inline void
HermitianFromEVD
( UpperOrLower uplo,
Matrix<F>& A,
const Matrix<BASE(F)>& w,
const Matrix<F>& Z )
{
DEBUG_ONLY(CallStackEntry cse("HermitianFromEVD"))
Matrix<F> Z1Copy, Y1;
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 = LockedView( Z, 0, k, m, nb );
auto w1 = LockedView( w, k, 0, nb, 1 );
Y1 = Z1Copy = Z1;
DiagonalScale( RIGHT, NORMAL, w1, Y1 );
Trrk( uplo, NORMAL, ADJOINT, F(1), Z1Copy, Y1, F(1), A );
}
}
void HermitianFromEVD
( UpperOrLower uplo,
Matrix<F>& A,
const Matrix<Base<F>>& w,
const Matrix<F>& Z )
{
DEBUG_CSE
Matrix<F> Z1Copy, Y1;
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 );
Y1 = Z1Copy = Z1;
DiagonalScale( RIGHT, NORMAL, w1, Y1 );
Trrk( uplo, NORMAL, ADJOINT, F(1), Z1Copy, Y1, F(1), A );
}
}
inline void
HPSDCholesky( UpperOrLower uplo, DistMatrix<Complex<R>,MC,MR>& A )
{
#ifndef RELEASE
PushCallStack("HPSDCholesky");
#endif
HPSDSquareRoot( uplo, A );
hpsd_cholesky::MakeExplicitlyHermitian( uplo, A );
const Grid& g = A.Grid();
if( uplo == LOWER )
{
DistMatrix<Complex<R>,MD,STAR> t(g);
LQ( A, t );
MakeTrapezoidal( LEFT, LOWER, 0, A );
}
else
{
DistMatrix<Complex<R>,MD,STAR> t(g);
QR( A, t );
MakeTrapezoidal( RIGHT, UPPER, 0, A );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void MakeExplicitlyHermitian( UpperOrLower uplo, DistMatrix<F,MC,MR>& A )
{
const Grid& g = A.Grid();
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> A11Adj(g);
DistMatrix<F,MR,MC> A11_MR_MC(g);
DistMatrix<F,MR,MC> A21_MR_MC(g);
DistMatrix<F,MR,MC> A12_MR_MC(g);
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
A11Adj.AlignWith( A11 );
A11_MR_MC.AlignWith( A11 );
A12_MR_MC.AlignWith( A21 );
A21_MR_MC.AlignWith( A12 );
//--------------------------------------------------------------------//
A11_MR_MC = A11;
A11Adj.ResizeTo( A11.Height(), A11.Width() );
Adjoint( A11_MR_MC.LocalMatrix(), A11Adj.LocalMatrix() );
if( uplo == LOWER )
{
MakeTrapezoidal( LEFT, UPPER, 1, A11Adj );
Axpy( (F)1, A11Adj, A11 );
A21_MR_MC = A21;
Adjoint( A21_MR_MC.LocalMatrix(), A12.LocalMatrix() );
}
else
{
MakeTrapezoidal( LEFT, LOWER, -1, A11Adj );
Axpy( (F)1, A11Adj, A11 );
A12_MR_MC = A12;
Adjoint( A12_MR_MC.LocalMatrix(), A21.LocalMatrix() );
}
//--------------------------------------------------------------------//
A21_MR_MC.FreeAlignments();
A12_MR_MC.FreeAlignments();
A11_MR_MC.FreeAlignments();
A11Adj.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
}
inline typename Base<F>::type
LogDetDivergence
( UpperOrLower uplo, const DistMatrix<F>& A, const DistMatrix<F>& B )
{
#ifndef RELEASE
PushCallStack("LogDetDivergence");
#endif
if( A.Grid() != B.Grid() )
throw std::logic_error("A and B must use the same grid");
if( A.Height() != A.Width() || B.Height() != B.Width() ||
A.Height() != B.Height() )
throw std::logic_error
("A and B must be square matrices of the same size");
typedef typename Base<F>::type R;
const int n = A.Height();
const Grid& g = A.Grid();
DistMatrix<F> ACopy( A );
DistMatrix<F> BCopy( B );
Cholesky( uplo, ACopy );
Cholesky( uplo, BCopy );
if( uplo == LOWER )
{
Trtrsm( LEFT, uplo, NORMAL, NON_UNIT, F(1), BCopy, ACopy );
}
else
{
MakeTrapezoidal( LEFT, uplo, 0, ACopy );
Trsm( LEFT, uplo, NORMAL, NON_UNIT, F(1), BCopy, ACopy );
}
MakeTrapezoidal( LEFT, uplo, 0, ACopy );
