NORMAL, Field(1), A, B, X );
else
Gemm( NORMAL, NORMAL, Field(1), A, B, X );
cholesky::SolveAfter( LOWER, NORMAL, Z, X );
}
else
{
Matrix<Field> U, V;
Matrix<Base<Field>> s;
if( orientation == NORMAL )
{
SVDCtrl<Base<Field>> ctrl;
ctrl.overwrite = false;
SVD( A, U, s, V, ctrl );
}
else
{
Matrix<Field> AAdj;
Adjoint( A, AAdj );
SVDCtrl<Base<Field>> ctrl;
ctrl.overwrite = true;
SVD( AAdj, U, s, V, ctrl );
}
auto sigmaMap =
[=]( const Base<Field>& sigma )
{ return sigma / (sigma*sigma + gamma*gamma); };
EntrywiseMap( s, MakeFunction(sigmaMap) );
Gemm( ADJOINT, NORMAL, Field(1), U, B, X );
DiagonalScale( LEFT, NORMAL, s, X );
U = X;
Gemm( NORMAL, NORMAL, Field(1), V, U, X );
}
}
else
{
LogicError("This case not yet supported");
}
}
Matrix3<T>
Matrix3<T>::Inverse( T determinant ) const
{
T det = ( (determinant == T()) ? Determinant( ) : determinant );
return Inverse( det, Adjoint( ) );
}
/**
* @brief Adjoint transform
*
* @param m To transform
* @return Transform
*/
template <class T> Matrix<T>
operator->* (const Matrix<T>& m) {
return Adjoint (m);
}
inline void
GolubReinschUpper
( DistMatrix<F>& A, DistMatrix<BASE(F),VR,STAR>& s, DistMatrix<F>& V )
{
#ifndef RELEASE
CallStackEntry entry("svd::GolubReinschUpper");
#endif
typedef BASE(F) Real;
const Int m = A.Height();
const Int n = A.Width();
const Int k = Min( m, n );
const Int offdiagonal = ( m>=n ? 1 : -1 );
const char uplo = ( m>=n ? 'U' : 'L' );
const Grid& g = A.Grid();
// Bidiagonalize A
DistMatrix<F,STAR,STAR> tP( g ), tQ( g );
Bidiag( A, tP, tQ );
// Grab copies of the diagonal and sub/super-diagonal of A
DistMatrix<Real,MD,STAR> d_MD_STAR(g), e_MD_STAR(g);
A.GetRealPartOfDiagonal( d_MD_STAR );
A.GetRealPartOfDiagonal( e_MD_STAR, offdiagonal );
// NOTE: lapack::BidiagQRAlg expects e to be of length k
DistMatrix<Real,STAR,STAR> d_STAR_STAR( d_MD_STAR ),
eHat_STAR_STAR( k, 1, g ),
e_STAR_STAR( g );
View( e_STAR_STAR, eHat_STAR_STAR, 0, 0, k-1, 1 );
e_STAR_STAR = e_MD_STAR;
// Initialize U and VAdj to the appropriate identity matrices
DistMatrix<F,VC,STAR> U_VC_STAR( g );
DistMatrix<F,STAR,VC> VAdj_STAR_VC( g );
U_VC_STAR.AlignWith( A );
VAdj_STAR_VC.AlignWith( V );
Identity( U_VC_STAR, m, k );
Identity( VAdj_STAR_VC, k, n );
// Compute the SVD of the bidiagonal matrix and accumulate the Givens
// rotations into our local portion of U and VAdj
Matrix<F>& ULoc = U_VC_STAR.Matrix();
Matrix<F>& VAdjLoc = VAdj_STAR_VC.Matrix();
lapack::BidiagQRAlg
( uplo, k, VAdjLoc.Width(), ULoc.Height(),
d_STAR_STAR.Buffer(), e_STAR_STAR.Buffer(),
VAdjLoc.Buffer(), VAdjLoc.LDim(),
ULoc.Buffer(), ULoc.LDim() );
// Make a copy of A (for the Householder vectors) and pull the necessary
// portions of U and VAdj into a standard matrix dist.
