inline void
TrsmLUNLarge
( UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLUNLarge");
#endif
const Grid& g = U.Grid();
// Matrix views
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);
DistMatrix<F> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,MR > X1_STAR_MR(g);
DistMatrix<F,STAR,VR > X1_STAR_VR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
U01_MC_STAR.AlignWith( X0 );
X1_STAR_MR.AlignWith( X0 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11; // U11[* ,* ] <- U11[MC,MR]
X1_STAR_VR = X1; // X1[* ,VR] <- X1[MC,MR]
// X1[* ,VR] := U11^-1[* ,* ] X1[* ,VR]
LocalTrsm
( LEFT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, X1_STAR_VR,
checkIfSingular );
X1_STAR_MR = X1_STAR_VR; // X1[* ,MR] <- X1[* ,VR]
X1 = X1_STAR_MR; // X1[MC,MR] <- X1[* ,MR]
U01_MC_STAR = U01; // U01[MC,* ] <- U01[MC,MR]
// X0[MC,MR] -= U01[MC,* ] X1[* ,MR]
LocalGemm( NORMAL, NORMAL, F(-1), U01_MC_STAR, X1_STAR_MR, F(1), X0 );
//--------------------------------------------------------------------//
U01_MC_STAR.FreeAlignments();
X1_STAR_MR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrsmLUNSmall
( UnitOrNonUnit diag,
F alpha, const DistMatrix<F,VC,STAR>& U, DistMatrix<F,VC,STAR>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLUNSmall");
if( U.Grid() != X.Grid() )
throw std::logic_error
("U and X must be distributed over the same grid");
if( U.Height() != U.Width() || U.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrsmLUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str() );
}
if( U.ColAlignment() != X.ColAlignment() )
throw std::logic_error("U and X are assumed to be aligned");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<F,VC,STAR>
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);
DistMatrix<F,VC,STAR> 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);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11; // U11[* ,* ] <- U11[VC,* ]
X1_STAR_STAR = X1; // X1[* ,* ] <- X1[VC,* ]
// X1[* ,* ] := U11^-1[* ,* ] X1[* ,* ]
LocalTrsm
( LEFT, UPPER, NORMAL, diag,
F(1), U11_STAR_STAR, X1_STAR_STAR, checkIfSingular );
X1 = X1_STAR_STAR;
// X0[VC,* ] -= U01[VC,* ] X1[* ,* ]
LocalGemm( NORMAL, NORMAL, F(-1), U01, X1_STAR_STAR, F(1), X0 );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TwoSidedTrsmUVar4
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar4");
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,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,MC > A01Trans_STAR_MC(g);
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,STAR,VC > A12_STAR_VC(g);
DistMatrix<F,STAR,MC > A12_STAR_MC(g);
DistMatrix<F,STAR,MR > A12_STAR_MR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MR, STAR> U12Trans_MR_STAR(g);
DistMatrix<F,VR, STAR> U12Trans_VR_STAR(g);
DistMatrix<F,STAR,VR > U12_STAR_VR(g);
DistMatrix<F,STAR,VC > U12_STAR_VC(g);
DistMatrix<F,STAR,MC > U12_STAR_MC(g);
DistMatrix<F,STAR,VR > Y12_STAR_VR(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 );
A01_VC_STAR.AlignWith( A02 );
A01Trans_STAR_MC.AlignWith( A02 );
A12_STAR_VR.AlignWith( A22 );
A12_STAR_VC.AlignWith( A22 );
A12_STAR_MC.AlignWith( A22 );
A12_STAR_MR.AlignWith( A22 );
U12Trans_MR_STAR.AlignWith( A02 );
U12Trans_VR_STAR.AlignWith( A02 );
U12_STAR_VR.AlignWith( A02 );
U12_STAR_VC.AlignWith( A22 );
U12_STAR_MC.AlignWith( A22 );
Y12_STAR_VR.AlignWith( A12 );
//--------------------------------------------------------------------//
// A01 := A01 inv(U11)
A01_VC_STAR = A01;
U11_STAR_STAR = U11;
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;
// A02 := A02 - A01 U12
A01Trans_STAR_MC.TransposeFrom( A01_VC_STAR );
