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 );
}
}
inline void
LockedPartitionRight( const DM& A, DM& AL, DM& AR, Int widthAL )
{
#ifndef RELEASE
PushCallStack("LockedPartitionRight [DistMatrix]");
if( widthAL < 0 )
throw std::logic_error("Width of left partition must be non-negative");
#endif
widthAL = std::min(widthAL,A.Width());
const Int widthAR = A.Width()-widthAL;
LockedView( AL, A, 0, 0, A.Height(), widthAL );
LockedView( AR, A, 0, widthAL, A.Height(), widthAR );
#ifndef RELEASE
PopCallStack();
#endif
}
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 );
}
}
inline void
LockedRepartitionRight
( const DM& AL, const DM& AR,
DM& A0, DM& A1, DM& A2, Int A1Width=Blocksize() )
{
DEBUG_ONLY(CallStackEntry cse("LockedRepartitionRight"))
LockedView( A0, AL );
LockedPartitionRight( AR, A1, A2, A1Width );
}
inline void
LockedRepartitionLeft
( const M& AL, const M& AR,
M& A0, M& A1, M& A2, Int A1Width=Blocksize() )
{
DEBUG_ONLY(CallStackEntry cse("LockedRepartitionLeft"))
LockedPartitionLeft( AL, A0, A1, A1Width );
LockedView( A2, AR );
}
inline void
SlideLockedPartitionRight
( DM& AL, DM& AR,
const DM& A0, const DM& A1, const DM& A2 )
{
DEBUG_ONLY(CallStackEntry cse("SlideLockedPartitionRight"))
LockedView1x2( AL, A0, A1 );
LockedView( AR, A2 );
}
inline void
LockedPartitionDown
( const DM& A, DM& AT,
DM& AB, Int heightAT )
{
#ifndef RELEASE
PushCallStack("LockedPartitionDown [DistMatrix]");
if( heightAT < 0 )
throw std::logic_error("Height of top partition must be non-negative");
#endif
heightAT = std::min(heightAT,A.Height());
const Int heightAB = A.Height()-heightAT;
LockedView( AT, A, 0, 0, heightAT, A.Width() );
LockedView( AB, A, heightAT, 0, heightAB, A.Width() );
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
SlideLockedPartitionLeft
( M& AL, M& AR,
const M& A0, const M& A1, const M& A2 )
{
DEBUG_ONLY(CallStackEntry cse("SlideLockedPartitionLeft"))
LockedView( AL, A0 );
LockedView1x2( AR, A1, A2 );
}
inline void
SlideLockedPartitionUp
( DM& AT, const DM& A0,
const DM& A1,
DM& AB, const DM& A2 )
{
DEBUG_ONLY(CallStackEntry cse("SlideLockedPartitionUp"))
LockedView( AT, A0 );
LockedView2x1( AB, A1, A2 );
}
inline void
SlideLockedPartitionDown
( M& AT, const M& A0,
const M& A1,
M& AB, const M& A2 )
{
DEBUG_ONLY(CallStackEntry cse("SlideLockedPartitionDown"))
LockedView2x1( AT, A0, A1 );
LockedView( AB, A2 );
}
inline void
LockedPartitionUp
( const M& A, M& AT,
M& AB, Int heightAB )
{
#ifndef RELEASE
PushCallStack("LockedPartitionUp [Matrix]");
if( heightAB < 0 )
throw std::logic_error
("Height of bottom partition must be non-negative");
#endif
heightAB = std::min(heightAB,A.Height());
const Int heightAT = A.Height()-heightAB;
LockedView( AT, A, 0, 0, heightAT, A.Width() );
LockedView( AB, A, heightAT, 0, heightAB, A.Width() );
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LockedRepartitionDown
( const M& AT, M& A0,
M& A1,
const M& AB, M& A2, Int A1Height=Blocksize() )
{
DEBUG_ONLY(CallStackEntry cse("LockedRepartitionDown"))
LockedView( A0, AT );
LockedPartitionDown( AB, A1, A2, A1Height );
}
inline void
LockedRepartitionUp
( const DM& AT, DM& A0,
DM& A1,
const DM& AB, DM& A2, Int A1Height=Blocksize() )
{
DEBUG_ONLY(CallStackEntry cse("LockedRepartitionUp"))
LockedPartitionUp( AT, A0, A1, A1Height );
