void LLNMedium
( const AbstractDistMatrix<F>& LPre,
AbstractDistMatrix<F>& XPre,
bool checkIfSingular )
{
DEBUG_CSE
const Int m = XPre.Height();
const Int bsize = Blocksize();
const Grid& g = LPre.Grid();
DistMatrixReadProxy<F,F,MC,MR> LProx( LPre );
DistMatrixReadWriteProxy<F,F,MC,MR> XProx( XPre );
auto& L = LProx.GetLocked();
auto& X = XProx.Get();
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,MR, STAR> X1Trans_MR_STAR(g);
for( Int k=0; k<m; k+=bsize )
{
const Int nbProp = Min(bsize,m-k);
const bool in2x2 = ( k+nbProp<m && L.Get(k+nbProp-1,k+nbProp) != F(0) );
const Int nb = ( in2x2 ? nbProp+1 : nbProp );
const Range<Int> ind1( k, k+nb ),
ind2( k+nb, m );
auto L11 = L( ind1, ind1 );
auto L21 = L( ind2, ind1 );
auto X1 = X( ind1, ALL );
auto X2 = X( ind2, ALL );
L11_STAR_STAR = L11; // L11[* ,* ] <- L11[MC,MR]
X1Trans_MR_STAR.AlignWith( X2 );
Transpose( X1, X1Trans_MR_STAR );
// X1^T[MR,* ] := X1^T[MR,* ] L11^-T[* ,* ]
// = (L11^-1[* ,* ] X1[* ,MR])^T
LocalQuasiTrsm
( RIGHT, LOWER, TRANSPOSE,
F(1), L11_STAR_STAR, X1Trans_MR_STAR, checkIfSingular );
Transpose( X1Trans_MR_STAR, X1 );
L21_MC_STAR.AlignWith( X2 );
L21_MC_STAR = L21; // L21[MC,* ] <- L21[MC,MR]
// X2[MC,MR] -= L21[MC,* ] X1[* ,MR]
LocalGemm
( NORMAL, TRANSPOSE, F(-1), L21_MC_STAR, X1Trans_MR_STAR, F(1), X2 );
}
}
void LLNLarge
( const AbstractDistMatrix<F>& LPre,
AbstractDistMatrix<F>& XPre,
bool checkIfSingular )
{
DEBUG_CSE
const Int m = XPre.Height();
const Int bsize = Blocksize();
const Grid& g = LPre.Grid();
DistMatrixReadProxy<F,F,MC,MR> LProx( LPre );
DistMatrixReadWriteProxy<F,F,MC,MR> XProx( XPre );
auto& L = LProx.GetLocked();
auto& X = XProx.Get();
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,STAR,MR > X1_STAR_MR(g);
DistMatrix<F,STAR,VR > X1_STAR_VR(g);
for( Int k=0; k<m; k+=bsize )
{
const Int nbProp = Min(bsize,m-k);
const bool in2x2 = ( k+nbProp<m && L.Get(k+nbProp-1,k+nbProp) != F(0) );
const Int nb = ( in2x2 ? nbProp+1 : nbProp );
const Range<Int> ind1( k, k+nb ),
ind2( k+nb, m );
auto L11 = L( ind1, ind1 );
auto L21 = L( ind2, ind1 );
auto X1 = X( ind1, ALL );
auto X2 = X( ind2, ALL );
// X1[* ,VR] := L11^-1[* ,* ] X1[* ,VR]
L11_STAR_STAR = L11;
X1_STAR_VR = X1;
LocalQuasiTrsm
( LEFT, LOWER, NORMAL, F(1), L11_STAR_STAR, X1_STAR_VR,
checkIfSingular );
X1_STAR_MR.AlignWith( X2 );
X1_STAR_MR = X1_STAR_VR; // X1[* ,MR] <- X1[* ,VR]
X1 = X1_STAR_MR; // X1[MC,MR] <- X1[* ,MR]
L21_MC_STAR.AlignWith( X2 );
L21_MC_STAR = L21; // L21[MC,* ] <- L21[MC,MR]
// X2[MC,MR] -= L21[MC,* ] X1[* ,MR]
LocalGemm( NORMAL, NORMAL, F(-1), L21_MC_STAR, X1_STAR_MR, F(1), X2 );
}
}
inline void
TrmmLLTCOld
( Orientation orientation,
UnitOrNonUnit diag,
T alpha,
const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLLTCOld");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( orientation == NORMAL )
throw std::logic_error("TrmmLLT expects a (Conjugate)Transpose option");