U12Trans_MR_STAR.TransposeFrom( U12 );
LocalGemm
( TRANSPOSE, TRANSPOSE,
F(-1), A01Trans_STAR_MC, U12Trans_MR_STAR, F(1), A02 );
// Y12 := A11 U12
U12Trans_VR_STAR = U12Trans_MR_STAR;
Zeros( U12_STAR_VR, A12.Height(), A12.Width() );
Transpose( U12Trans_VR_STAR.Matrix(), U12_STAR_VR.Matrix() );
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() );
// A12 := inv(U11)' A12
A12_STAR_VR = A12;
LocalTrsm
( LEFT, UPPER, ADJOINT, diag, F(1), U11_STAR_STAR, A12_STAR_VR );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12_STAR_VR, A12_STAR_VR );
// A22 := A22 - (A12' U12 + U12' A12)
A12_STAR_MR = A12_STAR_VR;
A12_STAR_VC = A12_STAR_VR;
U12_STAR_VC = U12_STAR_VR;
A12_STAR_MC = A12_STAR_VC;
U12_STAR_MC = U12_STAR_VC;
LocalTrr2k
( UPPER, ADJOINT, TRANSPOSE, ADJOINT,
F(-1), A12_STAR_MC, U12Trans_MR_STAR,
U12_STAR_MC, A12_STAR_MR,
F(1), A22 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12_STAR_VR, 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
TwoSidedTrsmUVar5
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrsmUVar5");
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>
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;
Y12_STAR_VR.ResizeTo( A12.Height(), A12.Width() );
Hemm
( LEFT, UPPER,
F(1), A11_STAR_STAR.LocalMatrix(), U12_STAR_VR.LocalMatrix(),
F(0), Y12_STAR_VR.LocalMatrix() );
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 );
//--------------------------------------------------------------------//
A12_STAR_MC.FreeAlignments();
A12_STAR_MR.FreeAlignments();
A12_STAR_VC.FreeAlignments();
A12_STAR_VR.FreeAlignments();
U12_STAR_MC.FreeAlignments();
U12_STAR_MR.FreeAlignments();
U12_STAR_VC.FreeAlignments();
U12_STAR_VR.FreeAlignments();
Y12.FreeAlignments();
Y12_STAR_VR.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
}
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
TrsmLUNMedium
( UnitOrNonUnit diag, F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrsmLUNMedium");
#endif
const Grid& g = U.Grid();
// Matrix views
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);
DistMatrix<F> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MR, STAR> X1Trans_MR_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
U01_MC_STAR.AlignWith( X0 );
X1Trans_MR_STAR.AlignWith( X0 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11; // U11[* ,* ] <- U11[MC,MR]
X1Trans_MR_STAR.TransposeFrom( X1 ); // X1[* ,MR] <- X1[MC,MR]
// X1[* ,MR] := U11^-1[* ,* ] X1[* ,MR]
//
// X1^T[MR,* ] := X1^T[MR,* ] U11^-T[* ,* ]
LocalTrsm
( RIGHT, UPPER, TRANSPOSE, diag,
F(1), U11_STAR_STAR, X1Trans_MR_STAR, checkIfSingular );
X1.TransposeFrom( X1Trans_MR_STAR );
U01_MC_STAR = U01; // U01[MC,* ] <- U01[MC,MR]
// X0[MC,MR] -= U01[MC,* ] X1[* ,MR]
LocalGemm
( NORMAL, TRANSPOSE, F(-1), U01_MC_STAR, X1Trans_MR_STAR, F(1), X0 );
//--------------------------------------------------------------------//
U01_MC_STAR.FreeAlignments();
X1Trans_MR_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
}
inline void
TrmmRUNC
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& U,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmRUNC");
if( U.Grid() != X.Grid() )
throw std::logic_error
("U and X must be distributed over the same grid");
if( U.Height() != U.Width() || X.Width() != U.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmRUNC: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T>
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);
DistMatrix<T> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<T,MR, STAR> U12Trans_MR_STAR(g);
DistMatrix<T,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<T,VC, STAR> X1_VC_STAR(g);
