inline void
TwoSidedTrsmUVar1( UnitOrNonUnit diag, Matrix<F>& A, const Matrix<F>& U )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrsmUVar1");
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
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
UTL, UTR, U00, U01, U02,
UBL, UBR, U10, U11, U12,
U20, U21, U22;
// Temporary products
Matrix<F> Y01;
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 );
//--------------------------------------------------------------------//
// Y01 := A00 U01
Zeros( A01.Height(), A01.Width(), Y01 );
Hemm( LEFT, UPPER, F(1), A00, U01, F(0), Y01 );
// A01 := inv(U00)' A01
Trsm( LEFT, UPPER, ADJOINT, diag, F(1), U00, A01 );
// A01 := A01 - 1/2 Y01
Axpy( F(-1)/F(2), Y01, A01 );
// A11 := A11 - (U01' A01 + A01' U01)
Her2k( UPPER, ADJOINT, F(-1), U01, A01, F(1), A11 );
// A11 := inv(U11)' A11 inv(U11)
TwoSidedTrsmUUnb( diag, A11, U11 );
// A01 := A01 - 1/2 Y01
Axpy( F(-1)/F(2), Y01, A01 );
// A01 := A01 inv(U11)
Trsm( RIGHT, UPPER, NORMAL, diag, F(1), U11, A01 );
//--------------------------------------------------------------------//
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
}
inline void
TwoSidedTrmmLVar2( UnitOrNonUnit diag, Matrix<F>& A, const Matrix<F>& L )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrmmLVar2");
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
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
LTL, LTR, L00, L01, L02,
LBL, LBR, L10, L11, L12,
L20, L21, L22;
// Temporary products
Matrix<F> Y21;
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 );
//--------------------------------------------------------------------//
// A10 := L11' A10
Trmm( LEFT, LOWER, ADJOINT, diag, F(1), L11, A10 );
// A10 := A10 + L21' A20
Gemm( ADJOINT, NORMAL, F(1), L21, A20, F(1), A10 );
// Y21 := A22 L21
Zeros( A21.Height(), A21.Width(), Y21 );
Hemm( LEFT, LOWER, F(1), A22, L21, F(0), Y21 );
// A21 := A21 L11
Trmm( RIGHT, LOWER, NORMAL, diag, F(1), L11, A21 );
// A21 := A21 + 1/2 Y21
Axpy( F(1)/F(2), Y21, A21 );
// A11 := L11' A11 L11
TwoSidedTrmmLUnb( diag, A11, L11 );
// A11 := A11 + (A21' L21 + L21' A21)
Her2k( LOWER, ADJOINT, F(1), A21, L21, F(1), A11 );
// A21 := A21 + 1/2 Y21
Axpy( F(1)/F(2), Y21, A21 );
//--------------------------------------------------------------------//
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
}
inline void
TwoSidedTrsmLVar2( UnitOrNonUnit diag, Matrix<F>& A, const Matrix<F>& L )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmLVar2");
if( A.Height() != A.Width() )
LogicError("A must be square");
if( L.Height() != L.Width() )
LogicError("Triangular matrices must be square");
if( A.Height() != L.Height() )
LogicError("A and L must be the same size");
#endif
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
LTL, LTR, L00, L01, L02,
LBL, LBR, L10, L11, L12,
L20, L21, L22;
// Temporary products
Matrix<F> X11;
Matrix<F> Y10;
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 );
//--------------------------------------------------------------------//
// Y10 := L10 A00
Zeros( Y10, L10.Height(), A00.Width() );
Hemm( RIGHT, LOWER, F(1), A00, L10, F(0), Y10 );
// A10 := A10 - 1/2 Y10