if( L.Height() != L.Width() || L.Height() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLTC: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,MC, STAR> L21_MC_STAR(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> D1AdjOrTrans_MR_STAR(g);
DistMatrix<T,MR, MC > D1AdjOrTrans_MR_MC(g);
DistMatrix<T,MC, MR > D1(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionDown
( X, XT,
XB, 0 );
while( XB.Height() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionDown
( XT, X0,
/**/ /**/
X1,
XB, X2 );
L21_MC_STAR.AlignWith( X2 );
D1AdjOrTrans_MR_STAR.AlignWith( X1 );
D1AdjOrTrans_MR_MC.AlignWith( X1 );
D1.AlignWith( X1 );
Zeros( X1.Width(), X1.Height(), D1AdjOrTrans_MR_STAR );
Zeros( X1.Height(), X1.Width(), D1 );
//--------------------------------------------------------------------//
X1_STAR_VR = X1;
L11_STAR_STAR = L11;
LocalTrmm
( LEFT, LOWER, orientation, diag, T(1), L11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
L21_MC_STAR = L21;
LocalGemm
( orientation, NORMAL,
T(1), X2, L21_MC_STAR, T(0), D1AdjOrTrans_MR_STAR );
D1AdjOrTrans_MR_MC.SumScatterFrom( D1AdjOrTrans_MR_STAR );
if( orientation == TRANSPOSE )
Transpose( D1AdjOrTrans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
else
Adjoint( D1AdjOrTrans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
Axpy( T(1), D1, X1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
D1AdjOrTrans_MR_MC.FreeAlignments();
D1AdjOrTrans_MR_STAR.FreeAlignments();
L21_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionDown
( XT, X0,
X1,
/**/ /**/
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmLLNC
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmLLNC");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( L.Height() != L.Width() || L.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLNC: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,MC, STAR> L21_MC_STAR(g);
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> X1Trans_MR_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
L21_MC_STAR.AlignWith( X2 );
X1Trans_MR_STAR.AlignWith( X2 );
X1_STAR_VR.AlignWith( X1 );
//--------------------------------------------------------------------//
L21_MC_STAR = L21;
X1Trans_MR_STAR.TransposeFrom( X1 );
LocalGemm
( NORMAL, TRANSPOSE, T(1), L21_MC_STAR, X1Trans_MR_STAR, T(1), X2 );
L11_STAR_STAR = L11;
X1_STAR_VR.TransposeFrom( X1Trans_MR_STAR );
LocalTrmm( LEFT, LOWER, NORMAL, diag, T(1), L11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
//--------------------------------------------------------------------//
L21_MC_STAR.FreeAlignments();
X1Trans_MR_STAR.FreeAlignments();
X1_STAR_VR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
}
void LLN
( UnitOrNonUnit diag,
F alpha,
const AbstractDistMatrix<F>& LPre,
AbstractDistMatrix<F>& XPre,
bool checkIfSingular )
{
DEBUG_CSE
const Int n = LPre.Height();