DistMatrix<T,MC, STAR> X1_MC_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionLeft( X, XL, XR, 0 );
while( XL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionLeft
( XL, /**/ XR,
X0, X1, /**/ X2 );
X1_MC_STAR.AlignWith( X2 );
U12Trans_MR_STAR.AlignWith( X2 );
X1_VC_STAR.AlignWith( X1 );
//--------------------------------------------------------------------//
X1_MC_STAR = X1;
U12Trans_MR_STAR.TransposeFrom( U12 );
LocalGemm
( NORMAL, TRANSPOSE, T(1), X1_MC_STAR, U12Trans_MR_STAR, T(1), X2 );
U11_STAR_STAR = U11;
X1_VC_STAR = X1_MC_STAR;
LocalTrmm
( RIGHT, UPPER, NORMAL, diag, T(1), U11_STAR_STAR, X1_VC_STAR );
X1 = X1_VC_STAR;
//--------------------------------------------------------------------//
X1_MC_STAR.FreeAlignments();
U12Trans_MR_STAR.FreeAlignments();
X1_VC_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionLeft
( XL, /**/ XR,
X0, /**/ X1, X2 );
}
}
inline void
TrdtrmmUVar1( Orientation orientation, DistMatrix<F>& U )
{
#ifndef RELEASE
PushCallStack("internal::TrdtrmmUVar1");
if( U.Height() != U.Width() )
throw std::logic_error("U must be square");
if( orientation == NORMAL )
throw std::logic_error("Orientation must be (conjugate-)transpose");
#endif
const Grid& g = U.Grid();
// Matrix views
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);
DistMatrix<F,MD,STAR> d1(g);
// Temporary distributions
DistMatrix<F,MC, STAR> S01_MC_STAR(g);
DistMatrix<F,VC, STAR> S01_VC_STAR(g);
DistMatrix<F,VR, STAR> U01_VR_STAR(g);
DistMatrix<F,STAR,MR > U01AdjOrTrans_STAR_MR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
S01_MC_STAR.AlignWith( U );
S01_VC_STAR.AlignWith( U );
U01_VR_STAR.AlignWith( U );
U01AdjOrTrans_STAR_MR.AlignWith( U );
PartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( UTL.Height() < U.Height() && UTL.Width() < U.Height() )
{
RepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
//--------------------------------------------------------------------//
U11.GetDiagonal( d1 );
S01_MC_STAR = U01;
S01_VC_STAR = S01_MC_STAR;
U01_VR_STAR = S01_VC_STAR;
if( orientation == TRANSPOSE )
{
DiagonalSolve( RIGHT, NORMAL, d1, U01_VR_STAR );
U01AdjOrTrans_STAR_MR.TransposeFrom( U01_VR_STAR );
}
else
{
DiagonalSolve( RIGHT, ADJOINT, d1, U01_VR_STAR );
U01AdjOrTrans_STAR_MR.AdjointFrom( U01_VR_STAR );
}
LocalTrrk( UPPER, F(1), S01_MC_STAR, U01AdjOrTrans_STAR_MR, F(1), U00 );
U11_STAR_STAR = U11;
LocalTrmm
( RIGHT, UPPER, ADJOINT, UNIT, F(1), U11_STAR_STAR, U01_VR_STAR );
U01 = U01_VR_STAR;
LocalTrdtrmm( orientation, UPPER, U11_STAR_STAR );
U11 = U11_STAR_STAR;
//--------------------------------------------------------------------//
d1.FreeAlignments();
SlidePartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrsmRUT
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmRUT");
if( orientation == NORMAL )
throw std::logic_error("TrsmRUT expects a (Conjugate)Transpose option");
#endif
const Grid& g = U.Grid();
// Matrix views
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);
DistMatrix<F> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<F,VR, STAR> U01_VR_STAR(g);
DistMatrix<F,STAR,MR > U01AdjOrTrans_STAR_MR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,VC, STAR> X1_VC_STAR(g);
DistMatrix<F,STAR,MC > X1Trans_STAR_MC(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionLeft( X, XL, XR, 0 );
while( XL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionLeft
( XL, /**/ XR,
X0, X1, /**/ X2 );
X1_VC_STAR.AlignWith( X0 );
X1Trans_STAR_MC.AlignWith( X0 );
U01_VR_STAR.AlignWith( X0 );
U01AdjOrTrans_STAR_MR.AlignWith( X0 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, UPPER, orientation, diag,
F(1), U11_STAR_STAR, X1_VC_STAR, checkIfSingular );