LockedView( A2, AB );
}
inline void
LockedRepartitionRight
( const DM& AL, const DM& AR,
DM& A0, DM& A1, DM& A2, Int A1Width )
{
#ifndef RELEASE
CallStackEntry cse("LockedRepartitionRight [DistMatrix]");
#endif
LockedView( A0, AL );
LockedPartitionRight( AR, A1, A2, A1Width );
}
inline void
LockedRepartitionLeft
( const M& AL, const M& AR,
M& A0, M& A1, M& A2, Int A1Width )
{
#ifndef RELEASE
CallStackEntry cse("LockedRepartitionLeft [Matrix]");
#endif
LockedPartitionLeft( AL, A0, A1, A1Width );
LockedView( A2, AR );
}
inline void
LockedRepartitionUp
( const DM& AT, DM& A0,
DM& A1,
const DM& AB, DM& A2, Int A1Height )
{
#ifndef RELEASE
CallStackEntry cse("LockedRepartitionUp [DistMatrix]");
#endif
LockedPartitionUp( AT, A0, A1, A1Height );
LockedView( A2, AB );
}
inline void
LockedRepartitionDown
( const M& AT, M& A0,
M& A1,
const M& AB, M& A2, Int A1Height )
{
#ifndef RELEASE
CallStackEntry cse("LockedRepartitionDown [Matrix]");
#endif
LockedView( A0, AT );
LockedPartitionDown( AB, A1, A2, A1Height );
}
inline void
SlideLockedPartitionDownDiagonal
( DM& ATL, DM& ATR, const DM& A00, const DM& A01, const DM& A02,
const DM& A10, const DM& A11, const DM& A12,
DM& ABL, DM& ABR, const DM& A20, const DM& A21, const DM& A22 )
{
DEBUG_ONLY(CallStackEntry cse("SlideLockedPartitionDownDiagonal"))
LockedView2x2( ATL, A00, A01,
A10, A11 );
LockedView2x1( ATR, A02, A12 );
LockedView1x2( ABL, A20, A21 );
LockedView( ABR, A22 );
}
inline void
LockedRepartitionDownDiagonal
( const M& ATL, const M& ATR, M& A00, M& A01, M& A02,
M& A10, M& A11, M& A12,
const M& ABL, const M& ABR, M& A20, M& A21, M& A22, Int bsize=Blocksize() )
{
DEBUG_ONLY(CallStackEntry cse("LockedRepartitionDownDiagonal"))
LockedView( A00, ATL );
LockedPartitionDownDiagonal( ABR, A11, A12,
A21, A22, bsize );
LockedPartitionDown( ABL, A10, A20, A11.Height() );
LockedPartitionRight( ATR, A01, A02, A11.Width() );
}
inline void
LockedPartitionDownLeftDiagonal
( const DM& A, DM& ATL, DM& ATR,
DM& ABL, DM& ABR, Int diagATL )
{
#ifndef RELEASE
PushCallStack("LockedPartitionDownLeftDiagonal [DistMatrix]");
if( diagATL < 0 )
throw std::logic_error("Top-left size must be non-negative");
#endif
const Int minDim = std::min(A.Height(),A.Width());
diagATL = std::min(diagATL,minDim);
const Int heightABR = A.Height()-diagATL;
const Int widthABR = A.Width()-diagATL;
LockedView( ATL, A, 0, 0, diagATL, diagATL );
LockedView( ATR, A, 0, diagATL, diagATL, widthABR );
LockedView( ABL, A, diagATL, 0, heightABR, diagATL );
LockedView( ABR, A, diagATL, diagATL, heightABR, widthABR );
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LockedPartitionUpRightDiagonal
( const DM& A, DM& ATL, DM& ATR,
DM& ABL, DM& ABR, Int diagABR )
{
#ifndef RELEASE
PushCallStack("LockedPartitionUpRightDiagonal [DistMatrix]");
if( diagABR < 0 )
throw std::logic_error("Bottom-right size must be non-negative");
#endif
const Int minDim = std::min(A.Height(),A.Width());
diagABR = std::min(diagABR,minDim);
const Int remHeight = A.Height()-diagABR;
const Int remWidth = A.Width()-diagABR;
LockedView( ATL, A, 0, 0, remHeight, remWidth );
LockedView( ATR, A, 0, remWidth, remHeight, diagABR );
LockedView( ABL, A, remHeight, 0, diagABR, remWidth );
LockedView( ABR, A, remHeight, remWidth, diagABR, diagABR );