DistMatrix<F> B( A );
if( m >= n )
{
DistMatrix<F> AT(g), AB(g);
DistMatrix<F,VC,STAR> UT_VC_STAR(g), UB_VC_STAR(g);
PartitionDown( A, AT, AB, n );
PartitionDown( U_VC_STAR, UT_VC_STAR, UB_VC_STAR, n );
AT = UT_VC_STAR;
MakeZeros( AB );
Adjoint( VAdj_STAR_VC, V );
}
else
{
DistMatrix<F> VT(g), VB(g);
DistMatrix<F,STAR,VC> VAdjL_STAR_VC(g), VAdjR_STAR_VC(g);
PartitionDown( V, VT, VB, m );
PartitionRight( VAdj_STAR_VC, VAdjL_STAR_VC, VAdjR_STAR_VC, m );
Adjoint( VAdjL_STAR_VC, VT );
MakeZeros( VB );
}
// Backtransform U and V
bidiag::ApplyU( LEFT, NORMAL, B, tQ, A );
bidiag::ApplyV( LEFT, NORMAL, B, tP, V );
// Copy out the appropriate subset of the singular values
s = d_STAR_STAR;
}
Matrix3D
Matrix3D::Inverse() const
{
return (1.0f / Determinant()) * Adjoint();
}
Int ADMM
( const Matrix<Real>& A,
const Matrix<Real>& b,
const Matrix<Real>& c,
Matrix<Real>& z,
const ADMMCtrl<Real>& ctrl )
{
EL_DEBUG_CSE
// Cache a custom partially-pivoted LU factorization of
// | rho*I A^H | = | B11 B12 |
// | A 0 | | B21 B22 |
// by (justifiably) avoiding pivoting in the first n steps of
// the factorization, so that
// [I,rho*I] = lu(rho*I).
// The factorization would then proceed with
// B21 := B21 U11^{-1} = A (rho*I)^{-1} = A/rho
// B12 := L11^{-1} B12 = I A^H = A^H.
// The Schur complement would then be
// B22 := B22 - B21 B12 = 0 - (A*A^H)/rho.
// We then factor said matrix with LU with partial pivoting and
// swap the necessary rows of B21 in order to implicitly commute
// the row pivots with the Gauss transforms in the manner standard
// for GEPP. Unless A A' is singular, pivoting should not be needed,
// as Cholesky factorization of the negative matrix should be valid.
//
// The result is the factorization
// | I 0 | | rho*I A^H | = | I 0 | | rho*I U12 |,
// | 0 P22 | | A 0 | | L21 L22 | | 0 U22 |
// where [L22,U22] are stored within B22.
Matrix<Real> U12, L21, B22, bPiv;
Adjoint( A, U12 );
L21 = A;
L21 *= 1/ctrl.rho;
Herk( LOWER, NORMAL, -1/ctrl.rho, A, B22 );
MakeHermitian( LOWER, B22 );
// TODO: Replace with sparse-direct Cholesky version?