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
U01_VR_STAR = U01;
if( orientation == ADJOINT )
U01AdjOrTrans_STAR_MR.AdjointFrom( U01_VR_STAR );
else
U01AdjOrTrans_STAR_MR.TransposeFrom( U01_VR_STAR );
// X0[MC,MR] -= X1[MC,* ] (U01[MR,* ])^(T/H)
// = X1^T[* ,MC] (U01^(T/H))[* ,MR]
LocalGemm
( TRANSPOSE, NORMAL,
F(-1), X1Trans_STAR_MC, U01AdjOrTrans_STAR_MR, F(1), X0 );
//--------------------------------------------------------------------//
X1_VC_STAR.FreeAlignments();
X1Trans_STAR_MC.FreeAlignments();
U01_VR_STAR.FreeAlignments();
U01AdjOrTrans_STAR_MR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionLeft
( XL, /**/ XR,
X0, /**/ X1, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LocalTrmmAccumulateRUN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
const DistMatrix<T,MC, MR >& U,
const DistMatrix<T,STAR,MC >& X_STAR_MC,
DistMatrix<T,MR, STAR>& ZTrans_MR_STAR )
{
#ifndef RELEASE
CallStackEntry entry("internal::LocalTrmmAccumulateRUN");
if( U.Grid() != X_STAR_MC.Grid() ||
X_STAR_MC.Grid() != ZTrans_MR_STAR.Grid() )
throw std::logic_error
("{U,X,Z} must be distributed over the same grid");
if( U.Height() != U.Width() ||
U.Height() != X_STAR_MC.Width() ||
U.Height() != ZTrans_MR_STAR.Height() )
{
std::ostringstream msg;
msg << "Nonconformal LocalTrmmAccumulateRUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X[* ,MC] ~ " << X_STAR_MC.Height() << " x "
<< X_STAR_MC.Width() << "\n"
<< " Z^H/T[MR,* ] ~ " << ZTrans_MR_STAR.Height() << " x "
<< ZTrans_MR_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( X_STAR_MC.RowAlignment() != U.ColAlignment() ||
ZTrans_MR_STAR.ColAlignment() != U.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T>
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);
DistMatrix<T> D11(g);
DistMatrix<T,STAR,MC>
XL_STAR_MC(g), XR_STAR_MC(g),
X0_STAR_MC(g), X1_STAR_MC(g), X2_STAR_MC(g);
DistMatrix<T,MR,STAR>
ZTTrans_MR_STAR(g), Z0Trans_MR_STAR(g),
ZBTrans_MR_STAR(g), Z1Trans_MR_STAR(g),
Z2Trans_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
LockedPartitionRight( X_STAR_MC, XL_STAR_MC, XR_STAR_MC, 0 );
PartitionDown
( ZTrans_MR_STAR, ZTTrans_MR_STAR,
ZBTrans_MR_STAR, 0 );
while( UTL.Height() < U.Height() )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
LockedRepartitionRight
( XL_STAR_MC, /**/ XR_STAR_MC,
X0_STAR_MC, /**/ X1_STAR_MC, X2_STAR_MC );
RepartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
/***************/ /***************/
Z1Trans_MR_STAR,
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
D11.AlignWith( U11 );
//--------------------------------------------------------------------//
D11 = U11;
MakeTriangular( UPPER, D11 );
if( diag == UNIT )
SetDiagonal( D11, T(1) );
LocalGemm
( orientation, orientation,
alpha, D11, X1_STAR_MC, T(1), Z1Trans_MR_STAR );
LocalGemm
( orientation, orientation,
alpha, U01, X0_STAR_MC, T(1), Z1Trans_MR_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlideLockedPartitionRight
( XL_STAR_MC, /**/ XR_STAR_MC,
X0_STAR_MC, X1_STAR_MC, /**/ X2_STAR_MC );
SlidePartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
Z1Trans_MR_STAR,
/***************/ /***************/
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
}
PopBlocksizeStack();
}
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
}
inline void
internal::LocalTrmmAccumulateLUN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
const DistMatrix<T,MC, MR >& U,
const DistMatrix<T,STAR,MR >& XAdjOrTrans_STAR_MR,
DistMatrix<T,MC, STAR>& Z_MC_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalTrmmAccumulateLUN");
if( U.Grid() != XAdjOrTrans_STAR_MR.Grid() ||
