inline void
internal::HegstLLVar4( DistMatrix<F,MC,MR>& A, const DistMatrix<F,MC,MR>& L )
{
#ifndef RELEASE
PushCallStack("internal::HegstLLVar4");
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>
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,STAR,MR > L10_STAR_MR(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 );
L10_STAR_MR.AlignWith( A00 );
L10_STAR_MC.AlignWith( A00 );
Y10_STAR_VR.AlignWith( A10 );
//--------------------------------------------------------------------//
// Y10 := A11 L10
A11_STAR_STAR = A11;
L10_STAR_VR = L10;
Y10_STAR_VR.ResizeTo( A10.Height(), A10.Width() );
Zero( Y10_STAR_VR );
Hemm
( LEFT, LOWER,
(F)0.5, 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, 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_MR = L10_STAR_VR;
L10_STAR_MC = L10_STAR_VR;
internal::LocalTrr2k
( LOWER, ADJOINT, ADJOINT,
(F)1, A10_STAR_MC, L10_STAR_MR,
L10_STAR_MC, A10_STAR_MR,
(F)1, A00 );
// A10 := A10 + 1/2 Y10
Axpy( (F)1, Y10_STAR_VR, A10_STAR_VR );
// A10 := L11' A10
L11_STAR_STAR = L11;
internal::LocalTrmm
( LEFT, LOWER, ADJOINT, NON_UNIT, (F)1, L11_STAR_STAR, A10_STAR_VR );
A10 = A10_STAR_VR;
// A20 := A20 + A21 L10
A21_MC_STAR = A21;
internal::LocalGemm
( NORMAL, NORMAL, (F)1, A21_MC_STAR, L10_STAR_MR, (F)1, A20 );
// A11 := L11' A11 L11
internal::LocalHegst
( LEFT, LOWER, A11_STAR_STAR, L11_STAR_STAR );
A11 = A11_STAR_STAR;
// A21 := A21 L11
A21_VC_STAR = A21_MC_STAR;
internal::LocalTrmm
( RIGHT, LOWER, NORMAL, NON_UNIT, (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();
L10_STAR_MR.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
}
inline void
TwoSidedTrmmLVar4
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& L )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrmmLVar4");
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,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 );
Zeros( Y10_STAR_VR, A10.Height(), A10.Width() );
Hemm
( LEFT, LOWER,
F(1), A11_STAR_STAR.LockedMatrix(), L10_STAR_VR.LockedMatrix(),
F(0), Y10_STAR_VR.Matrix() );
// 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;
//--------------------------------------------------------------------//
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
TrsvLT
( Orientation orientation, UnitOrNonUnit diag,
const DistMatrix<F>& L, DistMatrix<F>& x )
{
#ifndef RELEASE
PushCallStack("internal::TrsvLT");
if( L.Grid() != x.Grid() )
throw std::logic_error("{L,x} must be distributed over the same grid");
if( orientation == NORMAL )
throw std::logic_error("TrsvLT expects a (conjugate-)transpose option");
if( L.Height() != L.Width() )
throw std::logic_error("L 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( L.Width() != xLength )
throw std::logic_error("Nonconformal TrsvLT");
#endif
const Grid& g = L.Grid();
if( x.Width() == 1 )
{
// Matrix views
DistMatrix<F> L10(g), L11(g);
DistMatrix<F>
xT(g), x0(g),
xB(g), x1(g),
x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,MC, STAR> x1_MC_STAR(g);
DistMatrix<F,MC, MR > z1(g);
DistMatrix<F,MR, MC > z1_MR_MC(g);
DistMatrix<F,MR, STAR> z_MR_STAR(g);
// Views of z[MR,* ]
DistMatrix<F,MR,STAR> z0_MR_STAR(g),
z1_MR_STAR(g);
z_MR_STAR.AlignWith( L );
Zeros( x.Height(), 1, z_MR_STAR );
// Start the algorithm
PartitionUp
( x, xT,
xB, 0 );
while( xT.Height() > 0 )
{
RepartitionUp
( xT, x0,
x1,
/**/ /**/
xB, x2 );
const int n0 = x0.Height();
const int n1 = x1.Height();
LockedView( L10, L, n0, 0, n1, n0 );
LockedView( L11, L, n0, n0, n1, n1 );
View( z0_MR_STAR, z_MR_STAR, 0, 0, n0, 1 );
View( z1_MR_STAR, z_MR_STAR, n0, 0, n1, 1 );
x1_MC_STAR.AlignWith( L10 );
z1.AlignWith( x1 );
//----------------------------------------------------------------//
if( x2.Height() != 0 )
{
z1_MR_MC.SumScatterFrom( z1_MR_STAR );
z1 = z1_MR_MC;
Axpy( F(1), z1, x1 );
}
x1_STAR_STAR = x1;
L11_STAR_STAR = L11;
Trsv
( LOWER, orientation, diag,
L11_STAR_STAR.LockedMatrix(),
x1_STAR_STAR.Matrix() );
x1 = x1_STAR_STAR;
x1_MC_STAR = x1_STAR_STAR;
Gemv
( orientation, F(-1),
L10.LockedMatrix(),
x1_MC_STAR.LockedMatrix(),
F(1), z0_MR_STAR.Matrix() );
//----------------------------------------------------------------//
x1_MC_STAR.FreeAlignments();
z1.FreeAlignments();