Permutation P2;
LU( B22, P2 );
P2.PermuteRows( L21 );
bPiv = b;
P2.PermuteRows( bPiv );
// Possibly form the inverse of L22 U22
Matrix<Real> X22;
if( ctrl.inv )
{
X22 = B22;
MakeTrapezoidal( LOWER, X22 );
FillDiagonal( X22, Real(1) );
TriangularInverse( LOWER, UNIT, X22 );
Trsm( LEFT, UPPER, NORMAL, NON_UNIT, Real(1), B22, X22 );
}
Int numIter=0;
const Int m = A.Height();
const Int n = A.Width();
Matrix<Real> g, xTmp, y, t;
Zeros( g, m+n, 1 );
PartitionDown( g, xTmp, y, n );
Matrix<Real> x, u, zOld, xHat;
Zeros( z, n, 1 );
Zeros( u, n, 1 );
Zeros( t, n, 1 );
while( numIter < ctrl.maxIter )
{
zOld = z;
// Find x from
// | rho*I A^H | | x | = | rho*(z-u)-c |
// | A 0 | | y | | b |
// via our cached custom factorization:
//
// |x| = inv(U) inv(L) P' |rho*(z-u)-c|
// |y| |b |
// = |rho*I U12|^{-1} |I 0 | |I 0 | |rho*(z-u)-c|
// = |0 U22| |L21 L22| |0 P22'| |b |
// = " " |rho*(z-u)-c|
// | P22' b |
xTmp = z;
xTmp -= u;
xTmp *= ctrl.rho;
xTmp -= c;
y = bPiv;
Gemv( NORMAL, Real(-1), L21, xTmp, Real(1), y );
if( ctrl.inv )
{
Gemv( NORMAL, Real(1), X22, y, t );
y = t;
}
else
{
Trsv( LOWER, NORMAL, UNIT, B22, y );
Trsv( UPPER, NORMAL, NON_UNIT, B22, y );
}
Gemv( NORMAL, Real(-1), U12, y, Real(1), xTmp );
xTmp *= 1/ctrl.rho;
// xHat := alpha*x + (1-alpha)*zOld
xHat = xTmp;
xHat *= ctrl.alpha;
Axpy( 1-ctrl.alpha, zOld, xHat );
// z := pos(xHat+u)
z = xHat;
z += u;
LowerClip( z, Real(0) );
// u := u + (xHat-z)
u += xHat;
u -= z;
const Real objective = Dot( c, xTmp );
// rNorm := || x - z ||_2
t = xTmp;
t -= z;
const Real rNorm = FrobeniusNorm( t );
// sNorm := |rho| || z - zOld ||_2
t = z;
t -= zOld;
const Real sNorm = Abs(ctrl.rho)*FrobeniusNorm( t );
const Real epsPri = Sqrt(Real(n))*ctrl.absTol +
ctrl.relTol*Max(FrobeniusNorm(xTmp),FrobeniusNorm(z));
const Real epsDual = Sqrt(Real(n))*ctrl.absTol +
ctrl.relTol*Abs(ctrl.rho)*FrobeniusNorm(u);
if( ctrl.print )
{
t = xTmp;
LowerClip( t, Real(0) );
t -= xTmp;
const Real clipDist = FrobeniusNorm( t );
cout << numIter << ": "
<< "||x-z||_2=" << rNorm << ", "
<< "epsPri=" << epsPri << ", "
<< "|rho| ||z-zOld||_2=" << sNorm << ", "
<< "epsDual=" << epsDual << ", "
<< "||x-Pos(x)||_2=" << clipDist << ", "
<< "c'x=" << objective << endl;
}
if( rNorm < epsPri && sNorm < epsDual )
break;
++numIter;
}
if( ctrl.maxIter == numIter )
cout << "ADMM failed to converge" << endl;
x = xTmp;
return numIter;
}
/**
* @brief Backward transform
*
* @param m To transform
* @return Transform
*/
virtual Matrix< std::complex<T> >
operator->* (const Matrix< std::complex<T> >& m) const {
return Adjoint (m);
}
/**
* @brief Backward transform
*
* @param m To transform
* @return Transform
*/
inline virtual Matrix<CT>
operator->* (const Matrix<CT>& m) const {
return Adjoint (m);