XAdjOrTrans_STAR_MR.Grid() != Z_MC_STAR.Grid() )
throw std::logic_error
("{U,X,Z} must be distributed over the same grid");
if( U.Height() != U.Width() ||
U.Height() != XAdjOrTrans_STAR_MR.Width() ||
U.Height() != Z_MC_STAR.Height() ||
XAdjOrTrans_STAR_MR.Height() != Z_MC_STAR.Width() )
{
std::ostringstream msg;
msg << "Nonconformal LocalTrmmAccumulateLUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X^H/T[* ,MR] ~ " << XAdjOrTrans_STAR_MR.Height() << " x "
<< XAdjOrTrans_STAR_MR.Width() << "\n"
<< " Z[MC,* ] ~ " << Z_MC_STAR.Height() << " x "
<< Z_MC_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( XAdjOrTrans_STAR_MR.RowAlignment() != U.RowAlignment() ||
Z_MC_STAR.ColAlignment() != U.ColAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T,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);
DistMatrix<T,MC,MR> D11(g);
DistMatrix<T,STAR,MR>
XLAdjOrTrans_STAR_MR(g), XRAdjOrTrans_STAR_MR(g),
X0AdjOrTrans_STAR_MR(g), X1AdjOrTrans_STAR_MR(g),
X2AdjOrTrans_STAR_MR(g);
DistMatrix<T,MC,STAR>
ZT_MC_STAR(g), Z0_MC_STAR(g),
ZB_MC_STAR(g), Z1_MC_STAR(g),
Z2_MC_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
LockedPartitionRight
( XAdjOrTrans_STAR_MR, XLAdjOrTrans_STAR_MR, XRAdjOrTrans_STAR_MR, 0 );
PartitionDown
( Z_MC_STAR, ZT_MC_STAR,
ZB_MC_STAR, 0 );
while( UTL.Height() < U.Height() )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
LockedRepartitionRight
( XLAdjOrTrans_STAR_MR, /**/ XRAdjOrTrans_STAR_MR,
X0AdjOrTrans_STAR_MR, /**/ X1AdjOrTrans_STAR_MR, X2AdjOrTrans_STAR_MR
);
RepartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
/**********/ /**********/
Z1_MC_STAR,
ZB_MC_STAR, Z2_MC_STAR );
D11.AlignWith( U11 );
//--------------------------------------------------------------------//
D11 = U11;
MakeTrapezoidal( LEFT, UPPER, 0, D11 );
if( diag == UNIT )
SetDiagonalToOne( D11 );
internal::LocalGemm
( NORMAL, orientation, alpha, D11, X1AdjOrTrans_STAR_MR,
(T)1, Z1_MC_STAR );
internal::LocalGemm
( NORMAL, orientation, alpha, U01, X1AdjOrTrans_STAR_MR,
(T)1, Z0_MC_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlideLockedPartitionRight
( XLAdjOrTrans_STAR_MR, /**/ XRAdjOrTrans_STAR_MR,
X0AdjOrTrans_STAR_MR, X1AdjOrTrans_STAR_MR, /**/ X2AdjOrTrans_STAR_MR
);
SlidePartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
Z1_MC_STAR,
/**********/ /**********/
ZB_MC_STAR, Z2_MC_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
internal::TrmmLUNC
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T,MC,MR>& U,
DistMatrix<T,MC,MR>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLUNC");
if( U.Grid() != X.Grid() )
throw std::logic_error
("U and X must be distributed over the same grid");
if( U.Height() != U.Width() || U.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T,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);
DistMatrix<T,MC,MR> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<T,STAR,MC > U12_STAR_MC(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> D1Trans_MR_STAR(g);
DistMatrix<T,MR, MC > D1Trans_MR_MC(g);
DistMatrix<T,MC, MR > D1(g);
// Start the algorithm
Scal( alpha, X );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionDown
( X, XT,
XB, 0 );
while( XB.Height() > 0 )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
RepartitionDown
( XT, X0,
/**/ /**/
X1,
XB, X2 );
U12_STAR_MC.AlignWith( X2 );
D1Trans_MR_STAR.AlignWith( X1 );
D1Trans_MR_MC.AlignWith( X1 );
D1.AlignWith( X1 );
D1Trans_MR_STAR.ResizeTo( X1.Width(), X1.Height() );
D1.ResizeTo( X1.Height(), X1.Width() );
//--------------------------------------------------------------------//
X1_STAR_VR = X1;
U11_STAR_STAR = U11;
internal::LocalTrmm
