inline void
TrmmRUNA
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& U,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmRUNA");
if( U.Grid() != X.Grid() )
throw std::logic_error("{U,X} must be distributed over the same grid");
#endif
const Grid& g = U.Grid();
DistMatrix<T>
XT(g), X0(g),
XB(g), X1(g),
X2(g);
DistMatrix<T,STAR,VC > X1_STAR_VC(g);
DistMatrix<T,STAR,MC > X1_STAR_MC(g);
DistMatrix<T,MR, STAR> Z1Trans_MR_STAR(g);
DistMatrix<T,MR, MC > Z1Trans_MR_MC(g);
X1_STAR_VC.AlignWith( U );
X1_STAR_MC.AlignWith( U );
Z1Trans_MR_STAR.AlignWith( U );
PartitionDown
( X, XT,
XB, 0 );
while( XT.Height() < X.Height() )
{
RepartitionDown
( XT, X0,
/**/ /**/
X1,
XB, X2 );
Z1Trans_MR_MC.AlignWith( X1 );
//--------------------------------------------------------------------//
X1_STAR_VC = X1;
X1_STAR_MC = X1_STAR_VC;
Zeros( Z1Trans_MR_STAR, X1.Width(), X1.Height() );
LocalTrmmAccumulateRUN
( TRANSPOSE, diag, alpha, U, X1_STAR_MC, Z1Trans_MR_STAR );
Z1Trans_MR_MC.SumScatterFrom( Z1Trans_MR_STAR );
Transpose( Z1Trans_MR_MC.Matrix(), X1.Matrix() );
//--------------------------------------------------------------------//
Z1Trans_MR_MC.FreeAlignments();
SlidePartitionDown
( XT, X0,
X1,
/**/ /**/
XB, X2 );
}
}
inline void
Householder( Matrix<F>& A, Matrix<F>& t )
{
#ifndef RELEASE
CallStackEntry entry("lq::Householder");
#endif
t.ResizeTo( Min(A.Height(),A.Width()), 1 );
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02, ATopPan, ABottomPan,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
tT, t0,
tB, t1,
t2;
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2 );
View1x2( ATopPan, A11, A12 );
View1x2( ABottomPan, A21, A22 );
//--------------------------------------------------------------------//
PanelHouseholder( ATopPan, t1 );
ApplyQ( RIGHT, ADJOINT, ATopPan, t1, ABottomPan );
//--------------------------------------------------------------------//
SlidePartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
}
inline void
PanelLQ
( DistMatrix<Complex<Real> >& A,
DistMatrix<Complex<Real>,MD,STAR>& t )
{
#ifndef RELEASE
PushCallStack("internal::PanelLQ");
if( A.Grid() != t.Grid() )
throw std::logic_error("{A,t} must be distributed over the same grid");
if( t.Height() != std::min(A.Height(),A.Width()) || t.Width() != 1 )
throw std::logic_error
("t must be a vector of height equal to the minimum dimension of A");
if( !t.AlignedWithDiagonal( A, 0 ) )
throw std::logic_error("t must be aligned with A's main diagonal");
#endif
typedef Complex<Real> C;
const Grid& g = A.Grid();
// Matrix views
DistMatrix<C>
ATL(g), ATR(g), A00(g), a01(g), A02(g), aTopRow(g), ABottomPan(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g),
A20(g), a21(g), A22(g);
DistMatrix<C,MD,STAR>
tT(g), t0(g),
tB(g), tau1(g),
t2(g);
// Temporary distributions
DistMatrix<C> aTopRowConj(g);
DistMatrix<C,STAR,MR > aTopRowConj_STAR_MR(g);
DistMatrix<C,MC, STAR> z_MC_STAR(g);
PushBlocksizeStack( 1 );
PartitionDownLeftDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
RepartitionDown
( tT, t0,
/**/ /****/
tau1,
tB, t2 );
aTopRow.View1x2( alpha11, a12 );
ABottomPan.View1x2( a21, A22 );
aTopRowConj_STAR_MR.AlignWith( ABottomPan );