}
inline void
TwoSidedTrsmLVar2
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& L )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrsmLVar2");
if( A.Height() != A.Width() )
throw std::logic_error("A must be square");
if( L.Height() != L.Width() )
throw std::logic_error("Triangular matrices must be square");
if( A.Height() != L.Height() )
throw std::logic_error("A and L must be the same size");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<F>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
// Temporary distributions
DistMatrix<F,MR, STAR> A10Adj_MR_STAR(g);
DistMatrix<F,STAR,VR > A10_STAR_VR(g);
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,VC, STAR> A21_VC_STAR(g);
DistMatrix<F,MR, STAR> F10Adj_MR_STAR(g);
DistMatrix<F,MR, STAR> L10Adj_MR_STAR(g);
DistMatrix<F,VC, STAR> L10Adj_VC_STAR(g);
DistMatrix<F,STAR,MC > L10_STAR_MC(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> X11_MC_STAR(g);
DistMatrix<F,MC, STAR> X21_MC_STAR(g);
DistMatrix<F,MC, STAR> Y10Adj_MC_STAR(g);
DistMatrix<F,MR, MC > Y10Adj_MR_MC(g);
DistMatrix<F> X11(g);
DistMatrix<F> Y10Adj(g);
Matrix<F> Y10Local;
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
A10Adj_MR_STAR.AlignWith( L10 );
F10Adj_MR_STAR.AlignWith( A00 );
L10Adj_MR_STAR.AlignWith( A00 );
L10Adj_VC_STAR.AlignWith( A00 );
L10_STAR_MC.AlignWith( A00 );
X11.AlignWith( A11 );
X11_MC_STAR.AlignWith( L10 );
X21_MC_STAR.AlignWith( A20 );
Y10Adj_MC_STAR.AlignWith( A00 );
Y10Adj_MR_MC.AlignWith( A10 );
//--------------------------------------------------------------------//
// Y10 := L10 A00
L10Adj_MR_STAR.AdjointFrom( L10 );
L10Adj_VC_STAR = L10Adj_MR_STAR;
L10_STAR_MC.AdjointFrom( L10Adj_VC_STAR );
Y10Adj_MC_STAR.ResizeTo( A10.Width(), A10.Height() );
F10Adj_MR_STAR.ResizeTo( A10.Width(), A10.Height() );
Zero( Y10Adj_MC_STAR );
Zero( F10Adj_MR_STAR );
LocalSymmetricAccumulateRL
( ADJOINT,
F(1), A00, L10_STAR_MC, L10Adj_MR_STAR,
Y10Adj_MC_STAR, F10Adj_MR_STAR );
Y10Adj.SumScatterFrom( Y10Adj_MC_STAR );
Y10Adj_MR_MC = Y10Adj;
Y10Adj_MR_MC.SumScatterUpdate( F(1), F10Adj_MR_STAR );
Adjoint( Y10Adj_MR_MC.LockedLocalMatrix(), Y10Local );
// X11 := A10 L10'
X11_MC_STAR.ResizeTo( A11.Height(), A11.Width() );
LocalGemm
( NORMAL, NORMAL, F(1), A10, L10Adj_MR_STAR, F(0), X11_MC_STAR );
// A10 := A10 - Y10
Axpy( F(-1), Y10Local, A10.LocalMatrix() );
A10Adj_MR_STAR.AdjointFrom( A10 );
// A11 := A11 - (X11 + L10 A10') = A11 - (A10 L10' + L10 A10')
LocalGemm
( NORMAL, NORMAL, F(1), L10, A10Adj_MR_STAR, F(1), X11_MC_STAR );
X11.SumScatterFrom( X11_MC_STAR );
MakeTrapezoidal( LEFT, LOWER, 0, X11 );
Axpy( F(-1), X11, A11 );
// A10 := inv(L11) A10
L11_STAR_STAR = L11;
A10_STAR_VR.AdjointFrom( A10Adj_MR_STAR );
LocalTrsm