const Int bsize = Blocksize();
const Grid& g = LPre.Grid();
DistMatrixReadProxy<F,F,MC,MR> LProx( LPre );
DistMatrixReadWriteProxy<F,F,MC,MR> XProx( XPre );
auto& L = LProx.GetLocked();
auto& X = XProx.Get();
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g), X11_STAR_STAR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,STAR,MR > X10_STAR_MR(g), X11_STAR_MR(g);
DistMatrix<F,STAR,VR > X10_STAR_VR(g);
ScaleTrapezoid( alpha, LOWER, X );
for( Int k=0; k<n; k+=bsize )
{
const Int nb = Min(bsize,n-k);
const Range<Int> ind0( 0, k ),
ind1( k, k+nb ),
ind2( k+nb, n );
auto L11 = L( ind1, ind1 );
auto L21 = L( ind2, ind1 );
auto X10 = X( ind1, ind0 );
auto X11 = X( ind1, ind1 );
auto X20 = X( ind2, ind0 );
auto X21 = X( ind2, ind1 );
L11_STAR_STAR = L11;
X11_STAR_STAR = X11;
X10_STAR_VR = X10;
LocalTrsm
( LEFT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, X10_STAR_VR,
checkIfSingular );
Trstrm
( LEFT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, X11_STAR_STAR,
checkIfSingular );
X11 = X11_STAR_STAR;
X11_STAR_MR.AlignWith( X21 );
X11_STAR_MR = X11_STAR_STAR;
MakeTrapezoidal( LOWER, X11_STAR_MR );
X10_STAR_MR.AlignWith( X20 );
X10_STAR_MR = X10_STAR_VR;
X10 = X10_STAR_MR;
L21_MC_STAR.AlignWith( X20 );
L21_MC_STAR = L21;
LocalGemm
( NORMAL, NORMAL, F(-1), L21_MC_STAR, X10_STAR_MR, F(1), X20 );
LocalGemm
( NORMAL, NORMAL, F(-1), L21_MC_STAR, X11_STAR_MR, F(1), X21 );
}
}
inline void
TwoSidedTrsmLVar5
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& L )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmLVar5");
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
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,STAR> A11_STAR_STAR(g);
DistMatrix<F,MC, STAR> A21_MC_STAR(g);
DistMatrix<F,VC, STAR> A21_VC_STAR(g);
DistMatrix<F,VR, STAR> A21_VR_STAR(g);
DistMatrix<F,STAR,MR > A21Adj_STAR_MR(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,VC, STAR> L21_VC_STAR(g);
DistMatrix<F,VR, STAR> L21_VR_STAR(g);
DistMatrix<F,STAR,MR > L21Adj_STAR_MR(g);
DistMatrix<F,VC, STAR> Y21_VC_STAR(g);
DistMatrix<F> Y21(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 );
A21_MC_STAR.AlignWith( A22 );
A21_VC_STAR.AlignWith( A22 );
A21_VR_STAR.AlignWith( A22 );
A21Adj_STAR_MR.AlignWith( A22 );
L21_MC_STAR.AlignWith( A22 );
L21_VC_STAR.AlignWith( A22 );
L21_VR_STAR.AlignWith( A22 );
L21Adj_STAR_MR.AlignWith( A22 );
Y21.AlignWith( A21 );
Y21_VC_STAR.AlignWith( A22 );
//--------------------------------------------------------------------//
// A11 := inv(L11) A11 inv(L11)'
L11_STAR_STAR = L11;
A11_STAR_STAR = A11;
LocalTwoSidedTrsm( LOWER, diag, A11_STAR_STAR, L11_STAR_STAR );
A11 = A11_STAR_STAR;
// Y21 := L21 A11
L21_VC_STAR = L21;
Zeros( Y21_VC_STAR, A21.Height(), A21.Width() );
Hemm
( RIGHT, LOWER,
F(1), A11_STAR_STAR.Matrix(), L21_VC_STAR.Matrix(),
F(0), Y21_VC_STAR.Matrix() );
Y21 = Y21_VC_STAR;
// A21 := A21 inv(L11)'