z_MC_STAR.AlignWith( ABottomPan );
Zeros( ABottomPan.Height(), 1, z_MC_STAR );
//--------------------------------------------------------------------//
const C tau = Reflector( alpha11, a12 );
tau1.Set( 0, 0, tau );
const bool myDiagonalEntry = ( g.Row() == alpha11.ColAlignment() &&
g.Col() == alpha11.RowAlignment() );
C alpha = 0;
if( myDiagonalEntry )
{
alpha = alpha11.GetLocal(0,0);
alpha11.SetLocal(0,0,1);
}
Conjugate( aTopRow, aTopRowConj );
aTopRowConj_STAR_MR = aTopRowConj;
Gemv
( NORMAL,
C(1), ABottomPan.LockedLocalMatrix(),
aTopRowConj_STAR_MR.LockedLocalMatrix(),
C(0), z_MC_STAR.LocalMatrix() );
z_MC_STAR.SumOverRow();
Ger
( -Conj(tau),
z_MC_STAR.LockedLocalMatrix(),
aTopRowConj_STAR_MR.LockedLocalMatrix(),
ABottomPan.LocalMatrix() );
if( myDiagonalEntry )
alpha11.SetLocal(0,0,alpha);
//--------------------------------------------------------------------//
aTopRowConj_STAR_MR.FreeAlignments();
z_MC_STAR.FreeAlignments();
SlidePartitionDown
( tT, t0,
tau1,
/**/ /****/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
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
}
inline void
internal::ApplyPackedReflectorsLLVF
( Conjugation conjugation, int offset,
const DistMatrix<Complex<R>,MC,MR >& H,
const DistMatrix<Complex<R>,MD,STAR>& t,
DistMatrix<Complex<R>,MC,MR >& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsLLVF");
if( H.Grid() != t.Grid() || t.Grid() != A.Grid() )
throw std::logic_error
("{H,t,A} must be distributed over the same grid");
if( offset > 0 )
throw std::logic_error("Transforms cannot extend above matrix");
if( offset < -H.Height() )
throw std::logic_error("Transforms cannot extend below matrix");
if( H.Height() != A.Height() )
throw std::logic_error
("Height of transforms must equal height of target matrix");
if( t.Height() != H.DiagonalLength( offset ) )
throw std::logic_error("t must be the same length as H's offset diag.");
if( !t.AlignedWithDiagonal( H, offset ) )
throw std::logic_error("t must be aligned with H's 'offset' diagonal");
#endif
typedef Complex<R> C;
const Grid& g = H.Grid();
// Matrix views
DistMatrix<C,MC,MR>
HTL(g), HTR(g), H00(g), H01(g), H02(g), HPan(g), HPanCopy(g),
HBL(g), HBR(g), H10(g), H11(g), H12(g),
H20(g), H21(g), H22(g);
DistMatrix<C,MC,MR>
AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<C,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
DistMatrix<C,VC, STAR> HPan_VC_STAR(g);
DistMatrix<C,MC, STAR> HPan_MC_STAR(g);
DistMatrix<C,STAR,STAR> t1_STAR_STAR(g);
DistMatrix<C,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<C,STAR,MR > Z_STAR_MR(g);
DistMatrix<C,STAR,VR > Z_STAR_VR(g);
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
int HPanHeight = H11.Height() + H21.Height();
int HPanWidth = std::min( H11.Width(), std::max(HPanHeight+offset,0) );
HPan.LockedView( H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanWidth );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
HPan_MC_STAR.AlignWith( AB );
Z_STAR_MR.AlignWith( AB );
Z_STAR_VR.AlignWith( AB );
Z_STAR_MR.ResizeTo( HPan.Width(), AB.Width() );
SInv_STAR_STAR.ResizeTo( HPan.Width(), HPan.Width() );
Zero( SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, LOWER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