Axpy( F(-1)/F(2), Y10, A10 );
// A11 := A11 - (A10 L10' + L10 A10')
Her2k( LOWER, NORMAL, F(-1), A10, L10, F(1), A11 );
// A11 := inv(L11) A11 inv(L11)'
TwoSidedTrsmLUnb( diag, A11, L11 );
// A21 := A21 - A20 L10'
Gemm( NORMAL, ADJOINT, F(-1), A20, L10, F(1), A21 );
// A21 := A21 inv(L11)'
Trsm( RIGHT, LOWER, ADJOINT, diag, F(1), L11, A21 );
// A10 := A10 - 1/2 Y10
Axpy( F(-1)/F(2), Y10, A10 );
// A10 := inv(L11) A10
Trsm( LEFT, LOWER, NORMAL, diag, F(1), L11, A10 );
//--------------------------------------------------------------------//
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 );
}
}
inline void
TwoSidedTrsmUVar5( UnitOrNonUnit diag, Matrix<F>& A, const Matrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar5");
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
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
UTL, UTR, U00, U01, U02,
UBL, UBR, U10, U11, U12,
U20, U21, U22;
// Temporary products
Matrix<F> Y12;
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 );
//--------------------------------------------------------------------//
// A11 := inv(U11)' A11 inv(U11)
TwoSidedTrsmUUnb( diag, A11, U11 );
// Y12 := A11 U12
Zeros( Y12, A12.Height(), A12.Width() );
Hemm( LEFT, UPPER, F(1), A11, U12, F(0), Y12 );
// A12 := inv(U11)' A12
Trsm( LEFT, UPPER, ADJOINT, diag, F(1), U11, A12 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
// A22 := A22 - (A12' U12 + U12' A12)
Her2k( UPPER, ADJOINT, F(-1), A12, U12, F(1), A22 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
// A12 := A12 inv(U22)
Trsm( RIGHT, UPPER, NORMAL, diag, F(1), U22, A12 );
//--------------------------------------------------------------------//
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
TwoSidedTrsmUVar5
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar5");
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,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,MC > A12_STAR_MC(g);
DistMatrix<F,STAR,MR > A12_STAR_MR(g);
DistMatrix<F,STAR,VC > A12_STAR_VC(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,MR > U12_STAR_MR(g);
DistMatrix<F,STAR,VC > U12_STAR_VC(g);
DistMatrix<F,STAR,VR > U12_STAR_VR(g);
DistMatrix<F,STAR,VR > Y12_STAR_VR(g);
DistMatrix<F> Y12(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_MC.AlignWith( A22 );
A12_STAR_MR.AlignWith( A22 );
A12_STAR_VC.AlignWith( A22 );
A12_STAR_VR.AlignWith( A22 );
U12_STAR_MC.AlignWith( A22 );
U12_STAR_MR.AlignWith( A22 );
U12_STAR_VC.AlignWith( A22 );
U12_STAR_VR.AlignWith( A22 );
Y12.AlignWith( A12 );
Y12_STAR_VR.AlignWith( A12 );
//--------------------------------------------------------------------//
// A11 := inv(U11)' A11 inv(U11)
U11_STAR_STAR = U11;
A11_STAR_STAR = A11;
LocalTwoSidedTrsm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
// Y12 := A11 U12
U12_STAR_VR = U12;
Zeros( Y12_STAR_VR, A12.Height(), A12.Width() );
Hemm
( LEFT, UPPER,
F(1), A11_STAR_STAR.Matrix(), U12_STAR_VR.Matrix(),
F(0), Y12_STAR_VR.Matrix() );
Y12 = Y12_STAR_VR;
// 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;
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
// A22 := A22 - (A12' U12 + U12' A12)
A12_STAR_VR = A12;