SlidePartitionUp
( xT, x0,
/**/ /**/
x1,
xB, x2 );
}
}
else
{
// Matrix views
DistMatrix<F> L10(g), L11(g);
DistMatrix<F>
xL(g), xR(g),
x0(g), x1(g), x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,STAR,MC > x1_STAR_MC(g);
DistMatrix<F,STAR,MR > z_STAR_MR(g);
// Views of z[* ,MR], which will store updates to x
DistMatrix<F,STAR,MR> z0_STAR_MR(g),
z1_STAR_MR(g);
z_STAR_MR.AlignWith( L );
Zeros( 1, x.Width(), z_STAR_MR );
// Start the algorithm
PartitionLeft( x, xL, xR, 0 );
while( xL.Width() > 0 )
{
RepartitionLeft
( xL, /**/ xR,
x0, x1, /**/ x2 );
const int n0 = x0.Width();
const int n1 = x1.Width();
LockedView( L10, L, n0, 0, n1, n0 );
LockedView( L11, L, n0, n0, n1, n1 );
View( z0_STAR_MR, z_STAR_MR, 0, 0, 1, n0 );
View( z1_STAR_MR, z_STAR_MR, 0, n0, 1, n1 );
x1_STAR_MC.AlignWith( L10 );
//----------------------------------------------------------------//
if( x2.Width() != 0 )
x1.SumScatterUpdate( F(1), z1_STAR_MR );
x1_STAR_STAR = x1;
L11_STAR_STAR = L11;
Trsv
( LOWER, orientation, diag,
L11_STAR_STAR.LockedMatrix(),
x1_STAR_STAR.Matrix() );
x1 = x1_STAR_STAR;
x1_STAR_MC = x1_STAR_STAR;
Gemv
( orientation, F(-1),
L10.LockedMatrix(),
x1_STAR_MC.LockedMatrix(),
F(1), z0_STAR_MR.Matrix() );
//----------------------------------------------------------------//
x1_STAR_MC.FreeAlignments();
SlidePartitionLeft
( xL, /**/ xR,
x0, /**/ x1, x2 );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrsmRLT
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& L, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrsmRLT");
if( orientation == NORMAL )
LogicError("TrsmRLT expects a (Conjugate)Transpose option");
#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> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,VR, STAR> L21_VR_STAR(g);
DistMatrix<F,STAR,MR > L21AdjOrTrans_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
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
X1_VC_STAR.AlignWith( X2 );
X1Trans_STAR_MC.AlignWith( X2 );
L21_VR_STAR.AlignWith( X2 );
L21AdjOrTrans_STAR_MR.AlignWith( X2 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, LOWER, orientation, diag,
F(1), L11_STAR_STAR, X1_VC_STAR, checkIfSingular );
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
L21_VR_STAR = L21;
if( orientation == ADJOINT )
L21AdjOrTrans_STAR_MR.AdjointFrom( L21_VR_STAR );
else
L21AdjOrTrans_STAR_MR.TransposeFrom( L21_VR_STAR );
// X2[MC,MR] -= X1[MC,*] (L21[MR,*])^(T/H)
// = X1^T[* ,MC] (L21^(T/H))[*,MR]
LocalGemm
( TRANSPOSE, NORMAL,
F(-1), X1Trans_STAR_MC, L21AdjOrTrans_STAR_MR, F(1), X2 );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
}
inline void
LocalTrmmAccumulateLLT
( Orientation orientation, UnitOrNonUnit diag, T alpha,
const DistMatrix<T>& L,
const DistMatrix<T,MC,STAR>& X_MC_STAR,
DistMatrix<T,MR,STAR>& Z_MR_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalTrmmAccumulateLLT");
if( L.Grid() != X_MC_STAR.Grid() ||
X_MC_STAR.Grid() != Z_MR_STAR.Grid() )
throw std::logic_error
("{L,X,Z} must be distributed over the same grid");
if( L.Height() != L.Width() ||
L.Height() != X_MC_STAR.Height() ||
L.Height() != Z_MR_STAR.Height() )
{
std::ostringstream msg;
msg << "Nonconformal LocalTrmmAccumulateLLT: " << "\n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X[MC,* ] ~ " << X_MC_STAR.Height() << " x "
<< X_MC_STAR.Width() << "\n"
<< " Z[MR,* ] ` " << Z_MR_STAR.Height() << " x "
<< Z_MR_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( X_MC_STAR.ColAlignment() != L.ColAlignment() ||
Z_MR_STAR.ColAlignment() != L.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#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> D11(g);
DistMatrix<T,MC,STAR>
XT_MC_STAR(g), X0_MC_STAR(g),
XB_MC_STAR(g), X1_MC_STAR(g),
X2_MC_STAR(g);
DistMatrix<T,MR,STAR>
ZT_MR_STAR(g), Z0_MR_STAR(g),
ZB_MR_STAR(g), Z1_MR_STAR(g),
Z2_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
LockedPartitionDown
( X_MC_STAR, XT_MC_STAR,
XB_MC_STAR, 0 );
PartitionDown
( Z_MR_STAR, ZT_MR_STAR,
ZB_MR_STAR, 0 );
while( LTL.Height() < L.Height() )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
LockedRepartitionDown
( XT_MC_STAR, X0_MC_STAR,
/**********/ /**********/
X1_MC_STAR,
XB_MC_STAR, X2_MC_STAR );
RepartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
/**********/ /**********/
Z1_MR_STAR,
ZB_MR_STAR, Z2_MR_STAR );
D11.AlignWith( L11 );
//--------------------------------------------------------------------//
D11 = L11;
MakeTrapezoidal( LEFT, LOWER, 0, D11 );
if( diag == UNIT )
SetDiagonalToOne( D11 );
LocalGemm
( orientation, NORMAL, alpha, D11, X1_MC_STAR, T(1), Z1_MR_STAR );
LocalGemm
( orientation, NORMAL, alpha, L21, X2_MC_STAR, T(1), Z1_MR_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlideLockedPartitionDown
( XT_MC_STAR, X0_MC_STAR,
X1_MC_STAR,
/**********/ /**********/
XB_MC_STAR, X2_MC_STAR );
SlidePartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
Z1_MR_STAR,
/**********/ /**********/
ZB_MR_STAR, Z2_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
L00407C8F(
_unknown_ __eax, // r0
_unknown_ __ebx, // r1
_unknown_ __ecx, // r2
_unknown_ __edx, // r3
_unknown_ __edi, // r4
signed int __esi // r5
)
{
_unknown_ _t779; // _t779
_unknown_ _t780; // _t780
_unknown_ _t783; // _t783
_unknown_ _t784; // _t784
_unknown_ _t785; // _t785
_unknown_ _t786; // _t786
_unknown_ _t787; // _t787
_unknown_ _t792; // _t792
_unknown_ _t794; // _t794
_unknown_ _t796; // _t796
_unknown_ _t797; // _t797
_unknown_ _t798; // _t798
_unknown_ _t799; // _t799
_unknown_ _t800; // _t800
_unknown_ _t801; // _t801
_unknown_ _t802; // _t802
_unknown_ _t803; // _t803
_unknown_ _t805; // _t805
_unknown_ _t806; // _t806
_unknown_ _t807; // _t807
_unknown_ _t808; // _t808
_unknown_ _t810; // _t810
_unknown_ _t811; // _t811
_unknown_ _t812; // _t812
_unknown_ _t813; // _t813
_unknown_ _t815; // _t815
_unknown_ _t816; // _t816
_unknown_ _t817; // _t817
_unknown_ _t818; // _t818
_unknown_ _t823; // _t823
_unknown_ _t824; // _t824
signed int _t827; // _t827
_unknown_ _t828; // _t828
_unknown_ _t829; // _t829
_unknown_ _t831; // _t831
_unknown_ _t832; // _t832
_unknown_ _t834; // _t834
_unknown_ _t838; // _t838
_unknown_ _t840; // _t840
_unknown_ _t842; // _t842
_unknown_ _t843; // _t843
_unknown_ _t844; // _t844
_unknown_ _t846; // _t846
_unknown_ _t847; // _t847
_unknown_ _t848; // _t848
_unknown_ _t849; // _t849
_unknown_ _t850; // _t850
_unknown_ _t851; // _t851
_unknown_ _t853; // _t853
_unknown_ _t854; // _t854
L0:
while(1) {
_t809 = __edi;
__edx = __edx;
_t779 = __eax;
asm("sal ecx, cl");
asm("rcl ch, 0x1");
asm("sbb bl, dh");
_t827 = __esi << 12;
asm("sbb ebx, 0xc0761413");
_t794 = __ebx + 1459406168;
if(_t779 != -1204393742) {
goto L8;
} else {
goto L14;
}
}
L14:
asm("stc ");
asm("pushfd ");
goto L15;
L15:
_t810 = _t809 | -487951012;
asm("adc al, bl");
asm("rcr bl, 0x19");
_t779 = ~(_t779 << 24) >> (_t808 & 255);
asm("rol bx, cl");
__eflags = _t795 - -1256608298;
if(__eflags != 0) {
goto L21;
} else {
goto L16;
}
L16:
__esp = __ebp;
_pop(__ebp);
if(__eflags >= 0) {
goto L18;
} else {
goto L17;
}
L17:
_push(_t806);
asm("sal bh, 0x18");
_t784 = _t779 - _t806;
__eflags = _t784 & 65535 & __edx & 65535;
asm("rol dl, cl");
asm("sbb al, dh");
asm("ror dl, cl");
goto L51;
L51:
asm("sbb bl, 0x74");
_t808 = _t806 + -2883;
asm("sal ah, 0xd");
__eflags = _t794 - -1025494954;
if(_t794 == -1025494954) {
L52:
asm("lodsd ");
asm("fsub qword [ebp+0x25]");
__eflags = _t784 - 748244411;
}
goto L53;
L53:
_t795 = 36477261;
_t785 = _t784 >> (_t808 & 255);
asm("ror bx, cl");
asm("sbb edi, ecx");
__eflags = _t810 - 411730992;
goto L54;
L54:
_push( *((intOrPtr*)(_t785)));
__eflags = *((intOrPtr*)(__edx + -108720360)) & _t808 & 255;
*((intOrPtr*)(_t785)) = *((intOrPtr*)(_t785)) + _t785;
*((intOrPtr*)(_t785)) = *((intOrPtr*)(_t785)) + (_t808 & 255);
L55:
asm("int 0x28");
goto L56;
L56:
asm("sbb eax, edi");
asm("rcl al, 0x19");
_t796 = _t795 >> (_t808 & 255);
_t786 = _t785 >> 23;
asm("rcl eax, 0x18");
asm("sbb bh, 0x52");
goto L80;
L80:
_t811 = _t810 - 1;
__eflags = _t786 & -492243883;
asm("rcr bh, 0x19");
__eflags = _t786 & 0x2e;
goto L3;
L3:
asm("sbb ah, ch");
__eflags = _t796 & 0x6b;
asm("sal eax, 0x15");
asm("ror bl, cl");
goto L44;
L44:
asm("sal ax, cl");
asm("sbb bh, dl");
*((intOrPtr*)(_t808)) = __edx & 255;
asm("adc bl, 0x91");
goto L46;
L46:
_t779 = _t786 - 1891155626;
asm("sbb bl, 0x78");
asm("rcr ebx, 0x1d");