A21_VC_STAR = A21;
LocalTrsm
( RIGHT, LOWER, ADJOINT, diag, F(1), L11_STAR_STAR, A21_VC_STAR );
A21 = A21_VC_STAR;
// A21 := A21 - 1/2 Y21
Axpy( F(-1)/F(2), Y21, A21 );
// A22 := A22 - (L21 A21' + A21 L21')
A21_MC_STAR = A21;
L21_MC_STAR = L21;
A21_VC_STAR = A21_MC_STAR;
A21_VR_STAR = A21_VC_STAR;
L21_VR_STAR = L21_VC_STAR;
A21Adj_STAR_MR.AdjointFrom( A21_VR_STAR );
L21Adj_STAR_MR.AdjointFrom( L21_VR_STAR );
LocalTrr2k
( LOWER,
F(-1), L21_MC_STAR, A21Adj_STAR_MR,
A21_MC_STAR, L21Adj_STAR_MR,
F(1), A22 );
// A21 := A21 - 1/2 Y21
Axpy( F(-1)/F(2), Y21, A21 );
// A21 := inv(L22) A21
//
// This is the bottleneck because A21 only has blocksize columns
Trsm( LEFT, LOWER, NORMAL, diag, F(1), L22, 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 );
}
}
inline void
internal::HegstRLVar3( DistMatrix<F,MC,MR>& A, const DistMatrix<F,MC,MR>& L )
{
#ifndef RELEASE
PushCallStack("internal::HegstRLVar4");
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,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>
YTL(g), YTR(g), Y00(g), Y01(g), Y02(g),
YBL(g), YBR(g), Y10(g), Y11(g), Y12(g),
Y20(g), Y21(g), Y22(g);
DistMatrix<F,MC,MR>
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,MR > A11_STAR_MR(g);
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,VC, STAR> A21_VC_STAR(g);
DistMatrix<F,STAR,VR > A10_STAR_VR(g);
DistMatrix<F,STAR,MR > A10_STAR_MR(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,VR > L10_STAR_VR(g);
DistMatrix<F,STAR,MR > L10_STAR_MR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,STAR,STAR> X11_STAR_STAR(g);
DistMatrix<F,MC, STAR> X21_MC_STAR(g);
DistMatrix<F,MC, STAR> Z21_MC_STAR(g);
// We will use an entire extra matrix as temporary storage. If this is not
// acceptable, use HegstRLVar4 instead.
DistMatrix<F,MC,MR> Y(g);
Y.AlignWith( A );
Y.ResizeTo( A.Height(), A.Width() );
Zero( Y );
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDownDiagonal
( Y, YTL, YTR,
YBL, YBR, 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 );
RepartitionDownDiagonal
( YTL, /**/ YTR, Y00, /**/ Y01, Y02,
/*************/ /******************/
/**/ Y10, /**/ Y11, Y12,
YBL, /**/ YBR, Y20, /**/ Y21, Y22 );
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
A11_STAR_MR.AlignWith( Y21 );
A21_VC_STAR.AlignWith( A21 );
A10_STAR_VR.AlignWith( A10 );
A10_STAR_MR.AlignWith( A10 );
L10_STAR_VR.AlignWith( A10 );
L10_STAR_MR.AlignWith( A10 );
L21_MC_STAR.AlignWith( Y21 );
X21_MC_STAR.AlignWith( A20 );
Z21_MC_STAR.AlignWith( L20 );
//--------------------------------------------------------------------//
// A10 := A10 - 1/2 Y10
Axpy( (F)-0.5, Y10, A10 );
// A11 := A11 - (A10 L10' + L10 A10')
A10_STAR_VR = A10;
L10_STAR_VR = L10;
X11_STAR_STAR.ResizeTo( A11.Height(), A11.Width() );
Her2k
( LOWER, NORMAL,
(F)1, A10_STAR_VR.LocalMatrix(), L10_STAR_VR.LocalMatrix(),
(F)0, X11_STAR_STAR.LocalMatrix() );