HPan_VC_STAR = HPanCopy;
Herk
( UPPER, ADJOINT,
(C)1, HPan_VC_STAR.LockedLocalMatrix(),
(C)0, SInv_STAR_STAR.LocalMatrix() );
SInv_STAR_STAR.SumOverGrid();
t1_STAR_STAR = t1;
FixDiagonal( conjugation, t1_STAR_STAR, SInv_STAR_STAR );
HPan_MC_STAR = HPanCopy;
internal::LocalGemm
( ADJOINT, NORMAL, (C)1, HPan_MC_STAR, AB, (C)0, Z_STAR_MR );
Z_STAR_VR.SumScatterFrom( Z_STAR_MR );
internal::LocalTrsm
( LEFT, UPPER, ADJOINT, NON_UNIT, (C)1, SInv_STAR_STAR, Z_STAR_VR );
Z_STAR_MR = Z_STAR_VR;
internal::LocalGemm
( NORMAL, NORMAL, (C)-1, HPan_MC_STAR, Z_STAR_MR, (C)1, AB );
//--------------------------------------------------------------------//
HPan_MC_STAR.FreeAlignments();
Z_STAR_MR.FreeAlignments();
Z_STAR_VR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
ApplyPackedReflectorsLUVF
( int offset, const Matrix<R>& H, Matrix<R>& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsLUVF");
if( offset < 0 || offset > H.Height() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Height() )
throw std::logic_error
("Width of transforms must equal height of target matrix");
#endif
Matrix<R>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<R>
AT, A0, ATop,
AB, A1,
A2;
Matrix<R> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
const int HPanHeight = H01.Height() + H11.Height();
const int HPanOffset =
std::min( H11.Width(), std::max(offset-H00.Width(),0) );
const int HPanWidth = H11.Width()-HPanOffset;
HPan.LockedView( H, 0, H00.Width()+HPanOffset, HPanHeight, HPanWidth );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
ATop.View2x1( A0,
A1 );
Zeros( HPan.Width(), ATop.Width(), Z );
Zeros( HPan.Width(), HPan.Width(), SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, UPPER, offset, HPanCopy );
SetDiagonalToOne( RIGHT, offset, HPanCopy );
Syrk( LOWER, TRANSPOSE, R(1), HPanCopy, R(0), SInv );
HalveMainDiagonal( SInv );
Gemm( TRANSPOSE, NORMAL, R(1), HPanCopy, ATop, R(0), Z );
Trsm( LEFT, LOWER, NORMAL, NON_UNIT, R(1), SInv, Z );
Gemm( NORMAL, NORMAL, R(-1), HPanCopy, Z, R(1), ATop );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
ApplyPackedReflectorsLUVF
( int offset,
const DistMatrix<R>& H,
DistMatrix<R>& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsLUVF");
if( H.Grid() != A.Grid() )
throw std::logic_error("{H,A} must be distributed over the same grid");
if( offset < 0 || offset > H.Height() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Height() )
throw std::logic_error
("Width of transforms must equal height of target matrix");
#endif
const Grid& g = H.Grid();
DistMatrix<R>
HTL(g), HTR(g), H00(g), H01(g), H02(g), HPan(g),
HBL(g), HBR(g), H10(g), H11(g), H12(g),
H20(g), H21(g), H22(g);
DistMatrix<R>
AT(g), A0(g), ATop(g),
AB(g), A1(g),
A2(g);
DistMatrix<R> HPanCopy(g);
DistMatrix<R,VC, STAR> HPan_VC_STAR(g);
DistMatrix<R,MC, STAR> HPan_MC_STAR(g);
DistMatrix<R,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<R,STAR,MR > Z_STAR_MR(g);
DistMatrix<R,STAR,VR > Z_STAR_VR(g);
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
const int HPanHeight = H01.Height() + H11.Height();
const int HPanOffset =
std::min( H11.Width(), std::max(offset-H00.Width(),0) );