asm("rcl edi, 0xb");
asm("sbb eax, 0xd74ad4d3");
goto L72;
L72:
asm("adc di, bx");
_t797 = _t796 ^ 337739294;
asm("rcl ah, cl");
asm("rcr ebx, 0x16");
asm("sbb ebx, edi");
_t810 = _t811 & -527781969;
if(_t808 != 367264678) {
goto L23;
} else {
goto L73;
}
L73:
*((intOrPtr*)(_t831 + 299155711)) = *((intOrPtr*)(_t831 + 299155711)) - (_t797 & 255);
goto L74;
L74:
asm("rcl ah, 0x11");
asm("rcl bx, cl");
asm("rcl al, 0x17");
goto L6;
L6:
asm("adc al, dh");
_t809 = -451263470;
_t794 = -154469424;
if(__edx != 1078422774) {
goto L25;
} else {
goto L7;
}
L7:
goto L8;
L8:
_t828 = _t827 & -390891455;
L9:
1376606914();
_t806 = __esp;
_push(_t806);
_push(64);
L10();
goto L10;
L10:
_pop(__eax);
goto L35;
L35:
goto L36;
L36:
_pop(__edx);
*((intOrPtr*)(_t809)) = *((intOrPtr*)(_t809)) + (_t806 & 255);
asm("std ");
asm("invalid ");
L37:
_t806 = _t779;
_t780 = _t779 + *((intOrPtr*)(_t779 + 60));
_push( *((intOrPtr*)(_t780 + 80)));
_push(_t806);
_t779 = _t780 + 128;
_t844 = _t809 - 598847696;
if(_t844 != 0) {
goto L28;
} else {
goto L38;
}
L38:
if(_t844 != 0) {
goto L36;
} else {
goto L39;
}
L39:
_pop(__esi);
if(_t844 > 0) {
goto L32;
} else {
goto L40;
}
L40:
asm("salc ");
goto L41;
L41:
asm("adc edx, 0xfbed735a");
asm("ror bh, cl");
asm("sal ax, cl");
_t810 = _t809 >> 24;
asm("adc edi, 0xa556f6d2");
_t831 = (_t829 ^ 2084378932) & 1448525259;
if(_t806 != 1356731863) {
goto L17;
} else {
goto L42;
}
L42:
asm("sbb bl, 0x87");
asm("cwde ");
_t797 = -1010795954;
goto L43;
L43:
asm("rol bl, 0x16");
asm("adc ah, ch");
asm("adc ah, 0xb2");
asm("sbb ah, al");
_t798 = _t797 & _t831;
goto L49;
L49:
asm("ror edi, cl");
_t831 = _t831 - 1;
asm("sbb ebx, edx");
asm("rcr bl, cl");
_t812 = _t810 | -1932129093;
goto L22;
L22:
asm("rol eax, cl");
asm("rol al, 0xd");
_t799 = _t798 + 251010009;
goto L57;
L57:
asm("sal bh, 0x1c");
_t800 = _t799 >> (_t808 & 255);
_t787 = _t831;
if(__edx != -1607147456) {
goto L50;
} else {
goto L58;
}
L58:
asm("cli ");
asm("popad ");
asm("adc [edx+0xc065bafb], edi");
goto L59;
L59:
__edx = -1067074821;
asm("sbb cl, al");
asm("rcr eax, cl");
_t810 = __edx - _t831;
_t785 = _t779 << 14;
_t854 = _t808 - -1408886173;
if(_t854 != 0) {
goto L66;
} else {
goto L60;
}
L60:
if(_t854 <= 0) {
goto L54;
} else {
goto L61;
}
L61:
goto L62;
L62:
asm("sal edi, 0xf");
_t809 = _t823 ^ 764803179;
asm("rol bh, cl");
asm("sbb ax, si");
if(_t831 != 1788972044) {
goto L75;
} else {
goto L63;
}
L63:
asm("ficom word [esi+0x66]");
goto L64;
L64:
asm("rol bl, cl");
asm("adc ebx, 0xc570f7ae");
asm("sbb bl, bh");
asm("adc di, dx");
asm("rcl edi, 0xa");
asm("adc edi, 0x1157ea8");
goto L69;
L69:
asm("rol edi, 0x3");
_t809 = _t813 << (_t808 & 255) | _t808;
asm("sbb ebx, edx");
if(_t779 != 1935422773) {
goto L30;
} else {
goto L70;
}
L70:
_t779 = !_t779;
if(_t809 != -1556628160) {
goto L26;
} else {
goto L71;
}
L71:
asm("int 0x35");
goto L72;
L26:
__eflags = _t795 & 65535 & __edx & 65535;
__edx = __edx >> 26;
asm("rcl edi, 0x7");
_t779 = _t779 & _t831;
__eflags = _t808 - -250084760;
if(_t808 != -250084760) {
goto L59;
} else {
goto L27;
}
L27:
asm("scasd ");
asm("fscale ");
asm("invalid ");
goto L28;
L28:
_t783 = *((intOrPtr*)( *((intOrPtr*)(_t779)) + _t806 + 16));
*((intOrPtr*)(_t806 + _t783))();
_pop(__ecx);
_pop(__ecx);
asm("rcl dh, 0x12");
_t809 = _t809 >> 20;
asm("adc esi, ebx");
goto L45;
L45:
asm("sbb dh, 0xe9");
asm("adc dl, 0xbc");
asm("sbb edx, 0xa9309b8b");
asm("rol dl, cl");
__eflags = _t783 & 255 & _t783 & 255;
_t779 = _t783 - -1447259429;
_t794 = ~_t794;
goto L77;
L77:
_t806 = 0x779b97;
_t829 = _t828 >> (_t806 & 255);
asm("rcr bh, 0x1c");
asm("ror bl, cl");
__eflags = _t829 & -1765348044;
asm("sbb ebx, 0x805f12cc");
asm("rcr eax, cl");
goto L41;
L30:
__edx = __edx >> (_t808 & 255);
asm("rcl ax, cl");
asm("rcl bl, cl");
asm("rol ax, cl");
goto L24;
L24:
_t815 = _t809 << (_t808 & 255);
_t794 = -1779828323;
goto L31;
L31:
asm("sal eax, 0x13");
_t809 = _t815 + 362140062;
asm("sal eax, cl");
goto L32;
L32:
asm("lock mov di, dx");
asm("sbb ebx, 0x21877791");
__eflags = __edx - 651418901;
if(__edx != 651418901) {
goto L2;
} else {
goto L33;
}
L33:
asm("fsub dword [ss:edi]");
_t794 = _t794 + 1;
__esp = -1048539;
goto L34;
L34:
_t779 = _t779 & -4096;