MakeTrapezoidal( LEFT, LOWER, 0, X11_STAR_STAR );
A11.SumScatterUpdate( (F)-1, X11_STAR_STAR );
// A11 := inv(L11) A11 inv(L11)'
A11_STAR_STAR = A11;
L11_STAR_STAR = L11;
internal::LocalHegst( RIGHT, LOWER, A11_STAR_STAR, L11_STAR_STAR );
A11 = A11_STAR_STAR;
// A21 := A21 - A20 L10'
L10_STAR_MR = L10_STAR_VR;
X21_MC_STAR.ResizeTo( A21.Height(), A21.Width() );
internal::LocalGemm
( NORMAL, ADJOINT, (F)1, A20, L10_STAR_MR, (F)0, X21_MC_STAR );
A21.SumScatterUpdate( (F)-1, X21_MC_STAR );
// A21 := A21 inv(L11)'
A21_VC_STAR = A21;
internal::LocalTrsm
( RIGHT, LOWER, ADJOINT, NON_UNIT, (F)1, L11_STAR_STAR, A21_VC_STAR );
A21 = A21_VC_STAR;
// A10 := A10 - 1/2 Y10
Axpy( (F)-0.5, Y10, A10 );
// A10 := inv(L11) A10
A10_STAR_VR = A10;
internal::LocalTrsm
( LEFT, LOWER, NORMAL, NON_UNIT,
(F)1, L11_STAR_STAR, A10_STAR_VR );
// Y20 := Y20 + L21 A10
A10_STAR_MR = A10_STAR_VR;
A10 = A10_STAR_MR;
L21_MC_STAR = L21;
internal::LocalGemm
( NORMAL, NORMAL, (F)1, L21_MC_STAR, A10_STAR_MR, (F)1, Y20 );
// Y21 := L21 A11
//
// Symmetrize A11[* ,* ] by copying the lower triangle into the upper
// so that we can call a local gemm instead of worrying about
// reproducing a hemm with nonsymmetric local matrices.
{
const int height = A11_STAR_STAR.LocalHeight();
const int ldim = A11_STAR_STAR.LocalLDim();
F* A11Buffer = A11_STAR_STAR.LocalBuffer();
for( int i=1; i<height; ++i )
for( int j=0; j<i; ++j )
A11Buffer[j+i*ldim] = Conj(A11Buffer[i+j*ldim]);
}
A11_STAR_MR = A11_STAR_STAR;
internal::LocalGemm
( NORMAL, NORMAL, (F)1, L21_MC_STAR, A11_STAR_MR, (F)0, Y21 );
// Y21 := Y21 + L20 A10'
Z21_MC_STAR.ResizeTo( A21.Height(), A21.Width() );
internal::LocalGemm
( NORMAL, ADJOINT, (F)1, L20, A10_STAR_MR, (F)0, Z21_MC_STAR );
Y21.SumScatterUpdate( (F)1, Z21_MC_STAR );
//--------------------------------------------------------------------//
A11_STAR_MR.FreeAlignments();
A21_VC_STAR.FreeAlignments();
A10_STAR_VR.FreeAlignments();
A10_STAR_MR.FreeAlignments();
L10_STAR_VR.FreeAlignments();
L10_STAR_MR.FreeAlignments();
L21_MC_STAR.FreeAlignments();
X21_MC_STAR.FreeAlignments();
Z21_MC_STAR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlidePartitionDownDiagonal
( YTL, /**/ YTR, Y00, Y01, /**/ Y02,
/**/ Y10, Y11, /**/ Y12,
/*************/ /******************/
YBL, /**/ YBR, Y20, Y21, /**/ Y22 );
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/**********************************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrtrsmLLN
( UnitOrNonUnit diag, F alpha, const DistMatrix<F>& L, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrtrsmLLN");
#endif
const Grid& g = L.Grid();
// Matrix views
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);
DistMatrix<F>
XTL(g), XTR(g), X00(g), X01(g), X02(g),
XBL(g), XBR(g), X10(g), X11(g), X12(g),
X20(g), X21(g), X22(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,STAR,MR > X10_STAR_MR(g);