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LockedPartitionDownRightDiagonal
( const M& A, M& ATL, M& ATR,
M& ABL, M& ABR, Int diagATL )
{
#ifndef RELEASE
PushCallStack("LockedPartitionDownRightDiagonal [Matrix]");
if( diagATL < 0 )
throw std::logic_error("Top-left size must be non-negative");
#endif
const Int minDim = std::min( A.Height(), A.Width() );
diagATL = std::min(diagATL,minDim);
const Int sizeABR = minDim-diagATL;
const Int remHeight = A.Height()-sizeABR;
const Int remWidth = A.Width()-sizeABR;
LockedView( ATL, A, 0, 0, remHeight, remWidth );
LockedView( ATR, A, 0, remWidth, remHeight, sizeABR );
LockedView( ABL, A, remHeight, 0, sizeABR, remWidth );
LockedView( ABR, A, remHeight, remWidth, sizeABR, sizeABR );
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LockedRepartitionUpDiagonal
( const M& ATL, const M& ATR, M& A00, M& A01, M& A02,
M& A10, M& A11, M& A12,
const M& ABL, const M& ABR, M& A20, M& A21, M& A22, Int bsize=Blocksize() )
{
DEBUG_ONLY(CallStackEntry cse("LockedRepartitionUpDiagonal"))
LockedPartitionUpOffsetDiagonal
( ATL.Width()-ATL.Height(),
ATL, A00, A01,
A10, A11, bsize );
LockedPartitionUp( ATR, A02, A12, A11.Height() );
LockedPartitionLeft( ABL, A20, A21, A11.Width() );
LockedView( A22, ABR );
}
inline void
LockedRepartitionDownDiagonal
( const DM& ATL, const DM& ATR, DM& A00, DM& A01, DM& A02,
DM& A10, DM& A11, DM& A12,
const DM& ABL, const DM& ABR, DM& A20, DM& A21, DM& A22, Int bsize )
{
#ifndef RELEASE
CallStackEntry cse("LockedRepartitionDownDiagonal [DistMatrix]");
#endif
LockedView( A00, ATL );
LockedPartitionDownDiagonal( ABR, A11, A12,
A21, A22, bsize );
LockedPartitionDown( ABL, A10, A20, A11.Height() );
LockedPartitionRight( ATR, A01, A02, A11.Width() );
}
inline void
LockedRepartitionUpDiagonal
( const M& ATL, const M& ATR, M& A00, M& A01, M& A02,
M& A10, M& A11, M& A12,
const M& ABL, const M& ABR, M& A20, M& A21, M& A22, Int bsize )
{
#ifndef RELEASE
CallStackEntry cse("LockedRepartitionUpDiagonal [Matrix]");
#endif
LockedPartitionUpOffsetDiagonal
( ATL.Width()-ATL.Height(),
ATL, A00, A01,
A10, A11, bsize );
LockedPartitionUp( ATR, A02, A12, A11.Height() );
LockedPartitionLeft( ABL, A20, A21, A11.Width() );
LockedView( A22, ABR );
}
void ReshapeIntoGrid
( Int realSize, Int imagSize,
const ElementalMatrix<T>& x, ElementalMatrix<T>& X )
{
X.SetGrid( x.Grid() );
X.Resize( imagSize, realSize );
auto xSub = unique_ptr<ElementalMatrix<T>>
( x.Construct(x.Grid(),x.Root()) );
auto XSub = unique_ptr<ElementalMatrix<T>>
( X.Construct(X.Grid(),X.Root()) );
for( Int j=0; j<realSize; ++j )
{
View( *XSub, X, IR(0,imagSize), IR(j) );
LockedView( *xSub, x, IR(j*imagSize,(j+1)*imagSize), ALL );
Copy( *xSub, *XSub );
}
}
inline void
TrsvUN( UnitOrNonUnit diag, const DistMatrix<F>& U, DistMatrix<F>& x )
{
#ifndef RELEASE
PushCallStack("internal::TrsvUN");
if( U.Grid() != x.Grid() )
throw std::logic_error("{U,x} must be distributed over the same grid");
if( U.Height() != U.Width() )
throw std::logic_error("U must be square");
if( x.Width() != 1 && x.Height() != 1 )
throw std::logic_error("x must be a vector");
const int xLength = ( x.Width() == 1 ? x.Height() : x.Width() );
if( U.Width() != xLength )
throw std::logic_error("Nonconformal TrsvUN");
#endif
const Grid& g = U.Grid();
if( x.Width() == 1 )
{
// Matrix views