const int HPanWidth = H11.Width()-HPanOffset;
HPan.LockedView( H, 0, H00.Width()+HPanOffset, HPanHeight, HPanWidth );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
ATop.View2x1( A0,
A1 );
HPan_MC_STAR.AlignWith( ATop );
Z_STAR_MR.AlignWith( ATop );
Z_STAR_VR.AlignWith( ATop );
Zeros( HPan.Width(), ATop.Width(), Z_STAR_MR );
Zeros( HPan.Width(), HPan.Width(), SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, UPPER, offset, HPanCopy );
SetDiagonalToOne( RIGHT, offset, HPanCopy );
HPan_VC_STAR = HPanCopy;
Syrk
( LOWER, TRANSPOSE,
R(1), HPan_VC_STAR.LockedLocalMatrix(),
R(0), SInv_STAR_STAR.LocalMatrix() );
SInv_STAR_STAR.SumOverGrid();
HalveMainDiagonal( SInv_STAR_STAR );
HPan_MC_STAR = HPanCopy;
LocalGemm
( TRANSPOSE, NORMAL, R(1), HPan_MC_STAR, ATop, R(0), Z_STAR_MR );
Z_STAR_VR.SumScatterFrom( Z_STAR_MR );
LocalTrsm
( LEFT, LOWER, NORMAL, NON_UNIT,
R(1), SInv_STAR_STAR, Z_STAR_VR );
Z_STAR_MR = Z_STAR_VR;
LocalGemm( NORMAL, NORMAL, R(-1), HPan_MC_STAR, Z_STAR_MR, R(1), ATop );
//--------------------------------------------------------------------//
HPan_MC_STAR.FreeAlignments();
Z_STAR_MR.FreeAlignments();
Z_STAR_VR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelHouseholder( DistMatrix<F>& A, DistMatrix<F,MD,STAR>& t )
{
#ifndef RELEASE
CallStackEntry entry("lq::PanelHouseholder");
if( A.Grid() != t.Grid() )
LogicError("{A,t} must be distributed over the same grid");
if( t.Height() != Min(A.Height(),A.Width()) || t.Width() != 1 )
LogicError
("t must be a vector of height equal to the minimum dimension of A");
if( !t.AlignedWithDiagonal( A, 0 ) )
LogicError("t must be aligned with A's main diagonal");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), a01(g), A02(g), aTopRow(g), ABottomPan(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g),
A20(g), a21(g), A22(g);
DistMatrix<F,MD,STAR>
tT(g), t0(g),
tB(g), tau1(g),
t2(g);
// Temporary distributions
DistMatrix<F> aTopRowConj(g);
DistMatrix<F,STAR,MR > aTopRowConj_STAR_MR(g);
DistMatrix<F,MC, STAR> z_MC_STAR(g);
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22, 1 );
RepartitionDown
( tT, t0,
/**/ /****/
tau1,
tB, t2, 1 );
View1x2( aTopRow, alpha11, a12 );
View1x2( ABottomPan, a21, A22 );
aTopRowConj_STAR_MR.AlignWith( ABottomPan );
z_MC_STAR.AlignWith( ABottomPan );
//--------------------------------------------------------------------//
// Compute the Householder reflector
const F tau = Reflector( alpha11, a12 );
tau1.Set( 0, 0, tau );
// Apply the Householder reflector
const bool myDiagonalEntry = ( g.Row() == alpha11.ColAlignment() &&
g.Col() == alpha11.RowAlignment() );
F alpha = 0;
if( myDiagonalEntry )
{
alpha = alpha11.GetLocal(0,0);
alpha11.SetLocal(0,0,1);
}
Conjugate( aTopRow, aTopRowConj );
aTopRowConj_STAR_MR = aTopRowConj;
Zeros( z_MC_STAR, ABottomPan.Height(), 1 );
LocalGemv
( NORMAL, F(1), ABottomPan, aTopRowConj_STAR_MR, F(0), z_MC_STAR );
z_MC_STAR.SumOverRow();
Ger
( -Conj(tau),
z_MC_STAR.LockedMatrix(),
aTopRowConj_STAR_MR.LockedMatrix(),
ABottomPan.Matrix() );
if( myDiagonalEntry )
alpha11.SetLocal(0,0,alpha);
//--------------------------------------------------------------------//