( LEFT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, A10_STAR_VR );
A10 = A10_STAR_VR;
// A11 := inv(L11) A11 inv(L11)'
A11_STAR_STAR = A11;
LocalTwoSidedTrsm( LOWER, diag, A11_STAR_STAR, L11_STAR_STAR );
A11 = A11_STAR_STAR;
// A21 := A21 - A20 L10'
X21_MC_STAR.ResizeTo( A21.Height(), A21.Width() );
LocalGemm
( NORMAL, NORMAL, F(1), A20, L10Adj_MR_STAR, F(0), X21_MC_STAR );
A21.SumScatterUpdate( F(-1), X21_MC_STAR );
// A21 := A21 inv(L11)'
A21_VC_STAR = A21;
LocalTrsm
( RIGHT, LOWER, ADJOINT, diag, F(1), L11_STAR_STAR, A21_VC_STAR );
A21 = A21_VC_STAR;
//--------------------------------------------------------------------//
A10Adj_MR_STAR.FreeAlignments();
F10Adj_MR_STAR.FreeAlignments();
L10Adj_MR_STAR.FreeAlignments();
L10Adj_VC_STAR.FreeAlignments();
L10_STAR_MC.FreeAlignments();
X11.FreeAlignments();
X11_MC_STAR.FreeAlignments();
X21_MC_STAR.FreeAlignments();
Y10Adj_MC_STAR.FreeAlignments();
Y10Adj_MR_MC.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/**********************************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
/*! compute inverse matrix */
COMPILER_FORCEINLINE const LinearSpace3 Inverse() const { return Rcp(Det())*Adjoint(); }
SVDInfo ScaLAPACKHelper
( const AbstractDistMatrix<F>& APre,
AbstractDistMatrix<F>& UPre,
AbstractDistMatrix<Base<F>>& sPre,
AbstractDistMatrix<F>& V,
const SVDCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
AssertScaLAPACKSupport();
SVDInfo info;
#ifdef EL_HAVE_SCALAPACK
typedef Base<F> Real;
DistMatrix<F,MC,MR,BLOCK> A( APre );
DistMatrixWriteProxy<Real,Real,STAR,STAR> sProx(sPre);
DistMatrixWriteProxy<F,F,MC,MR,BLOCK> UProx(UPre);
auto& s = sProx.Get();
auto& U = UProx.Get();
const int m = A.Height();
const int n = A.Width();
const int k = Min(m,n);
auto approach = ctrl.bidiagSVDCtrl.approach;
if( approach == THIN_SVD || approach == COMPACT_SVD )
{
Zeros( U, m, k );
DistMatrix<F,MC,MR,BLOCK> VH( A.Grid() );
Zeros( VH, k, n );
s.Resize( k, 1 );
auto descA = FillDesc( A );
auto descU = FillDesc( U );
auto descVH = FillDesc( VH );
scalapack::SVD
( m, n,
A.Buffer(), descA.data(),
s.Buffer(),
U.Buffer(), descU.data(),
VH.Buffer(), descVH.data() );
const bool compact = ( approach == COMPACT_SVD );
if( compact )
{
const Real twoNorm = ( k==0 ? Real(0) : s.Get(0,0) );
const Real thresh =
bidiag_svd::APosterioriThreshold
( m, n, twoNorm, ctrl.bidiagSVDCtrl );
Int rank = k;
for( Int j=0; j<k; ++j )
{
if( s.Get(j,0) <= thresh )
{
rank = j;
break;
}
}
s.Resize( rank, 1 );
U.Resize( m, rank );
VH.Resize( rank, n );
}
Adjoint( VH, V );
}
else
LogicError
("Only Thin and Compact singular value options currently supported");
#endif
return info;
}
inline void
TwoSidedTrsmUVar2
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar2");