A12_STAR_VC = A12_STAR_VR;
U12_STAR_VC = U12_STAR_VR;
A12_STAR_MC = A12_STAR_VC;
U12_STAR_MC = U12_STAR_VC;
A12_STAR_MR = A12_STAR_VR;
U12_STAR_MR = U12_STAR_VR;
LocalTrr2k
( UPPER, ADJOINT, ADJOINT,
F(-1), U12_STAR_MC, A12_STAR_MR,
A12_STAR_MC, U12_STAR_MR,
F(1), A22 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
// A12 := A12 inv(U22)
//
// This is the bottleneck because A12 only has blocksize rows
Trsm( RIGHT, UPPER, NORMAL, diag, F(1), U22, A12 );
//--------------------------------------------------------------------//
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
TwoSidedTrmmLVar4
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& L )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrmmLVar4");
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,STAR,VR > A10_STAR_VR(g);
DistMatrix<F,STAR,MR > A10_STAR_MR(g);
DistMatrix<F,STAR,MC > A10_STAR_MC(g);
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,VC, STAR> A21_VC_STAR(g);
DistMatrix<F,MC, STAR> A21_MC_STAR(g);
DistMatrix<F,STAR,VR > L10_STAR_VR(g);
DistMatrix<F,MR, STAR> L10Adj_MR_STAR(g);
DistMatrix<F,STAR,MC > L10_STAR_MC(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,VR > Y10_STAR_VR(g);
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 );
A10_STAR_VR.AlignWith( A00 );
A10_STAR_MR.AlignWith( A00 );
A10_STAR_MC.AlignWith( A00 );
A21_MC_STAR.AlignWith( A20 );
L10_STAR_VR.AlignWith( A00 );
L10Adj_MR_STAR.AlignWith( A00 );
L10_STAR_MC.AlignWith( A00 );
Y10_STAR_VR.AlignWith( A10 );
//--------------------------------------------------------------------//
// Y10 := A11 L10
A11_STAR_STAR = A11;
L10Adj_MR_STAR.AdjointFrom( L10 );
L10_STAR_VR.AdjointFrom( L10Adj_MR_STAR );
Y10_STAR_VR.ResizeTo( A10.Height(), A10.Width() );
Zero( Y10_STAR_VR );
Hemm
( LEFT, LOWER,
F(1), A11_STAR_STAR.LockedLocalMatrix(),
L10_STAR_VR.LockedLocalMatrix(),
F(0), Y10_STAR_VR.LocalMatrix() );
// A10 := A10 + 1/2 Y10
A10_STAR_VR = A10;
Axpy( F(1)/F(2), Y10_STAR_VR, A10_STAR_VR );
// A00 := A00 + (A10' L10 + L10' A10)
A10_STAR_MR = A10_STAR_VR;
A10_STAR_MC = A10_STAR_VR;
L10_STAR_MC = L10_STAR_VR;
LocalTrr2k
( LOWER, ADJOINT, ADJOINT, ADJOINT,
F(1), A10_STAR_MC, L10Adj_MR_STAR,
L10_STAR_MC, A10_STAR_MR,
F(1), A00 );
// A10 := A10 + 1/2 Y10
Axpy( F(1)/F(2), Y10_STAR_VR, A10_STAR_VR );
// A10 := L11' A10
L11_STAR_STAR = L11;
LocalTrmm
( LEFT, LOWER, ADJOINT, diag, F(1), L11_STAR_STAR, A10_STAR_VR );
A10 = A10_STAR_VR;
// A20 := A20 + A21 L10
A21_MC_STAR = A21;
LocalGemm
( NORMAL, ADJOINT, F(1), A21_MC_STAR, L10Adj_MR_STAR, F(1), A20 );
// A11 := L11' A11 L11
LocalTwoSidedTrmm( LOWER, diag, A11_STAR_STAR, L11_STAR_STAR );
A11 = A11_STAR_STAR;
// A21 := A21 L11
A21_VC_STAR = A21_MC_STAR;
LocalTrmm
( RIGHT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, A21_VC_STAR );
A21 = A21_VC_STAR;
//--------------------------------------------------------------------//
A10_STAR_VR.FreeAlignments();
A10_STAR_MR.FreeAlignments();
A10_STAR_MC.FreeAlignments();
A21_MC_STAR.FreeAlignments();
L10_STAR_VR.FreeAlignments();
L10Adj_MR_STAR.FreeAlignments();
L10_STAR_MC.FreeAlignments();
Y10_STAR_VR.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
}