SlidePartitionDown
( tT, t0,
tau1,
/**/ /****/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
}
inline void
Householder( DistMatrix<F>& A, DistMatrix<F,MD,STAR>& t )
{
#ifndef RELEASE
CallStackEntry entry("qr::Householder");
if( A.Grid() != t.Grid() )
LogicError("{A,s} must be distributed over the same grid");
#endif
const Grid& g = A.Grid();
if( t.Viewing() )
{
if( !t.AlignedWithDiagonal( A ) )
LogicError("t was not aligned with A");
}
else
{
t.AlignWithDiagonal( A );
}
t.ResizeTo( Min(A.Height(),A.Width()), 1 );
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g), ALeftPan(g), ARightPan(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<F,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2 );
View2x1
( ALeftPan, A11,
A21 );
View2x1
( ARightPan, A12,
A22 );
//--------------------------------------------------------------------//
PanelHouseholder( ALeftPan, t1 );
ApplyQ( LEFT, ADJOINT, ALeftPan, t1, ARightPan );
//--------------------------------------------------------------------//
SlidePartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
}
inline void
internal::PanelLU
( DistMatrix<F, STAR,STAR>& A,
DistMatrix<F, MC, STAR>& B,
DistMatrix<int,STAR,STAR>& p,
int pivotOffset )
{
#ifndef RELEASE
PushCallStack("internal::PanelLU");
if( A.Grid() != p.Grid() || p.Grid() != B.Grid() )
throw std::logic_error
("Matrices must be distributed over the same grid");
if( A.Width() != B.Width() )
throw std::logic_error("A and B must be the same width");
if( A.Height() != p.Height() || p.Width() != 1 )
throw std::logic_error("p must be a vector that conforms with A");
#endif
const Grid& g = A.Grid();
const int r = g.Height();
const int colShift = B.ColShift();
const int colAlignment = B.ColAlignment();
// Matrix views
DistMatrix<F,STAR,STAR>
ATL(g), ATR(g), A00(g), a01(g), A02(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g),
A20(g), a21(g), A22(g);
DistMatrix<F,MC,STAR>
BL(g), BR(g),
B0(g), b1(g), B2(g);
DistMatrix<int,STAR,STAR>
pT(g), p0(g),
pB(g), psi1(g),
p2(g);
const int width = A.Width();
const int numBytes = (width+1)*sizeof(F)+sizeof(int);
std::vector<byte> sendData(numBytes);
std::vector<byte> recvData(numBytes);
// Extract pointers to send and recv data
F* sendBufFloat = (F*) &sendData[0];
F* recvBufFloat = (F*) &recvData[0];
int* sendBufInt = (int*) &sendData[(width+1)*sizeof(F)];
int* recvBufInt = (int*) &recvData[(width+1)*sizeof(F)];
// Start the algorithm
PushBlocksizeStack( 1 );
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionRight( B, BL, BR, 0 );
PartitionDown
( p, pT,
pB, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
RepartitionRight
( BL, /**/ BR,
B0, /**/ b1, B2 );
RepartitionDown
( pT, p0,
/**/ /****/
psi1,
pB, p2 );
//--------------------------------------------------------------------//
// Store the index/value of the pivot candidate in A
F pivotValue = alpha11.GetLocalEntry(0,0);
int pivotIndex = a01.Height();
for( int i=0; i<a21.Height(); ++i )
{
F value = a21.GetLocalEntry(i,0);
if( FastAbs(value) > FastAbs(pivotValue) )
{
pivotValue = value;
pivotIndex = a01.Height() + i + 1;
}
}
// Update the pivot candidate to include local data from B