if( A.Height() != A.Width() )
LogicError("A must be square");
if( U.Height() != U.Width() )
LogicError("Triangular matrices must be square");
if( A.Height() != U.Height() )
LogicError("A and U must be the same size");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
// Temporary distributions
DistMatrix<F,MC, STAR> A01_MC_STAR(g);
DistMatrix<F,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,MC, STAR> F01_MC_STAR(g);
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,VR, STAR> U01_VR_STAR(g);
DistMatrix<F,STAR,MR > U01Adj_STAR_MR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,MR > X11_STAR_MR(g);
DistMatrix<F,MR, STAR> X12Adj_MR_STAR(g);
DistMatrix<F,MR, MC > X12Adj_MR_MC(g);
DistMatrix<F,MR, MC > Y01_MR_MC(g);
DistMatrix<F,MR, STAR> Y01_MR_STAR(g);
DistMatrix<F> X11(g);
DistMatrix<F> Y01(g);
Matrix<F> X12Local;
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
A01_MC_STAR.AlignWith( U01 );
Y01.AlignWith( A01 );
Y01_MR_STAR.AlignWith( A00 );
U01_MC_STAR.AlignWith( A00 );
U01_VR_STAR.AlignWith( A00 );
U01Adj_STAR_MR.AlignWith( A00 );
X11_STAR_MR.AlignWith( U01 );
X11.AlignWith( A11 );
X12Adj_MR_STAR.AlignWith( A02 );
X12Adj_MR_MC.AlignWith( A12 );
F01_MC_STAR.AlignWith( A00 );
//--------------------------------------------------------------------//
// Y01 := A00 U01
U01_MC_STAR = U01;
U01_VR_STAR = U01_MC_STAR;
U01Adj_STAR_MR.AdjointFrom( U01_VR_STAR );
Zeros( Y01_MR_STAR, A01.Height(), A01.Width() );
Zeros( F01_MC_STAR, A01.Height(), A01.Width() );
LocalSymmetricAccumulateLU
( ADJOINT,
F(1), A00, U01_MC_STAR, U01Adj_STAR_MR, F01_MC_STAR, Y01_MR_STAR );
Y01_MR_MC.SumScatterFrom( Y01_MR_STAR );
Y01 = Y01_MR_MC;
Y01.SumScatterUpdate( F(1), F01_MC_STAR );
// X11 := U01' A01
LocalGemm( ADJOINT, NORMAL, F(1), U01_MC_STAR, A01, X11_STAR_MR );
// A01 := A01 - Y01
Axpy( F(-1), Y01, A01 );
A01_MC_STAR = A01;
// A11 := A11 - triu(X11 + A01' U01) = A11 - (U01 A01 + A01' U01)
LocalGemm( ADJOINT, NORMAL, F(1), A01_MC_STAR, U01, F(1), X11_STAR_MR );
X11.SumScatterFrom( X11_STAR_MR );
MakeTriangular( UPPER, X11 );
Axpy( F(-1), X11, A11 );
// A01 := A01 inv(U11)
U11_STAR_STAR = U11;
A01_VC_STAR = A01_MC_STAR;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, A01_VC_STAR );
A01 = A01_VC_STAR;
// A11 := inv(U11)' A11 inv(U11)
A11_STAR_STAR = A11;
LocalTwoSidedTrsm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
// A12 := A12 - A02' U01
LocalGemm( ADJOINT, NORMAL, F(1), A02, U01_MC_STAR, X12Adj_MR_STAR );
X12Adj_MR_MC.SumScatterFrom( X12Adj_MR_STAR );
Adjoint( X12Adj_MR_MC.LockedMatrix(), X12Local );
Axpy( F(-1), X12Local, A12.Matrix() );
// A12 := inv(U11)' A12
A12_STAR_VR = A12;
LocalTrsm
( LEFT, UPPER, ADJOINT, diag, F(1), U11_STAR_STAR, A12_STAR_VR );