goto L35;
L2:
asm("rcr bl, cl");
asm("ror ebx, 0x19");
asm("ror dh, 0x1a");
_t779 = (_t779 | 1339361818) >> (_t808 & 255);
asm("ror bl, 0x10");
L65:
asm("sal bh, 0x10");
asm("rcr edi, cl");
_t797 = ~_t794;
asm("rcl eax, 0x10");
goto L74;
L75:
_t795 = _t795 + 374830029;
asm("adc edi, 0xe4617f53");
asm("ror bl, 0x1a");
_t779 = _t779 + 1;
__eflags = _t779 - 1288929701;
if(_t779 != 1288929701) {
goto L15;
} else {
goto L76;
}
L76:
_pop(ds);
asm("movsb ");
goto L77;
L66:
asm("sal ah, 0x5");
asm("rcr dx, cl");
asm("adc al, 0xff");
asm("rol esi, cl");
goto L79;
L79:
asm("rcr bl, cl");
return;
goto L82;
L82:
L50:
_t813 = _t812 & -2047814454;
_t795 = _t800 + __edx;
asm("sal ebx, 0x14");
asm("sbb bh, ah");
_t779 = _t787 - -1935422773;
asm("rol bh, cl");
goto L64;
L11:
__eax = __eax - 1;
goto L34;
L5:
asm("daa ");
_t779 = _t779 - -104532248;
asm("rcl ax, cl");
asm("sbb bx, cx");
goto L29;
L29:
_t809 = _t809 | -1152305042;
goto L30;
L25:
asm("adc ax, cx");
asm("ror al, cl");
_t795 = ~_t794;
asm("rcr ebx, cl");
goto L67;
L67:
asm("rcl bh, 0x5");
asm("ror ah, cl");
_t810 = (-451263470 | _t831) - 1;
__eflags = _t831 - 1336743640;
if(_t831 != 1336743640) {
goto L47;
} else {
goto L68;
}
L68:
_t795 = _t795 - 1;
_t779 = _t779 - 1501076930;
__eflags = _t795 & 255 & __edx & 255;
goto L69;
L47:
asm("rcl al, 0x9");
asm("rcr bl, cl");
asm("rol dh, 0x16");
asm("rcl di, cl");
asm("sal bh, 0x1");
asm("sal al, 0x9");
asm("sal bx, cl");
__eflags = _t810 - 2010318303;
if(_t810 != 2010318303) {
goto L81;
} else {
goto L48;
}
L48:
asm("adc cl, [edi+0x79]");
asm("popad ");
goto L49;
L81:
asm("rcl bl, 0x14");
asm("adc al, dh");
_t809 = _t810 << 12 ^ -2100988332;
asm("adc al, bh");
__eflags = _t779 - 51549907;
if (_t779 != 51549907) goto L1;
goto L1;
L1:
asm("adc al, bh");
_t779 = (_t779 | -1641774524) >> 23;
_t823 = (_t809 & -990957014 ^ 937748571) >> (_t808 & 255);
asm("rol al, 0xe");
goto L62;
L23:
_t808 = _t808 + 1;
asm("sbb bh, 0xf9");
asm("rcl eax, cl");
asm("rol di, cl");
__eflags = _t797 - -525952445;
asm("rcr edi, 0x19");
goto L43;
L78:
__ebx = __ebx | __esi;
asm("rol bh, 0x19");
__esi = -1112598767;
__eflags = __edx & 65535 & __esi & 65535;
asm("rol eax, 0x6");
asm("rcr bl, 0x11");
asm("adc dh, dl");
goto L19;
L19:
_t779 = _t792 << 30;
__edx = __edx | 1630326992;
asm("sbb ax, di");
_t795 = -1948230174;
asm("sal edi, 0x9");
__eflags = _t808 - -2003135051;
if(_t808 != -2003135051) {
goto L12;
} else {
goto L20;
}
L20:
asm("sbb esp, ebx");
*((intOrPtr*)(_t831 + -956054714)) = __esp;
goto L21;
L21:
_t785 = _t779 + _t810;
asm("ror eax, cl");
asm("rol bl, 0x17");
asm("rol edi, 0x14");
goto L56;
L12:
_t831 = _t831 ^ -1112597252;
_t779 = _t779 & _t831;
__eflags = _t810 & 65535 & _t779 & 65535;
asm("sbb bl, cl");
asm("sbb ah, 0xcd");
_t809 = 538325363;
asm("sbb bl, dl");
__eflags = _t831 - 1654797031;
if(_t831 != 1654797031) {
goto L4;
} else {
goto L13;
}
L13:
*((intOrPtr*)(_t779)) = *((intOrPtr*)(_t779)) >> 1;
asm("xlatb ");
*((intOrPtr*)(-755903498)) = *((intOrPtr*)(-755903498)) & _t795;
goto L0;
L4:
__edx = __edx ^ 2032913572;
_t795 = 0xfffffffffbf9dfea;
__eflags = _t795 & -399694041;
goto L5;
L18:
_t792 = _t779 - *((intOrPtr*)(__esp + _t795 * 4 - 77));
_t808 = _t808 + _t792;
goto L19;
}
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
TrsmLLTSmall
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F,STAR,VR>& L, DistMatrix<F,VR,STAR>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLLTSmall");
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("TrsmLLT expects a (Conjugate)Transpose option");
if( L.Height() != L.Width() || L.Height() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrsmLLT: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( L.RowAlignment() != X.ColAlignment() )
throw std::logic_error("L and X must be aligned");
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<F,STAR,VR>
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,VR,STAR> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> X1_STAR_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 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11; // L11[* ,* ] <- L11[* ,VR]
X1_STAR_STAR = X1; // X1[* ,* ] <- X1[VR,* ]