for( int i=0; i<B.LocalHeight(); ++i )
{
F value = b1.GetLocalEntry(i,0);
if( FastAbs(value) > FastAbs(pivotValue) )
{
pivotValue = value;
pivotIndex = A.Height() + colShift + i*r;
}
}
// Fill the send buffer with:
// [ pivotValue | pivotRow | pivotIndex ]
if( pivotIndex < A.Height() )
{
sendBufFloat[0] = A.GetLocalEntry(pivotIndex,a10.Width());
const int ALDim = A.LocalLDim();
const F* ABuffer = A.LocalBuffer(pivotIndex,0);
for( int j=0; j<width; ++j )
sendBufFloat[j+1] = ABuffer[j*ALDim];
}
else
{
const int localIndex = ((pivotIndex-A.Height())-colShift)/r;
sendBufFloat[0] = b1.GetLocalEntry(localIndex,0);
const int BLDim = B.LocalLDim();
const F* BBuffer = B.LocalBuffer(localIndex,0);
for( int j=0; j<width; ++j )
sendBufFloat[j+1] = BBuffer[j*BLDim];
}
*sendBufInt = pivotIndex;
// Communicate to establish the pivot information
mpi::AllReduce
( &sendData[0], &recvData[0], numBytes, PivotOp<F>(), g.ColComm() );
// Update the pivot vector
const int maxIndex = *recvBufInt;
p.SetLocalEntry(a01.Height(),0,maxIndex+pivotOffset);
// Copy the current row into the pivot row
if( maxIndex < A.Height() )
{
const int ALDim = A.LocalLDim();
F* ASetBuffer = A.LocalBuffer(maxIndex,0);
const F* AGetBuffer = A.LocalBuffer(A00.Height(),0);
for( int j=0; j<width; ++j )
ASetBuffer[j*ALDim] = AGetBuffer[j*ALDim];
}
else
{
const int ownerRank = (colAlignment+(maxIndex-A.Height())) % r;
if( g.Row() == ownerRank )
{
const int localIndex = ((maxIndex-A.Height())-colShift) / r;
const int ALDim = A.LocalLDim();
const int BLDim = B.LocalLDim();
F* BBuffer = B.LocalBuffer(localIndex,0);
const F* ABuffer = A.LocalBuffer(A00.Height(),0);
for( int j=0; j<width; ++j )
BBuffer[j*BLDim] = ABuffer[j*ALDim];
}
}
// Copy the pivot row into the current row
{
F* ABuffer = A.LocalBuffer(A00.Height(),0);
const int ALDim = A.LocalLDim();
for( int j=0; j<width; ++j )
ABuffer[j*ALDim] = recvBufFloat[j+1];
}
// Now we can perform the update of the current panel
F alpha = alpha11.GetLocalEntry(0,0);
if( alpha == (F)0 )
throw SingularMatrixException();
F alpha11Inv = ((F)1) / alpha;
Scal( alpha11Inv, a21.LocalMatrix() );
Scal( alpha11Inv, b1.LocalMatrix() );
Geru( (F)-1, a21.LocalMatrix(), a12.LocalMatrix(), A22.LocalMatrix() );
Geru( (F)-1, b1.LocalMatrix(), a12.LocalMatrix(), B2.LocalMatrix() );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
SlidePartitionRight
( BL, /**/ BR,
B0, b1, /**/ B2 );
SlidePartitionDown
( pT, p0,
psi1,
/**/ /****/
pB, p2 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LUHF
( Conjugation conjugation, Int offset,
const Matrix<F>& H, const Matrix<F>& t, Matrix<F>& A )
{
#ifndef RELEASE
CallStackEntry cse("apply_packed_reflectors::LUHF");
// TODO: Proper dimension checks
if( t.Height() != H.DiagonalLength(offset) )
LogicError("t must be the same length as H's offset diag");
#endif
Matrix<F>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<F>
AT, A0,
AB, A1,
A2;
Matrix<F>
tT, t0,
tB, t1,
t2;
Matrix<F> SInv, Z;
LockedPartitionDownOffsetDiagonal
( offset,
H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2 );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2, H11.Height() );