( LEFT, UPPER, NORMAL, diag, (T)1, U11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
U12_STAR_MC = U12;
internal::LocalGemm
( TRANSPOSE, TRANSPOSE, (T)1, X2, U12_STAR_MC, (T)0, D1Trans_MR_STAR );
D1Trans_MR_MC.SumScatterFrom( D1Trans_MR_STAR );
Transpose( D1Trans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
Axpy( (T)1, D1, X1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
D1Trans_MR_MC.FreeAlignments();
D1Trans_MR_STAR.FreeAlignments();
U12_STAR_MC.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlidePartitionDown
( XT, X0,
X1,
/**/ /**/
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TwoSidedTrmmUVar5
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrmmUVar5");
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>
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,MC, STAR> A01_MC_STAR(g);
DistMatrix<F,MR, STAR> A01_MR_STAR(g);
DistMatrix<F,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,MR, STAR> U01_MR_STAR(g);
DistMatrix<F,VC, STAR> U01_VC_STAR(g);
DistMatrix<F,VC, STAR> Y01_VC_STAR(g);
DistMatrix<F> Y01(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 );
A01_MC_STAR.AlignWith( A00 );
A01_MR_STAR.AlignWith( A00 );
A01_VC_STAR.AlignWith( A00 );
U01_MC_STAR.AlignWith( A00 );
U01_MR_STAR.AlignWith( A00 );
U01_VC_STAR.AlignWith( A00 );
Y01.AlignWith( A01 );
Y01_VC_STAR.AlignWith( A01 );
//--------------------------------------------------------------------//
// Y01 := U01 A11
A11_STAR_STAR = A11;
U01_VC_STAR = U01;
Y01_VC_STAR.ResizeTo( A01.Height(), A01.Width() );
Hemm
( RIGHT, UPPER,
F(1), A11_STAR_STAR.LocalMatrix(), U01_VC_STAR.LocalMatrix(),
F(0), Y01_VC_STAR.LocalMatrix() );
Y01 = Y01_VC_STAR;
// A01 := U00 A01
Trmm( LEFT, UPPER, NORMAL, diag, F(1), U00, A01 );
// A01 := A01 + 1/2 Y01
Axpy( F(1)/F(2), Y01, A01 );
// A00 := A00 + (U01 A01' + A01 U01')
A01_MC_STAR = A01;
U01_MC_STAR = U01;
A01_VC_STAR = A01_MC_STAR;
A01_MR_STAR = A01_VC_STAR;
U01_MR_STAR = U01_MC_STAR;
LocalTrr2k
( UPPER, ADJOINT, ADJOINT,
F(1), U01_MC_STAR, A01_MR_STAR,
A01_MC_STAR, U01_MR_STAR,
F(1), A00 );
// A01 := A01 + 1/2 Y01
Axpy( F(1)/F(2), Y01_VC_STAR, A01_VC_STAR );
// A01 := A01 U11'
U11_STAR_STAR = U11;
LocalTrmm
( RIGHT, UPPER, ADJOINT, diag, F(1), U11_STAR_STAR, A01_VC_STAR );
A01 = A01_VC_STAR;
// A11 := U11 A11 U11'
LocalTwoSidedTrmm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
//--------------------------------------------------------------------//
A01_MC_STAR.FreeAlignments();
A01_MR_STAR.FreeAlignments();
A01_VC_STAR.FreeAlignments();
U01_MC_STAR.FreeAlignments();
U01_MR_STAR.FreeAlignments();
U01_VC_STAR.FreeAlignments();
Y01.FreeAlignments();
Y01_VC_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
}
inline void
internal::TrsvUN
( UnitOrNonUnit diag,
const DistMatrix<F,MC,MR>& U,
DistMatrix<F,MC,MR>& 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,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);
DistMatrix<F,MC,MR>
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> z0_MC_STAR(g);
// Start the algorithm
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( x, xT,
xB, 0 );
while( xT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( xT, x0,
x1,
/**/ /**/
xB, x2 );
x1_MR_STAR.AlignWith( U01 );
z0_MC_STAR.AlignWith( U01 );
z0_MC_STAR.ResizeTo( x0.Height(), 1 );
//----------------------------------------------------------------//
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)0, z0_MC_STAR.LocalMatrix() );
x0.SumScatterUpdate( (F)1, z0_MC_STAR );
//----------------------------------------------------------------//
x1_MR_STAR.FreeAlignments();
z0_MC_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( xT, x0,