DistMatrix<F,STAR,VR > X10_STAR_VR(g);
DistMatrix<F,STAR,MR > X11_STAR_MR(g);
DistMatrix<F,STAR,STAR> X11_STAR_STAR(g);
// Start the algorithm
ScaleTrapezoid( alpha, LOWER, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionDownDiagonal
( X, XTL, XTR,
XBL, XBR, 0 );
while( XBR.Height() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionDownDiagonal
( XTL, /**/ XTR, X00, /**/ X01, X02,
/*************/ /******************/
/**/ X10, /**/ X11, X12,
XBL, /**/ XBR, X20, /**/ X21, X22 );
L21_MC_STAR.AlignWith( X20 );
X10_STAR_MR.AlignWith( X20 );
X11_STAR_MR.AlignWith( X21 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11;
X11_STAR_STAR = X11;
X10_STAR_VR = X10;
LocalTrsm
( LEFT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, X10_STAR_VR,
checkIfSingular );
LocalTrtrsm
( LEFT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, X11_STAR_STAR,
checkIfSingular );
X11 = X11_STAR_STAR;
X11_STAR_MR = X11_STAR_STAR;
MakeTriangular( LOWER, X11_STAR_MR );
X10_STAR_MR = X10_STAR_VR;
X10 = X10_STAR_MR;
L21_MC_STAR = L21;
LocalGemm
( NORMAL, NORMAL, F(-1), L21_MC_STAR, X10_STAR_MR, F(1), X20 );
LocalGemm
( NORMAL, NORMAL, F(-1), L21_MC_STAR, X11_STAR_MR, F(1), X21 );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionDownDiagonal
( XTL, /**/ XTR, X00, X01, /**/ X02,
/**/ X10, X11, /**/ X12,
/*************/ /******************/
XBL, /**/ XBR, X20, X21, /**/ X22 );
}
}
inline void
TwoSidedTrmmLVar2
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<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
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,STAR> A11_STAR_STAR(g);
DistMatrix<F,VC, STAR> A21_VC_STAR(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,STAR,MR > L21Adj_STAR_MR(g);
DistMatrix<F,VC, STAR> L21_VC_STAR(g);
DistMatrix<F,VR, STAR> L21_VR_STAR(g);
DistMatrix<F,STAR,MR > X10_STAR_MR(g);
DistMatrix<F,STAR,STAR> X11_STAR_STAR(g);
DistMatrix<F,MC, STAR> Z21_MC_STAR(g);
DistMatrix<F,MR, STAR> Z21_MR_STAR(g);
DistMatrix<F,MR, MC > Z21_MR_MC(g);
DistMatrix<F> Y21(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 );
A21_VC_STAR.AlignWith( A22 );
L21_MC_STAR.AlignWith( A20 );
L21_VC_STAR.AlignWith( A22 );
L21_VR_STAR.AlignWith( A22 );
L21Adj_STAR_MR.AlignWith( A22 );
X10_STAR_MR.AlignWith( A10 );
Y21.AlignWith( A21 );
Z21_MC_STAR.AlignWith( A22 );
Z21_MR_STAR.AlignWith( A22 );
//--------------------------------------------------------------------//
// A10 := L11' A10
L11_STAR_STAR = L11;
A10_STAR_VR = A10;
LocalTrmm
( LEFT, LOWER, ADJOINT, diag, F(1), L11_STAR_STAR, A10_STAR_VR );
A10 = A10_STAR_VR;
// A10 := A10 + L21' A20
L21_MC_STAR = L21;
X10_STAR_MR.ResizeTo( A10.Height(), A10.Width() );
LocalGemm( ADJOINT, NORMAL, F(1), L21_MC_STAR, A20, F(0), X10_STAR_MR );
A10.SumScatterUpdate( F(1), X10_STAR_MR );
// Y21 := A22 L21
L21_VC_STAR = L21_MC_STAR;
L21_VR_STAR = L21_VC_STAR;
L21Adj_STAR_MR.AdjointFrom( L21_VR_STAR );