const R frobNorm = Norm( ACopy, FROBENIUS_NORM );
R logDet;
R localLogDet(0);
DistMatrix<F,MD,STAR> d(g);
ACopy.GetDiagonal( d );
if( d.InDiagonal() )
{
const int nLocalDiag = d.LocalHeight();
for( int iLocal=0; iLocal<nLocalDiag; ++iLocal )
{
const R delta = RealPart(d.GetLocal(iLocal,0));
localLogDet += 2*Log(delta);
}
}
mpi::AllReduce( &localLogDet, &logDet, 1, mpi::SUM, g.VCComm() );
const R logDetDiv = frobNorm*frobNorm - logDet - R(n);
#ifndef RELEASE
PopCallStack();
#endif
return logDetDiv;
}
inline void
HermitianFromEVD
( UpperOrLower uplo,
Matrix<F>& A,
const Matrix<BASE(F)>& w,
const Matrix<F>& Z )
{
#ifndef RELEASE
CallStackEntry entry("HermitianFromEVD");
#endif
typedef BASE(F) R;
Matrix<F> ZL, ZR,
Z0, Z1, Z2;
Matrix<R> wT, w0,
wB, w1,
w2;
Matrix<F> Z1Copy, Y1;
A.ResizeTo( Z.Height(), Z.Height() );
if( uplo == LOWER )
MakeTrapezoidal( UPPER, A, 1 );
else
MakeTrapezoidal( LOWER, A, -1 );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
//--------------------------------------------------------------------//
Y1 = Z1Copy = Z1;
DiagonalScale( RIGHT, NORMAL, w1, Y1 );
Trrk( uplo, NORMAL, ADJOINT, F(1), Z1Copy, Y1, F(1), A );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
}
void Explicit( AbstractDistMatrix<F>& L, AbstractDistMatrix<F>& APre )
{
EL_DEBUG_CSE
const Grid& g = APre.Grid();
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
DistMatrix<F,MD,STAR> householderScalars(g);
DistMatrix<Base<F>,MD,STAR> signature(g);
LQ( A, householderScalars, signature );
const Int m = A.Height();
const Int n = A.Width();
const Int minDim = Min(m,n);
auto AL = A( IR(0,m), IR(0,minDim) );
Copy( AL, L );
MakeTrapezoidal( LOWER, L );
// TODO: Replace this with an in-place expansion of Q
DistMatrix<F> Q(g);
Identity( Q, A.Height(), A.Width() );
lq::ApplyQ( RIGHT, NORMAL, A, householderScalars, signature, Q );
Copy( Q, APre );
}
inline void
SingularValuesUpper
( DistMatrix<Complex<Real> >& A,
DistMatrix<Real,VR,STAR>& s,
double heightRatio=1.2 )
{
#ifndef RELEASE
PushCallStack("svd::SingularValuesUpper");
if( heightRatio <= 1.0 )
throw std::logic_error("Nonsensical switchpoint for SingularValues");
#endif
typedef Complex<Real> C;
const Grid& g = A.Grid();
const int m = A.Height();
const int n = A.Width();
if( m >= heightRatio*n )
{
DistMatrix<C,MD,STAR> t(g);
QR( A, t );
DistMatrix<C> AT(g),
AB(g);
PartitionDown
( A, AT,
AB, n );
MakeTrapezoidal( LEFT, UPPER, 0, AT );
SimpleSingularValuesUpper( AT, s );
}
else
{
SimpleSingularValuesUpper( A, s );
}
#ifndef RELEASE
PopCallStack();
#endif
}
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 );
}
}
inline void
RLHF( int offset, const Matrix<R>& H, Matrix<R>& A )
{
#ifndef RELEASE
CallStackEntry entry("apply_packed_reflectors::RLHF");
if( offset > 0 || offset < -H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Width() )
throw std::logic_error
("Width of transforms must equal width of target matrix");
#endif
Matrix<R>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<R> ALeft;
Matrix<R> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
const int HPanWidth = H10.Width() + H11.Width();
const int HPanOffset =
std::min( H11.Height(), std::max(-offset-H00.Height(),0) );
const int HPanHeight = H11.Height()-HPanOffset;
LockedView