A12 = A12_STAR_VR;
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
}
}
inline void
HemmRUA
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::HemmRUA");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
#endif
const Grid& g = A.Grid();
DistMatrix<T>
BT(g), B0(g),
BB(g), B1(g),
B2(g);
DistMatrix<T>
CT(g), C0(g),
CB(g), C1(g),
C2(g);
DistMatrix<T,MR, STAR> B1Adj_MR_STAR(g);
DistMatrix<T,VC, STAR> B1Adj_VC_STAR(g);
DistMatrix<T,STAR,MC > B1_STAR_MC(g);
DistMatrix<T,MC, STAR> Z1Adj_MC_STAR(g);
DistMatrix<T,MR, STAR> Z1Adj_MR_STAR(g);
DistMatrix<T,MR, MC > Z1Adj_MR_MC(g);
DistMatrix<T> Z1Adj(g);
B1Adj_MR_STAR.AlignWith( A );
B1Adj_VC_STAR.AlignWith( A );
B1_STAR_MC.AlignWith( A );
Z1Adj_MC_STAR.AlignWith( A );
Z1Adj_MR_STAR.AlignWith( A );
Matrix<T> Z1Local;
Scale( beta, C );
LockedPartitionDown
( B, BT,
BB, 0 );
PartitionDown
( C, CT,
CB, 0 );
while( CT.Height() < C.Height() )
{
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
RepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
Z1Adj_MR_MC.AlignWith( C1 );
Zeros( C1.Width(), C1.Height(), Z1Adj_MC_STAR );
Zeros( C1.Width(), C1.Height(), Z1Adj_MR_STAR );
//--------------------------------------------------------------------//
B1Adj_MR_STAR.AdjointFrom( B1 );
B1Adj_VC_STAR = B1Adj_MR_STAR;
B1_STAR_MC.AdjointFrom( B1Adj_VC_STAR );
LocalSymmetricAccumulateRU
( ADJOINT, alpha, A, B1_STAR_MC, B1Adj_MR_STAR,
Z1Adj_MC_STAR, Z1Adj_MR_STAR );
Z1Adj.SumScatterFrom( Z1Adj_MC_STAR );
Z1Adj_MR_MC = Z1Adj;
Z1Adj_MR_MC.SumScatterUpdate( T(1), Z1Adj_MR_STAR );
Adjoint( Z1Adj_MR_MC.LockedLocalMatrix(), Z1Local );
Axpy( T(1), Z1Local, C1.LocalMatrix() );
//--------------------------------------------------------------------//
Z1Adj_MR_MC.FreeAlignments();
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
internal::HegstLUVar2( DistMatrix<F,MC,MR>& A, const DistMatrix<F,MC,MR>& U )
{
#ifndef RELEASE
PushCallStack("internal::HegstLUVar2");
if( A.Height() != A.Width() )
throw std::logic_error("A must be square");
if( U.Height() != U.Width() )
throw std::logic_error("Triangular matrices must be square");
if( A.Height() != U.Height() )
throw std::logic_error("A and U must be the same size");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F,MC,MR>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<F,MC,MR>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
// Temporary distributions
DistMatrix<F,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,MC > U12_STAR_MC(g);
DistMatrix<F,STAR,VR > U12_STAR_VR(g);
DistMatrix<F,MR, STAR> U12Adj_MR_STAR(g);
DistMatrix<F,VC, STAR> U12Adj_VC_STAR(g);
DistMatrix<F,MC, STAR> X01_MC_STAR(g);
DistMatrix<F,STAR,STAR> X11_STAR_STAR(g);
DistMatrix<F,MC, MR > Y12(g);
DistMatrix<F,MC, MR > Z12Adj(g);