Z21_MC_STAR.ResizeTo( A21.Height(), A21.Width() );
Z21_MR_STAR.ResizeTo( A21.Height(), A21.Width() );
Zero( Z21_MC_STAR );
Zero( Z21_MR_STAR );
LocalSymmetricAccumulateLL
( ADJOINT,
F(1), A22, L21_MC_STAR, L21Adj_STAR_MR, Z21_MC_STAR, Z21_MR_STAR );
Z21_MR_MC.SumScatterFrom( Z21_MR_STAR );
Y21 = Z21_MR_MC;
Y21.SumScatterUpdate( F(1), Z21_MC_STAR );
// A21 := A21 L11
A21_VC_STAR = A21;
LocalTrmm
( RIGHT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, A21_VC_STAR );
A21 = A21_VC_STAR;
// A21 := A21 + 1/2 Y21
Axpy( F(1)/F(2), Y21, A21 );
// A11 := L11' A11 L11
A11_STAR_STAR = A11;
LocalTwoSidedTrmm( LOWER, diag, A11_STAR_STAR, L11_STAR_STAR );
A11 = A11_STAR_STAR;
// A11 := A11 + (A21' L21 + L21' A21)
A21_VC_STAR = A21;
X11_STAR_STAR.ResizeTo( A11.Height(), A11.Width() );
Her2k
( LOWER, ADJOINT,
F(1), A21_VC_STAR.LocalMatrix(), L21_VC_STAR.LocalMatrix(),
F(0), X11_STAR_STAR.LocalMatrix() );
A11.SumScatterUpdate( F(1), X11_STAR_STAR );
// A21 := A21 + 1/2 Y21
Axpy( F(1)/F(2), Y21, A21 );
//--------------------------------------------------------------------//
A21_VC_STAR.FreeAlignments();
L21_MC_STAR.FreeAlignments();
L21_VC_STAR.FreeAlignments();
L21_VR_STAR.FreeAlignments();
L21Adj_STAR_MR.FreeAlignments();
X10_STAR_MR.FreeAlignments();
Y21.FreeAlignments();
Z21_MC_STAR.FreeAlignments();
Z21_MR_STAR.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
}
inline void
TriangularInverseLVar3( UnitOrNonUnit diag, DistMatrix<F>& L )
{
#ifndef RELEASE
PushCallStack("internal::TriangularInverseLVar3");
if( L.Height() != L.Width() )
throw std::logic_error("Nonsquare matrices cannot be triangular");
#endif
const Grid& g = L.Grid();
// Matrix views
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,MR > L10_STAR_MR(g);
DistMatrix<F,STAR,VR > L10_STAR_VR(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> L21_MC_STAR(g);
DistMatrix<F,VC, STAR> L21_VC_STAR(g);
// Start the algorithm
PartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
while( LTL.Height() < L.Height() )
{
RepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
L10_STAR_MR.AlignWith( L20 );
L21_MC_STAR.AlignWith( L20 );
//--------------------------------------------------------------------//
L10_STAR_VR = L10;
L11_STAR_STAR = L11;
LocalTrsm
( LEFT, LOWER, NORMAL, diag, F(-1), L11_STAR_STAR, L10_STAR_VR );
L21_MC_STAR = L21;
L10_STAR_MR = L10_STAR_VR;
LocalGemm
( NORMAL, NORMAL, F(1), L21_MC_STAR, L10_STAR_MR, F(1), L20 );
L10 = L10_STAR_MR;
L21_VC_STAR = L21_MC_STAR;
LocalTrsm
( RIGHT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, L21_VC_STAR );
LocalTriangularInverse( LOWER, diag, L11_STAR_STAR );
L11 = L11_STAR_STAR;
L21 = L21_VC_STAR;
//--------------------------------------------------------------------//
L10_STAR_MR.FreeAlignments();
L21_MC_STAR.FreeAlignments();
SlidePartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