/**/ /**/
x1,
xB, x2 );
}
}
else
{
// Matrix views
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);
DistMatrix<F,MC,MR>
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,STAR,MC > z0_STAR_MC(g);
DistMatrix<F,MR, MC > z0_MR_MC(g);
DistMatrix<F,MC, MR > z0(g);
// Start the algorithm
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionLeft( x, xL, xR, 0 );
while( xL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionLeft
( xL, /**/ xR,
x0, x1, /**/ x2 );
x1_STAR_MR.AlignWith( U01 );
z0_STAR_MC.AlignWith( U01 );
z0.AlignWith( x0 );
z0_STAR_MC.ResizeTo( 1, x0.Width() );
//----------------------------------------------------------------//
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)0, z0_STAR_MC.LocalMatrix() );
z0_MR_MC.SumScatterFrom( z0_STAR_MC );
z0 = z0_MR_MC;
Axpy( (F)1, z0, x0 );
//----------------------------------------------------------------//
x1_STAR_MR.FreeAlignments();
z0_STAR_MC.FreeAlignments();
z0.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionLeft
( xL, /**/ xR,
x0, /**/ x1, x2 );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
UVar3( UnitOrNonUnit diag, DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("triangular_inverse::UVar3");
if( U.Height() != U.Width() )
LogicError("Nonsquare matrices cannot be triangular");
#endif
const Grid& g = U.Grid();
// Matrix views
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,VC, STAR> U01_VC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,VR > U12_STAR_VR(g);
DistMatrix<F,STAR,MC > U01Trans_STAR_MC(g);
DistMatrix<F,MR, STAR> U12Trans_MR_STAR(g);
// Start the algorithm
PartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( UBR.Height() < U.Height() )
{
RepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
U01Trans_STAR_MC.AlignWith( U02 );
U12Trans_MR_STAR.AlignWith( U02 );
//--------------------------------------------------------------------//
U01_VC_STAR = U01;
U11_STAR_STAR = U11;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(-1), U11_STAR_STAR, U01_VC_STAR );
// We transpose before the communication to avoid cache-thrashing
// in the unpacking stage.
U12Trans_MR_STAR.TransposeFrom( U12 );
U01Trans_STAR_MC.TransposeFrom( U01_VC_STAR );
LocalGemm
( TRANSPOSE, TRANSPOSE,
F(1), U01Trans_STAR_MC, U12Trans_MR_STAR, F(1), U02 );
U01.TransposeFrom( U01Trans_STAR_MC );
U12_STAR_VR.TransposeFrom( U12Trans_MR_STAR );
LocalTrsm
( LEFT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, U12_STAR_VR );
LocalTriangularInverse( UPPER, diag, U11_STAR_STAR );
U11 = U11_STAR_STAR;
U12 = U12_STAR_VR;
//--------------------------------------------------------------------//
SlidePartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
}
}
inline void
TrsmRUN
( UnitOrNonUnit diag, F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmRUN");
#endif
const Grid& g = U.Grid();
// Matrix views
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);
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,MR > U12_STAR_MR(g);
DistMatrix<F,VC, STAR> X1_VC_STAR(g);
DistMatrix<F,STAR,MC > X1Trans_STAR_MC(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
X1_VC_STAR.AlignWith( X2 );
X1Trans_STAR_MC.AlignWith( X2 );
U12_STAR_MR.AlignWith( X2 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, X1_VC_STAR,
checkIfSingular );
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
U12_STAR_MR = U12;
// X2[MC,MR] -= X1[MC,* ] U12[* ,MR]
// = X1^T[* ,MC] U12[* ,MR]
LocalGemm
( TRANSPOSE, NORMAL, F(-1), X1Trans_STAR_MC, U12_STAR_MR, F(1), X2 );
//--------------------------------------------------------------------//
X1_VC_STAR.FreeAlignments();