DistMatrix<F> U01(g),
U11(g);
DistMatrix<F>
xT(g), x0(g),
xB(g), x1(g),
x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,MR, STAR> x1_MR_STAR(g);
DistMatrix<F,MC, STAR> z_MC_STAR(g);
// Views of z[MC,* ], which will store updates to x
DistMatrix<F,MC,STAR> z0_MC_STAR(g),
z1_MC_STAR(g);
z_MC_STAR.AlignWith( U );
Zeros( x.Height(), 1, z_MC_STAR );
// Start the algorithm
PartitionUp
( x, xT,
xB, 0 );
while( xT.Height() > 0 )
{
RepartitionUp
( xT, x0,
x1,
/**/ /**/
xB, x2 );
const int n0 = x0.Height();
const int n1 = x1.Height();
LockedView( U01, U, 0, n0, n0, n1 );
LockedView( U11, U, n0, n0, n1, n1 );
View( z0_MC_STAR, z_MC_STAR, 0, 0, n0, 1 );
View( z1_MC_STAR, z_MC_STAR, n0, 0, n1, 1 );
x1_MR_STAR.AlignWith( U01 );
//----------------------------------------------------------------//
if( x2.Height() != 0 )
x1.SumScatterUpdate( F(1), z1_MC_STAR );
x1_STAR_STAR = x1;
U11_STAR_STAR = U11;
Trsv
( UPPER, NORMAL, diag,
U11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_MR_STAR = x1_STAR_STAR;
Gemv
( NORMAL, F(-1),
U01.LockedLocalMatrix(),
x1_MR_STAR.LockedLocalMatrix(),
F(1), z0_MC_STAR.LocalMatrix() );
//----------------------------------------------------------------//
x1_MR_STAR.FreeAlignments();
SlidePartitionUp
( xT, x0,
/**/ /**/
x1,
xB, x2 );
}
}
else
{
// Matrix views
DistMatrix<F> U01(g),
U11(g);
DistMatrix<F>
xL(g), xR(g),
x0(g), x1(g), x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,STAR,MR > x1_STAR_MR(g);
DistMatrix<F,MC, MR > z1(g);
DistMatrix<F,MR, MC > z1_MR_MC(g);
DistMatrix<F,STAR,MC > z_STAR_MC(g);
// Views of z[* ,MC]
DistMatrix<F,STAR,MC> z0_STAR_MC(g),
z1_STAR_MC(g);
z_STAR_MC.AlignWith( U );
Zeros( 1, x.Width(), z_STAR_MC );
// Start the algorithm
PartitionLeft( x, xL, xR, 0 );
while( xL.Width() > 0 )
{
RepartitionLeft
( xL, /**/ xR,
x0, x1, /**/ x2 );
const int n0 = x0.Width();
const int n1 = x1.Width();
LockedView( U01, U, 0, n0, n0, n1 );
LockedView( U11, U, n0, n0, n1, n1 );
View( z0_STAR_MC, z_STAR_MC, 0, 0, 1, n0 );
View( z1_STAR_MC, z_STAR_MC, 0, n0, 1, n1 );
x1_STAR_MR.AlignWith( U01 );
z1.AlignWith( x1 );
//----------------------------------------------------------------//
if( x2.Width() != 0 )
{
z1_MR_MC.SumScatterFrom( z1_STAR_MC );
z1 = z1_MR_MC;
Axpy( F(1), z1, x1 );
}
x1_STAR_STAR = x1;
U11_STAR_STAR = U11;
Trsv
( UPPER, NORMAL, diag,
U11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_STAR_MR = x1_STAR_STAR;
Gemv
( NORMAL, F(-1),
U01.LockedLocalMatrix(),
x1_STAR_MR.LockedLocalMatrix(),
F(1), z0_STAR_MC.LocalMatrix() );
//----------------------------------------------------------------//
x1_STAR_MR.FreeAlignments();
z1.FreeAlignments();
SlidePartitionLeft
( xL, /**/ xR,
x0, /**/ x1, x2 );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline Matrix<T> LockedView( const Matrix<T>& B )
{
Matrix<T> A;
LockedView( A, B );
return A;
}
inline void
RLHF
( Conjugation conjugation, int offset,
const DistMatrix<Complex<R> >& H,
const DistMatrix<Complex<R>,MD,STAR>& t,
DistMatrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("apply_packed_reflectors::RLHF");
if( H.Grid() != t.Grid() || t.Grid() != A.Grid() )
throw std::logic_error
("{H,t,A} must be distributed over the same grid");
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");
if( t.Height() != H.DiagonalLength( offset ) )