( HPan, H, H00.Height()+HPanOffset, 0, HPanHeight, HPanWidth );
View( ALeft, A, 0, 0, A.Height(), HPanWidth );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, LOWER, offset, HPanCopy );
SetDiagonal( RIGHT, offset, HPanCopy, R(1) );
Syrk( UPPER, NORMAL, R(1), HPanCopy, SInv );
HalveMainDiagonal( SInv );
Gemm( NORMAL, TRANSPOSE, R(1), ALeft, HPanCopy, Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, R(1), SInv, Z );
Gemm( NORMAL, NORMAL, R(-1), Z, HPanCopy, R(1), ALeft );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
}
}
void ExplicitTriang( Matrix<F>& A )
{
DEBUG_ONLY(CSE cse("rq::ExplicitTriang"))
Matrix<F> t;
Matrix<Base<F>> d;
Householder( A, t, d );
MakeTrapezoidal( UPPER, A, A.Width()-A.Height() );
}
void ExplicitTriang( ElementalMatrix<F>& A )
{
DEBUG_ONLY(CSE cse("rq::ExplicitTriang"))
DistMatrix<F,MD,STAR> t(A.Grid());
DistMatrix<Base<F>,MD,STAR> d(A.Grid());
Householder( A, t, d );
MakeTrapezoidal( UPPER, A, A.Width()-A.Height() );
}
void ExplicitTriang( AbstractDistMatrix<F>& A )
{
DEBUG_ONLY(CallStackEntry cse("rq::ExplicitTriang"))
DistMatrix<F,MD,STAR> t(A.Grid());
DistMatrix<Base<F>,MD,STAR> d(A.Grid());
Householder( A, t, d );
MakeTrapezoidal( UPPER, A, A.Width()-A.Height() );
}
inline typename Base<F>::type
LogDetDivergence( UpperOrLower uplo, const Matrix<F>& A, const Matrix<F>& B )
{
#ifndef RELEASE
PushCallStack("LogDetDivergence");
#endif
if( A.Height() != A.Width() || B.Height() != B.Width() ||
A.Height() != B.Height() )
throw std::logic_error
("A and B must be square matrices of the same size");
typedef typename Base<F>::type R;
const int n = A.Height();
Matrix<F> ACopy( A );
Matrix<F> BCopy( B );
Cholesky( uplo, ACopy );
Cholesky( uplo, BCopy );
if( uplo == LOWER )
{
Trtrsm( LEFT, uplo, NORMAL, NON_UNIT, F(1), BCopy, ACopy );
}
else
{
MakeTrapezoidal( LEFT, uplo, 0, ACopy );
Trsm( LEFT, uplo, NORMAL, NON_UNIT, F(1), BCopy, ACopy );
}
MakeTrapezoidal( LEFT, uplo, 0, ACopy );
const R frobNorm = Norm( ACopy, FROBENIUS_NORM );
Matrix<F> d;
ACopy.GetDiagonal( d );
R logDet(0);
for( int i=0; i<n; ++i )
logDet += 2*Log( RealPart(d.Get(i,0)) );
const R logDetDiv = frobNorm*frobNorm - logDet - R(n);
#ifndef RELEASE
PopCallStack();
#endif
return logDetDiv;
}
void ExplicitTriang( Matrix<F>& A )
{
DEBUG_ONLY(CallStackEntry cse("lq::ExplicitTriang"))
Matrix<F> t;
Matrix<Base<F>> d;
LQ( A, t, d );
const Int m = A.Height();
const Int n = A.Width();
const Int minDim = Min(m,n);
A.Resize( m, minDim );
MakeTrapezoidal( LOWER, A );
}
void ExplicitTriang( Matrix<F>& A )
{
EL_DEBUG_CSE
Matrix<F> householderScalars;
Matrix<Base<F>> signature;
LQ( A, householderScalars, signature );
const Int m = A.Height();
const Int n = A.Width();
const Int minDim = Min(m,n);
A.Resize( m, minDim );
MakeTrapezoidal( LOWER, A );
}
inline void
ExplicitLQHelper( Matrix<Real>& L, Matrix<Real>& A )
{
LQ( A );
L = A;
MakeTrapezoidal( LEFT, LOWER, 0, L );
// TODO: Replace this with an in-place expansion of Q
Matrix<Real> Q;
Identity( A.Height(), A.Width(), Q );
ApplyPackedReflectors( RIGHT, UPPER, HORIZONTAL, BACKWARD, 0, A, Q );
A = Q;
}
void ExplicitTriang( AbstractDistMatrix<F>& A )
{
EL_DEBUG_CSE
const Grid& g = A.Grid();
DistMatrix<F,MD,STAR> householderScalars(g);
DistMatrix<Base<F>,MD,STAR> signature(g);