X1Trans_STAR_MC.FreeAlignments();
U12_STAR_MR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TwoSidedTrsmUVar1
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar1");
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,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,VC, STAR> U01_VC_STAR(g);
DistMatrix<F,VR, STAR> U01_VR_STAR(g);
DistMatrix<F,STAR,MR > U01Adj_STAR_MR(g);
DistMatrix<F,STAR,STAR> X11_STAR_STAR(g);
DistMatrix<F,MR, MC > Z01_MR_MC(g);
DistMatrix<F,MC, STAR> Z01_MC_STAR(g);
DistMatrix<F,MR, STAR> Z01_MR_STAR(g);
DistMatrix<F> Y01(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 );
A01_VC_STAR.AlignWith( A01 );
U01_MC_STAR.AlignWith( A00 );
U01_VR_STAR.AlignWith( A00 );
U01_VC_STAR.AlignWith( A00 );
U01Adj_STAR_MR.AlignWith( A00 );
Y01.AlignWith( A01 );
Z01_MR_MC.AlignWith( A01 );
Z01_MC_STAR.AlignWith( A00 );
Z01_MR_STAR.AlignWith( A00 );
//--------------------------------------------------------------------//
// Y01 := A00 U01
U01_MC_STAR = U01;
U01_VR_STAR = U01_MC_STAR;
U01Adj_STAR_MR.AdjointFrom( U01_VR_STAR );
Zeros( Z01_MC_STAR, A01.Height(), A01.Width() );
Zeros( Z01_MR_STAR, A01.Height(), A01.Width() );
LocalSymmetricAccumulateLU
( ADJOINT,
F(1), A00, U01_MC_STAR, U01Adj_STAR_MR, Z01_MC_STAR, Z01_MR_STAR );
Z01_MR_MC.SumScatterFrom( Z01_MR_STAR );
Y01 = Z01_MR_MC;
Y01.SumScatterUpdate( F(1), Z01_MC_STAR );
// A01 := inv(U00)' A01
//
// This is the bottleneck because A01 only has blocksize columns
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)
A01_VC_STAR = A01;
U01_VC_STAR = U01_MC_STAR;
Zeros( X11_STAR_STAR, A11.Height(), A11.Width() );
Her2k
( UPPER, ADJOINT,
F(-1), A01_VC_STAR.Matrix(), U01_VC_STAR.Matrix(),
F(0), X11_STAR_STAR.Matrix() );
A11.SumScatterUpdate( F(1), X11_STAR_STAR );
// A11 := inv(U11)' A11 inv(U11)
A11_STAR_STAR = A11;
U11_STAR_STAR = U11;
LocalTwoSidedTrsm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
// A01 := A01 - 1/2 Y01
Axpy( F(-1)/F(2), Y01, A01 );
// A01 := A01 inv(U11)
A01_VC_STAR = A01;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, A01_VC_STAR );
A01 = A01_VC_STAR;
//--------------------------------------------------------------------//
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
TrtrmmUVar1( Orientation orientation, DistMatrix<T>& U )
{
#ifndef RELEASE
PushCallStack("internal::TrtrmmUVar1");
if( U.Height() != U.Width() )
throw std::logic_error("U must be square");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T>
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<T,MC, STAR> U01_MC_STAR(g);
DistMatrix<T,VC, STAR> U01_VC_STAR(g);
DistMatrix<T,VR, STAR> U01_VR_STAR(g);
DistMatrix<T,STAR,MR > U01AdjOrTrans_STAR_MR(g);
DistMatrix<T,STAR,STAR> U11_STAR_STAR(g);
U01_MC_STAR.AlignWith( U );
U01_VC_STAR.AlignWith( U );
U01_VR_STAR.AlignWith( U );
U01AdjOrTrans_STAR_MR.AlignWith( U );
PartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( UTL.Height() < U.Height() && UTL.Width() < U.Height() )
{
RepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
//--------------------------------------------------------------------//
U01_MC_STAR = U01;
U01_VC_STAR = U01_MC_STAR;
U01_VR_STAR = U01_VC_STAR;
if( orientation == ADJOINT )
U01AdjOrTrans_STAR_MR.AdjointFrom( U01_VR_STAR );
else
U01AdjOrTrans_STAR_MR.TransposeFrom( U01_VR_STAR );
LocalTrrk( UPPER, T(1), U01_MC_STAR, U01AdjOrTrans_STAR_MR, T(1), U00 );
U11_STAR_STAR = U11;
LocalTrmm
( RIGHT, UPPER, orientation, NON_UNIT,
T(1), U11_STAR_STAR, U01_VC_STAR );
U01 = U01_VC_STAR;
LocalTrtrmm( orientation, UPPER, U11_STAR_STAR );
U11 = U11_STAR_STAR;
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}