LockedView1x2( HPan, H11, H12 );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTriangular( UPPER, HPanCopy );
SetDiagonal( HPanCopy, F(1) );
Herk( LOWER, NORMAL, F(1), HPanCopy, SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, NORMAL, F(1), HPanCopy, AB, Z );
Trsm( LEFT, LOWER, NORMAL, NON_UNIT, F(1), SInv, Z );
Gemm( ADJOINT, NORMAL, F(-1), HPanCopy, Z, F(1), AB );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
}
inline void
LUHF
( Conjugation conjugation, Int offset,
const DistMatrix<F>& H, const DistMatrix<F,MD,STAR>& t, DistMatrix<F>& A )
{
#ifndef RELEASE
CallStackEntry cse("apply_packed_reflectors::LUHF");
if( H.Grid() != t.Grid() || t.Grid() != A.Grid() )
LogicError("{H,t,A} must be distributed over the same grid");
// TODO: Proper dimension checks
if( t.Height() != H.DiagonalLength(offset) )
LogicError("t must be the same length as H's offset diag");
if( !t.AlignedWithDiagonal( H, offset ) )
LogicError("t must be aligned with H's offset diagonal");
#endif
const Grid& g = H.Grid();
DistMatrix<F>
HTL(g), HTR(g), H00(g), H01(g), H02(g), HPan(g), HPanCopy(g),
HBL(g), HBR(g), H10(g), H11(g), H12(g),
H20(g), H21(g), H22(g);
DistMatrix<F>
AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<F,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
DistMatrix<F,STAR,VR > HPan_STAR_VR(g);
DistMatrix<F,STAR,MC > HPan_STAR_MC(g);
DistMatrix<F,STAR,STAR> t1_STAR_STAR(g);
DistMatrix<F,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<F,STAR,MR > Z_STAR_MR(g);
DistMatrix<F,STAR,VR > Z_STAR_VR(g);
LockedPartitionDownOffsetDiagonal
( offset,
H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2 );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2, H11.Height() );
LockedView1x2( HPan, H11, H12 );
HPan_STAR_MC.AlignWith( AB );
Z_STAR_MR.AlignWith( AB );
Z_STAR_VR.AlignWith( AB );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTriangular( UPPER, HPanCopy );
SetDiagonal( HPanCopy, F(1) );
HPan_STAR_VR = HPanCopy;
Zeros( SInv_STAR_STAR, HPan.Height(), HPan.Height() );
Herk
( LOWER, NORMAL,
F(1), HPan_STAR_VR.LockedMatrix(),
F(0), SInv_STAR_STAR.Matrix() );
SInv_STAR_STAR.SumOverGrid();
t1_STAR_STAR = t1;
FixDiagonal( conjugation, t1_STAR_STAR, SInv_STAR_STAR );
HPan_STAR_MC = HPan_STAR_VR;
LocalGemm( NORMAL, NORMAL, F(1), HPan_STAR_MC, AB, Z_STAR_MR );
Z_STAR_VR.SumScatterFrom( Z_STAR_MR );
LocalTrsm
( LEFT, LOWER, NORMAL, NON_UNIT, F(1), SInv_STAR_STAR, Z_STAR_VR );
Z_STAR_MR = Z_STAR_VR;
LocalGemm( ADJOINT, NORMAL, F(-1), HPan_STAR_MC, Z_STAR_MR, F(1), AB );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
}
inline void
LocalTrmmAccumulateRUN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
const DistMatrix<T,MC, MR >& U,
const DistMatrix<T,STAR,MC >& X_STAR_MC,
DistMatrix<T,MR, STAR>& ZTrans_MR_STAR )
{
#ifndef RELEASE
CallStackEntry entry("internal::LocalTrmmAccumulateRUN");
if( U.Grid() != X_STAR_MC.Grid() ||
X_STAR_MC.Grid() != ZTrans_MR_STAR.Grid() )
throw std::logic_error
("{U,X,Z} must be distributed over the same grid");
if( U.Height() != U.Width() ||
U.Height() != X_STAR_MC.Width() ||
U.Height() != ZTrans_MR_STAR.Height() )
{
std::ostringstream msg;
msg << "Nonconformal LocalTrmmAccumulateRUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X[* ,MC] ~ " << X_STAR_MC.Height() << " x "