// X1[* ,* ] := L11^-[T/H][* ,* ] X1[* ,* ]
LocalTrsm
( LEFT, LOWER, orientation, diag,
F(1), L11_STAR_STAR, X1_STAR_STAR, checkIfSingular );
X1 = X1_STAR_STAR;
// X0[VR,* ] -= L10[* ,VR]^(T/H) X1[* ,* ]
LocalGemm( orientation, NORMAL, F(-1), L10, X1_STAR_STAR, F(1), X0 );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrsmRLN
( UnitOrNonUnit diag, F alpha, const DistMatrix<F>& L, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmRLN");
#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> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<F,MR, STAR> L10Trans_MR_STAR(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,MC > X1Trans_STAR_MC(g);
DistMatrix<F,VC, STAR> X1_VC_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionLeft( X, XL, XR, 0 );
while( XL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
RepartitionLeft
( XL, /**/ XR,
X0, X1, /**/ X2 );
X1Trans_STAR_MC.AlignWith( X0 );
L10Trans_MR_STAR.AlignWith( X0 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, X1_VC_STAR,
checkIfSingular );
// X0[MC,MR] -= X1[MC,* ] L10[*,MR]
// = X1^T[* ,MC] L10^T[MR,* ]
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
L10Trans_MR_STAR.TransposeFrom( L10 );
LocalGemm
( TRANSPOSE, TRANSPOSE,
F(-1), X1Trans_STAR_MC, L10Trans_MR_STAR, F(1), X0 );
//--------------------------------------------------------------------//
X1Trans_STAR_MC.FreeAlignments();
L10Trans_MR_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionLeft
( XL, /**/ XR,
X0, /**/ X1, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
internal::TrsvLN
( UnitOrNonUnit diag,
const DistMatrix<F,MC,MR>& L,
DistMatrix<F,MC,MR>& x )
{
#ifndef RELEASE
PushCallStack("internal::TrsvLN");
if( L.Grid() != x.Grid() )
throw std::logic_error("{L,x} must be distributed over the same grid");
if( L.Height() != L.Width() )
throw std::logic_error("L 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( L.Width() != xLength )
throw std::logic_error("Nonconformal TrsvLN");
#endif
const Grid& g = L.Grid();
if( x.Width() == 1 )
{
// Matrix views
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);
DistMatrix<F,MC,MR>
xT(g), x0(g),
xB(g), x1(g),
x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,MR, STAR> x1_MR_STAR(g);
DistMatrix<F,MC, STAR> z2_MC_STAR(g);
// Start the algorithm
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 );
x1_MR_STAR.AlignWith( L21 );
z2_MC_STAR.AlignWith( L21 );
z2_MC_STAR.ResizeTo( x2.Height(), 1 );
//----------------------------------------------------------------//
x1_STAR_STAR = x1;
L11_STAR_STAR = L11;
Trsv
( LOWER, NORMAL, diag,
L11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_MR_STAR = x1_STAR_STAR;
Gemv
( NORMAL, (F)-1,
L21.LockedLocalMatrix(),
x1_MR_STAR.LockedLocalMatrix(),
(F)0, z2_MC_STAR.LocalMatrix() );
x2.SumScatterUpdate( (F)1, z2_MC_STAR );
//----------------------------------------------------------------//
x1_MR_STAR.FreeAlignments();
z2_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionDown
( xT, x0,
x1,
/**/ /**/
xB, x2 );
}
}
else
{
// Matrix views
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);
DistMatrix<F,MC,MR>
xL(g), xR(g),
x0(g), x1(g), x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,STAR,MR > x1_STAR_MR(g);
DistMatrix<F,STAR,MC > z2_STAR_MC(g);
DistMatrix<F,MR, MC > z2_MR_MC(g);
DistMatrix<F,MC, MR > z2(g);
// Start the algorithm
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionRight( x, xL, xR, 0 );
while( xR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionRight
( xL, /**/ xR,
x0, /**/ x1, x2 );
x1_STAR_MR.AlignWith( L21 );
z2_STAR_MC.AlignWith( L21 );
z2.AlignWith( x2 );
z2_STAR_MC.ResizeTo( 1, x2.Width() );
//----------------------------------------------------------------//
x1_STAR_STAR = x1;
L11_STAR_STAR = L11;
Trsv
( LOWER, NORMAL, diag,
L11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_STAR_MR = x1_STAR_STAR;
Gemv
( NORMAL, (F)-1,
L21.LockedLocalMatrix(),
x1_STAR_MR.LockedLocalMatrix(),
(F)0, z2_STAR_MC.LocalMatrix() );
z2_MR_MC.SumScatterFrom( z2_STAR_MC );
z2 = z2_MR_MC;
Axpy( (F)1, z2, x2 );
//----------------------------------------------------------------//
x1_STAR_MR.FreeAlignments();
z2_STAR_MC.FreeAlignments();