LQ( A, householderScalars, signature );
const Int m = A.Height();
const Int n = A.Width();
const Int minDim = Min(m,n);
A.Resize( m, minDim );
MakeTrapezoidal( LOWER, A );
}
inline void
HPSDCholesky( UpperOrLower uplo, DistMatrix<R,MC,MR>& A )
{
#ifndef RELEASE
PushCallStack("HPSDCholesky");
#endif
HPSDSquareRoot( uplo, A );
hpsd_cholesky::MakeExplicitlyHermitian( uplo, A );
if( uplo == LOWER )
{
LQ( A );
MakeTrapezoidal( LEFT, LOWER, 0, A );
}
else
{
QR( A );
MakeTrapezoidal( RIGHT, UPPER, 0, A );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void ExplicitTriang( Matrix<F>& A, const QRCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
Matrix<F> householderScalars;
Matrix<Base<F>> signature;
if( ctrl.colPiv )
{
Permutation Omega;
BusingerGolub( A, householderScalars, signature, Omega, ctrl );
}
else
Householder( A, householderScalars, signature );
A.Resize( householderScalars.Height(), A.Width() );
MakeTrapezoidal( UPPER, A );
}
void ExplicitTriang( Matrix<F>& A, const QRCtrl<Base<F>>& ctrl )
{
DEBUG_ONLY(CSE cse("qr::ExplicitTriang"))
Matrix<F> t;
Matrix<Base<F>> d;
if( ctrl.colPiv )
{
Permutation Omega;
BusingerGolub( A, t, d, Omega, ctrl );
}
else
Householder( A, t, d );
A.Resize( t.Height(), A.Width() );
MakeTrapezoidal( UPPER, A );
}
void ExplicitTriang( AbstractDistMatrix<F>& A, const QRCtrl<Base<F>>& ctrl )
{
DEBUG_ONLY(CallStackEntry cse("qr::ExplicitTriang"))
DistMatrix<F,MD,STAR> t(A.Grid());
DistMatrix<Base<F>,MD,STAR> d(A.Grid());
if( ctrl.colPiv )
{
DistMatrix<Int,VC,STAR> p(A.Grid());
BusingerGolub( A, t, d, p, ctrl );
}
else
Householder( A, t, d );
A.Resize( t.Height(), A.Width() );
MakeTrapezoidal( UPPER, A );
}
void ExplicitTriang( Matrix<F>& A, const QRCtrl<Base<F>>& ctrl )
{
DEBUG_ONLY(CallStackEntry cse("qr::ExplicitTriang"))
Matrix<F> t;
Matrix<Base<F>> d;
if( ctrl.colPiv )
{
Matrix<Int> p;
BusingerGolub( A, t, d, p, ctrl );
}
else
Householder( A, t, d );
A.Resize( t.Height(), A.Width() );
MakeTrapezoidal( UPPER, A );
}
void ExplicitTriang( ElementalMatrix<F>& A, const QRCtrl<Base<F>>& ctrl )
{
DEBUG_ONLY(CSE cse("qr::ExplicitTriang"))
DistMatrix<F,MD,STAR> t(A.Grid());
DistMatrix<Base<F>,MD,STAR> d(A.Grid());
if( ctrl.colPiv )
{
DistPermutation Omega(A.Grid());
BusingerGolub( A, t, d, Omega, ctrl );
}
else
Householder( A, t, d );
A.Resize( t.Height(), A.Width() );
MakeTrapezoidal( UPPER, A );
}
void ExplicitTriang( ElementalMatrix<F>& A, const QRCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
DistMatrix<F,MD,STAR> householderScalars(A.Grid());
DistMatrix<Base<F>,MD,STAR> signature(A.Grid());
if( ctrl.colPiv )
{
DistPermutation Omega(A.Grid());
BusingerGolub( A, householderScalars, signature, Omega, ctrl );
}
else
Householder( A, householderScalars, signature );
A.Resize( householderScalars.Height(), A.Width() );
MakeTrapezoidal( UPPER, A );
}
inline void
ExplicitLQHelper
( DistMatrix<Complex<Real> >& L, DistMatrix<Complex<Real> >& A )
{
const Grid& g = A.Grid();
DistMatrix<Complex<Real>,MD,STAR> t( g );
LQ( A, t );
L = A;
MakeTrapezoidal( LEFT, LOWER, 0, L );
// TODO: Replace this with an in-place expansion of Q
DistMatrix<Complex<Real> > Q( g );
Identity( A.Height(), A.Width(), Q );
ApplyPackedReflectors
( RIGHT, UPPER, HORIZONTAL, BACKWARD, UNCONJUGATED, 0, A, t, Q );