throw std::logic_error("t must be the same length as H's offset diag");
if( !t.AlignedWithDiagonal( H, offset ) )
throw std::logic_error("t must be aligned with H's 'offset' diagonal");
#endif
typedef Complex<R> C;
const Grid& g = H.Grid();
DistMatrix<C>
HTL(g), HTR(g), H00(g), H01(g), H02(g), HPan(g), HPanCopy(g),
HBL(g), HBR(g), H10(g), H11(g), H12(g),
H20(g), H21(g), H22(g);
DistMatrix<C> ALeft(g);
DistMatrix<C,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
DistMatrix<C,STAR,VR > HPan_STAR_VR(g);
DistMatrix<C,STAR,MR > HPan_STAR_MR(g);
DistMatrix<C,STAR,STAR> t1_STAR_STAR(g);
DistMatrix<C,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<C,STAR,MC > ZAdj_STAR_MC(g);
DistMatrix<C,STAR,VC > ZAdj_STAR_VC(g);
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 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 );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanHeight );
View( ALeft, A, 0, 0, A.Height(), HPanWidth );
HPan_STAR_MR.AlignWith( ALeft );
ZAdj_STAR_MC.AlignWith( ALeft );
ZAdj_STAR_VC.AlignWith( ALeft );
Zeros( HPan.Height(), ALeft.Height(), ZAdj_STAR_MC );
Zeros( HPan.Height(), HPan.Height(), SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, LOWER, offset, HPanCopy );
SetDiagonal( RIGHT, offset, HPanCopy, C(1) );
HPan_STAR_VR = HPanCopy;
Herk
( UPPER, NORMAL,
C(1), HPan_STAR_VR.LockedMatrix(),
C(0), SInv_STAR_STAR.Matrix() );
SInv_STAR_STAR.SumOverGrid();
t1_STAR_STAR = t1;
FixDiagonal( conjugation, t1_STAR_STAR, SInv_STAR_STAR );
HPan_STAR_MR = HPan_STAR_VR;
LocalGemm
( NORMAL, ADJOINT,
C(1), HPan_STAR_MR, ALeft, C(0), ZAdj_STAR_MC );
ZAdj_STAR_VC.SumScatterFrom( ZAdj_STAR_MC );
LocalTrsm
( LEFT, UPPER, ADJOINT, NON_UNIT,
C(1), SInv_STAR_STAR, ZAdj_STAR_VC );
ZAdj_STAR_MC = ZAdj_STAR_VC;
LocalGemm
( ADJOINT, NORMAL,
C(-1), ZAdj_STAR_MC, HPan_STAR_MR, C(1), ALeft );
//--------------------------------------------------------------------//
HPan_STAR_MR.FreeAlignments();
ZAdj_STAR_MC.FreeAlignments();
ZAdj_STAR_VC.FreeAlignments();
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
RLHF
( Conjugation conjugation, int offset,
const Matrix<Complex<R> >& H,
const Matrix<Complex<R> >& t,
Matrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("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");
if( t.Height() != H.DiagonalLength( offset ) )
throw std::logic_error("t must be the same length as H's offset diag");
#endif
typedef Complex<R> C;
Matrix<C>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<C> ALeft;
Matrix<C>
tT, t0,
tB, t1,
t2;
Matrix<C> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 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 );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanHeight );
View( ALeft, A, 0, 0, A.Height(), HPanWidth );
Zeros( ALeft.Height(), HPan.Height(), Z );
Zeros( HPan.Height(), HPan.Height(), SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, LOWER, offset, HPanCopy );
SetDiagonal( RIGHT, offset, HPanCopy, C(1) );
Herk( UPPER, NORMAL, C(1), HPanCopy, C(0), SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, ADJOINT, C(1), ALeft, HPanCopy, C(0), Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, C(1), SInv, Z );
Gemm( NORMAL, NORMAL, C(-1), Z, HPanCopy, C(1), ALeft );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}