z2.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionRight
( xL, /**/ xR,
x0, x1, /**/ x2 );
}
}
#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
}
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 );
}
}
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
TrsmLLTLarge
( Orientation orientation, UnitOrNonUnit diag,
F alpha, const DistMatrix<F>& L, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmLLTLarge");
if( orientation == NORMAL )
throw std::logic_error("TrsmLLT expects a (Conjugate)Transpose option");
#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> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<F,STAR,MC > L10_STAR_MC(g);
DistMatrix<F,STAR,STAR> L11_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
( 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 );
L10_STAR_MC.AlignWith( X0 );
X1_STAR_MR.AlignWith( X0 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11; // L11[* ,* ] <- L11[MC,MR]
X1_STAR_VR = X1; // X1[* ,VR] <- X1[MC,MR]
// X1[* ,VR] := L11^-[T/H][* ,* ] X1[* ,VR]
LocalTrsm
( LEFT, LOWER, orientation, diag, F(1), L11_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]
L10_STAR_MC = L10; // L10[* ,MC] <- L10[MC,MR]
// X0[MC,MR] -= (L10[* ,MC])^(T/H) X1[* ,MR]
// = L10^[T/H][MC,* ] X1[* ,MR]
LocalGemm
( orientation, NORMAL, F(-1), L10_STAR_MC, X1_STAR_MR, F(1), X0 );
//--------------------------------------------------------------------//
L10_STAR_MC.FreeAlignments();
X1_STAR_MR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmRLNCOld
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmRLNCOld");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( L.Height() != L.Width() || X.Width() != L.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmRLNC: \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> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,MR, STAR> L21_MR_STAR(g);
DistMatrix<T,VC, STAR> X1_VC_STAR(g);
DistMatrix<T,MC, STAR> D1_MC_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
L21_MR_STAR.AlignWith( X2 );
D1_MC_STAR.AlignWith( X1 );
Zeros( X1.Height(), X1.Width(), D1_MC_STAR );
//--------------------------------------------------------------------//
X1_VC_STAR = X1;
L11_STAR_STAR = L11;
LocalTrmm
( RIGHT, LOWER, NORMAL, diag, T(1), L11_STAR_STAR, X1_VC_STAR );
X1 = X1_VC_STAR;
L21_MR_STAR = L21;
LocalGemm( NORMAL, NORMAL, T(1), X2, L21_MR_STAR, T(0), D1_MC_STAR );
X1.SumScatterUpdate( T(1), D1_MC_STAR );
//--------------------------------------------------------------------//
L21_MR_STAR.FreeAlignments();
D1_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmLLTA
( Orientation orientation,
UnitOrNonUnit diag,
T alpha,
const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLLTA");
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
("TrmmLLTA expects a (Conjugate)Transpose option");
if( L.Height() != L.Width() || L.Height() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLTA: \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>
XL(g), XR(g),
X0(g), X1(g), X2(g);
DistMatrix<T,MC,STAR> X1_MC_STAR(g);
DistMatrix<T,MR,STAR> Z1_MR_STAR(g);
DistMatrix<T,MR,MC > Z1_MR_MC(g);
X1_MC_STAR.AlignWith( L );
Z1_MR_STAR.AlignWith( L );
PartitionRight( X, XL, XR, 0 );
while( XL.Width() < X.Width() )
{
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
Zeros( X1.Height(), X1.Width(), Z1_MR_STAR );
//--------------------------------------------------------------------//
X1_MC_STAR = X1;
LocalTrmmAccumulateLLT
( orientation, diag, alpha, L, X1_MC_STAR, Z1_MR_STAR );
Z1_MR_MC.SumScatterFrom( Z1_MR_STAR );
X1 = Z1_MR_MC;
//--------------------------------------------------------------------//
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TwoSidedTrsmLVar2
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& L )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrsmLVar2");
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,MR, STAR> A10Adj_MR_STAR(g);
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,MR, STAR> F10Adj_MR_STAR(g);
DistMatrix<F,MR, STAR> L10Adj_MR_STAR(g);
DistMatrix<F,VC, STAR> L10Adj_VC_STAR(g);
DistMatrix<F,STAR,MC > L10_STAR_MC(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,MC, STAR> X11_MC_STAR(g);