DistMatrix<F,MR, MC > Z12Adj_MR_MC(g);
DistMatrix<F,MC, STAR> Z12Adj_MC_STAR(g);
DistMatrix<F,MR, STAR> Z12Adj_MR_STAR(g);
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
A12_STAR_VR.AlignWith( A12 );
U12_STAR_MC.AlignWith( A22 );
U12_STAR_VR.AlignWith( A12 );
U12Adj_MR_STAR.AlignWith( A22 );
U12Adj_VC_STAR.AlignWith( A22 );
X01_MC_STAR.AlignWith( A01 );
Y12.AlignWith( A12 );
Z12Adj.AlignWith( A12 );
Z12Adj_MR_MC.AlignWith( A12 );
Z12Adj_MC_STAR.AlignWith( A22 );
Z12Adj_MR_STAR.AlignWith( A22 );
//--------------------------------------------------------------------//
// A01 := A01 U11'
U11_STAR_STAR = U11;
A01_VC_STAR = A01;
internal::LocalTrmm
( RIGHT, UPPER, ADJOINT, NON_UNIT, (F)1, U11_STAR_STAR, A01_VC_STAR );
A01 = A01_VC_STAR;
// A01 := A01 + A02 U12'
U12Adj_MR_STAR.AdjointFrom( U12 );
X01_MC_STAR.ResizeTo( A01.Height(), A01.Width() );
internal::LocalGemm
( NORMAL, NORMAL,
(F)1, A02, U12Adj_MR_STAR, (F)0, X01_MC_STAR );
A01.SumScatterUpdate( (F)1, X01_MC_STAR );
// Y12 := U12 A22
U12Adj_VC_STAR = U12Adj_MR_STAR;
U12_STAR_MC.AdjointFrom( U12Adj_VC_STAR );
Z12Adj_MC_STAR.ResizeTo( A12.Width(), A12.Height() );
Z12Adj_MR_STAR.ResizeTo( A12.Width(), A12.Height() );
Zero( Z12Adj_MC_STAR );
Zero( Z12Adj_MR_STAR );
internal::LocalSymmetricAccumulateRU
( ADJOINT,
(F)1, A22, U12_STAR_MC, U12Adj_MR_STAR,
Z12Adj_MC_STAR, Z12Adj_MR_STAR );
Z12Adj.SumScatterFrom( Z12Adj_MC_STAR );
Z12Adj_MR_MC = Z12Adj;
Z12Adj_MR_MC.SumScatterUpdate( (F)1, Z12Adj_MR_STAR );
Y12.ResizeTo( A12.Height(), A12.Width() );
Adjoint( Z12Adj_MR_MC.LockedLocalMatrix(), Y12.LocalMatrix() );
// A12 := U11 A12
A12_STAR_VR = A12;
U11_STAR_STAR = U11;
internal::LocalTrmm
( LEFT, UPPER, NORMAL, NON_UNIT, (F)1, U11_STAR_STAR, A12_STAR_VR );
A12 = A12_STAR_VR;
// A12 := A12 + 1/2 Y12
Axpy( (F)0.5, Y12, A12 );
// A11 := U11 A11 U11'
A11_STAR_STAR = A11;
internal::LocalHegst( LEFT, UPPER, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
// A11 := A11 + (A12 U12' + U12 A12')
A12_STAR_VR = A12;
U12_STAR_VR = U12;
X11_STAR_STAR.ResizeTo( A11.Height(), A11.Width() );
Her2k
( UPPER, NORMAL,
(F)1, A12_STAR_VR.LocalMatrix(), U12_STAR_VR.LocalMatrix(),
(F)0, X11_STAR_STAR.LocalMatrix() );
A11.SumScatterUpdate( (F)1, X11_STAR_STAR );
// A12 := A12 + 1/2 Y12
Axpy( (F)0.5, Y12, A12 );
//--------------------------------------------------------------------//
A12_STAR_VR.FreeAlignments();
U12_STAR_MC.FreeAlignments();
U12_STAR_VR.FreeAlignments();
U12Adj_MR_STAR.FreeAlignments();
U12Adj_VC_STAR.FreeAlignments();
X01_MC_STAR.FreeAlignments();
Y12.FreeAlignments();
Z12Adj.FreeAlignments();
Z12Adj_MR_MC.FreeAlignments();
Z12Adj_MC_STAR.FreeAlignments();
Z12Adj_MR_STAR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