A = Q;
}
void Explicit
( ElementalMatrix<F>& APre,
ElementalMatrix<F>& R,
bool thinQR,
const QRCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
const Grid& g = A.Grid();
DistMatrix<F,MD,STAR> householderScalars(g);
DistMatrix<Base<F>,MD,STAR> signature(g);
if( ctrl.colPiv )
{
DistPermutation Omega(g);
QR( A, householderScalars, signature, Omega, ctrl );
}
else
QR( A, householderScalars, signature );
const Int m = A.Height();
const Int n = A.Width();
const Int numIts = householderScalars.Height();
auto AT = A( IR(0,numIts), IR(0,n) );
Copy( AT, R );
MakeTrapezoidal( UPPER, R );
if( thinQR )
{
A.Resize( m, numIts );
ExpandPackedReflectors
( LOWER, VERTICAL, CONJUGATED, 0, A, householderScalars );
DiagonalScale( RIGHT, NORMAL, signature, A );
}
else
{
auto ACopy = A;
// TODO: Use an extension of ExpandPackedReflectors to make this faster
Identity( A, A.Height(), A.Height() );
qr::ApplyQ( LEFT, NORMAL, ACopy, householderScalars, signature, A );
}
}
void Explicit
( ElementalMatrix<F>& APre,
ElementalMatrix<F>& R,
ElementalMatrix<Int>& OmegaFull,
bool thinQR,
const QRCtrl<Base<F>>& ctrl )
{
DEBUG_ONLY(CSE cse("qr::Explicit"))
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
const Grid& g = A.Grid();
DistMatrix<F,MD,STAR> t(g);
DistMatrix<Base<F>,MD,STAR> d(g);
DistPermutation Omega(g);
QR( A, t, d, Omega, ctrl );
const Int m = A.Height();
const Int n = A.Width();
const Int numIts = t.Height();
auto AT = A( IR(0,numIts), IR(0,n) );
Copy( AT, R );
MakeTrapezoidal( UPPER, R );
if( thinQR )
{
A.Resize( m, numIts );
ExpandPackedReflectors( LOWER, VERTICAL, CONJUGATED, 0, A, t );
DiagonalScale( RIGHT, NORMAL, d, A );
}
else
{
auto ACopy = A;
// TODO: Use an extension of ExpandPackedReflectors to make this faster
Identity( A, A.Height(), A.Height() );
qr::ApplyQ( LEFT, NORMAL, ACopy, t, d, A );
}
Omega.ExplicitMatrix( OmegaFull );
}
void Explicit( Matrix<F>& L, Matrix<F>& A )
{
EL_DEBUG_CSE
Matrix<F> householderScalars;
Matrix<Base<F>> signature;
LQ( A, householderScalars, signature );
const Int m = A.Height();
const Int n = A.Width();
const Int minDim = Min(m,n);
auto AL = A( IR(0,m), IR(0,minDim) );
L = AL;
MakeTrapezoidal( LOWER, L );
// TODO: Replace this with an in-place expansion of Q
Matrix<F> Q;
Identity( Q, A.Height(), A.Width() );
lq::ApplyQ( RIGHT, NORMAL, A, householderScalars, signature, Q );
A = Q;
}
void Explicit
( Matrix<F>& A,
Matrix<F>& R,
bool thinQR,
const QRCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
Matrix<F> householderScalars;
Matrix<Base<F>> signature;
if( ctrl.colPiv )
{
Permutation Omega;
QR( A, householderScalars, signature, Omega, ctrl );
}
else
QR( A, householderScalars, signature );
const Int m = A.Height();
const Int n = A.Width();
const Int numIts = householderScalars.Height();
auto AT = A( IR(0,numIts), IR(0,n) );
R = AT;
MakeTrapezoidal( UPPER, R );
if( thinQR )
{
A.Resize( m, numIts );
ExpandPackedReflectors
( LOWER, VERTICAL, CONJUGATED, 0, A, householderScalars );
DiagonalScale( RIGHT, NORMAL, signature, A );
}
else
{
auto ACopy = A;
// TODO: Use an extension of ExpandPackedReflectors to make this faster
Identity( A, A.Height(), A.Height() );
qr::ApplyQ( LEFT, NORMAL, ACopy, householderScalars, signature, A );
}
}