DistMatrix<F,MC, STAR> X21_MC_STAR(g);
DistMatrix<F,MC, STAR> Y10Adj_MC_STAR(g);
DistMatrix<F,MR, MC > Y10Adj_MR_MC(g);
DistMatrix<F> X11(g);
DistMatrix<F> Y10Adj(g);
Matrix<F> Y10Local;
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 );
A10Adj_MR_STAR.AlignWith( L10 );
F10Adj_MR_STAR.AlignWith( A00 );
L10Adj_MR_STAR.AlignWith( A00 );
L10Adj_VC_STAR.AlignWith( A00 );
L10_STAR_MC.AlignWith( A00 );
X11.AlignWith( A11 );
X11_MC_STAR.AlignWith( L10 );
X21_MC_STAR.AlignWith( A20 );
Y10Adj_MC_STAR.AlignWith( A00 );
Y10Adj_MR_MC.AlignWith( A10 );
//--------------------------------------------------------------------//
// Y10 := L10 A00
L10Adj_MR_STAR.AdjointFrom( L10 );
L10Adj_VC_STAR = L10Adj_MR_STAR;
L10_STAR_MC.AdjointFrom( L10Adj_VC_STAR );
Y10Adj_MC_STAR.ResizeTo( A10.Width(), A10.Height() );
F10Adj_MR_STAR.ResizeTo( A10.Width(), A10.Height() );
Zero( Y10Adj_MC_STAR );
Zero( F10Adj_MR_STAR );
LocalSymmetricAccumulateRL
( ADJOINT,
F(1), A00, L10_STAR_MC, L10Adj_MR_STAR,
Y10Adj_MC_STAR, F10Adj_MR_STAR );
Y10Adj.SumScatterFrom( Y10Adj_MC_STAR );
Y10Adj_MR_MC = Y10Adj;
Y10Adj_MR_MC.SumScatterUpdate( F(1), F10Adj_MR_STAR );
Adjoint( Y10Adj_MR_MC.LockedLocalMatrix(), Y10Local );
// X11 := A10 L10'
X11_MC_STAR.ResizeTo( A11.Height(), A11.Width() );
LocalGemm
( NORMAL, NORMAL, F(1), A10, L10Adj_MR_STAR, F(0), X11_MC_STAR );
// A10 := A10 - Y10
Axpy( F(-1), Y10Local, A10.LocalMatrix() );
A10Adj_MR_STAR.AdjointFrom( A10 );
// A11 := A11 - (X11 + L10 A10') = A11 - (A10 L10' + L10 A10')
LocalGemm
( NORMAL, NORMAL, F(1), L10, A10Adj_MR_STAR, F(1), X11_MC_STAR );
X11.SumScatterFrom( X11_MC_STAR );
MakeTrapezoidal( LEFT, LOWER, 0, X11 );
Axpy( F(-1), X11, A11 );
// A10 := inv(L11) A10
L11_STAR_STAR = L11;
A10_STAR_VR.AdjointFrom( A10Adj_MR_STAR );
LocalTrsm
( LEFT, LOWER, NORMAL, diag, F(1), L11_STAR_STAR, A10_STAR_VR );
A10 = A10_STAR_VR;
// A11 := inv(L11) A11 inv(L11)'
A11_STAR_STAR = A11;
LocalTwoSidedTrsm( LOWER, diag, A11_STAR_STAR, L11_STAR_STAR );
A11 = A11_STAR_STAR;
// A21 := A21 - A20 L10'
X21_MC_STAR.ResizeTo( A21.Height(), A21.Width() );
LocalGemm
( NORMAL, NORMAL, F(1), A20, L10Adj_MR_STAR, F(0), X21_MC_STAR );
A21.SumScatterUpdate( F(-1), X21_MC_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;
//--------------------------------------------------------------------//
A10Adj_MR_STAR.FreeAlignments();
F10Adj_MR_STAR.FreeAlignments();
L10Adj_MR_STAR.FreeAlignments();
L10Adj_VC_STAR.FreeAlignments();
L10_STAR_MC.FreeAlignments();
X11.FreeAlignments();
X11_MC_STAR.FreeAlignments();
X21_MC_STAR.FreeAlignments();
Y10Adj_MC_STAR.FreeAlignments();
Y10Adj_MR_MC.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
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
TrdtrmmLVar1( Orientation orientation, DistMatrix<F>& L )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrdtrmmLVar1");
if( L.Height() != L.Width() )
LogicError("L must be square");
if( orientation == NORMAL )
LogicError("Orientation must be (conjugate-)transpose");
#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,MD,STAR> d1(g);
// Temporary distributions
DistMatrix<F,STAR,VR > L10_STAR_VR(g);
DistMatrix<F,STAR,VC > S10_STAR_VC(g);
DistMatrix<F,STAR,MC > S10_STAR_MC(g);
DistMatrix<F,STAR,MR > L10_STAR_MR(g);
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
L10_STAR_VR.AlignWith( L );
S10_STAR_VC.AlignWith( L );
S10_STAR_MC.AlignWith( L );
L10_STAR_MR.AlignWith( L );
PartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
while( LTL.Height() < L.Height() && LTL.Width() < L.Height() )
{
RepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
//--------------------------------------------------------------------//
L11.GetDiagonal( d1 );
L10_STAR_VR = L10;
S10_STAR_VC = L10_STAR_VR;
S10_STAR_MC = S10_STAR_VC;
DiagonalSolve( LEFT, NORMAL, d1, L10_STAR_VR, true );
L10_STAR_MR = L10_STAR_VR;
LocalTrrk
( LOWER, orientation, F(1), S10_STAR_MC, L10_STAR_MR, F(1), L00 );
L11_STAR_STAR = L11;
LocalTrmm
( LEFT, LOWER, orientation, UNIT, F(1), L11_STAR_STAR, L10_STAR_VR );
L10 = L10_STAR_VR;
LocalTrdtrmm( orientation, LOWER, L11_STAR_STAR );
L11 = L11_STAR_STAR;
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
}
}