<< X_STAR_MC.Width() << "\n"
<< " Z^H/T[MR,* ] ~ " << ZTrans_MR_STAR.Height() << " x "
<< ZTrans_MR_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( X_STAR_MC.RowAlignment() != U.ColAlignment() ||
ZTrans_MR_STAR.ColAlignment() != U.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<T> D11(g);
DistMatrix<T,STAR,MC>
XL_STAR_MC(g), XR_STAR_MC(g),
X0_STAR_MC(g), X1_STAR_MC(g), X2_STAR_MC(g);
DistMatrix<T,MR,STAR>
ZTTrans_MR_STAR(g), Z0Trans_MR_STAR(g),
ZBTrans_MR_STAR(g), Z1Trans_MR_STAR(g),
Z2Trans_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
LockedPartitionRight( X_STAR_MC, XL_STAR_MC, XR_STAR_MC, 0 );
PartitionDown
( ZTrans_MR_STAR, ZTTrans_MR_STAR,
ZBTrans_MR_STAR, 0 );
while( UTL.Height() < U.Height() )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
LockedRepartitionRight
( XL_STAR_MC, /**/ XR_STAR_MC,
X0_STAR_MC, /**/ X1_STAR_MC, X2_STAR_MC );
RepartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
/***************/ /***************/
Z1Trans_MR_STAR,
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
D11.AlignWith( U11 );
//--------------------------------------------------------------------//
D11 = U11;
MakeTriangular( UPPER, D11 );
if( diag == UNIT )
SetDiagonal( D11, T(1) );
LocalGemm
( orientation, orientation,
alpha, D11, X1_STAR_MC, T(1), Z1Trans_MR_STAR );
LocalGemm
( orientation, orientation,
alpha, U01, X0_STAR_MC, T(1), Z1Trans_MR_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlideLockedPartitionRight
( XL_STAR_MC, /**/ XR_STAR_MC,
X0_STAR_MC, X1_STAR_MC, /**/ X2_STAR_MC );
SlidePartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
Z1Trans_MR_STAR,
/***************/ /***************/
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
}
PopBlocksizeStack();
}
inline void
HemmRUA
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::HemmRUA");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
#endif
const Grid& g = A.Grid();
DistMatrix<T>
BT(g), B0(g),
BB(g), B1(g),
B2(g);
DistMatrix<T>
CT(g), C0(g),
CB(g), C1(g),
C2(g);
DistMatrix<T,MR, STAR> B1Adj_MR_STAR(g);
DistMatrix<T,VC, STAR> B1Adj_VC_STAR(g);
DistMatrix<T,STAR,MC > B1_STAR_MC(g);
DistMatrix<T,MC, STAR> Z1Adj_MC_STAR(g);
DistMatrix<T,MR, STAR> Z1Adj_MR_STAR(g);
DistMatrix<T,MR, MC > Z1Adj_MR_MC(g);
DistMatrix<T> Z1Adj(g);
B1Adj_MR_STAR.AlignWith( A );
B1Adj_VC_STAR.AlignWith( A );
B1_STAR_MC.AlignWith( A );
Z1Adj_MC_STAR.AlignWith( A );
Z1Adj_MR_STAR.AlignWith( A );
Matrix<T> Z1Local;
Scale( beta, C );
LockedPartitionDown
( B, BT,
BB, 0 );
PartitionDown
( C, CT,
CB, 0 );
while( CT.Height() < C.Height() )
{
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
RepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
Z1Adj_MR_MC.AlignWith( C1 );
Zeros( C1.Width(), C1.Height(), Z1Adj_MC_STAR );
Zeros( C1.Width(), C1.Height(), Z1Adj_MR_STAR );
//--------------------------------------------------------------------//
B1Adj_MR_STAR.AdjointFrom( B1 );
B1Adj_VC_STAR = B1Adj_MR_STAR;
B1_STAR_MC.AdjointFrom( B1Adj_VC_STAR );
LocalSymmetricAccumulateRU
( ADJOINT, alpha, A, B1_STAR_MC, B1Adj_MR_STAR,
Z1Adj_MC_STAR, Z1Adj_MR_STAR );
Z1Adj.SumScatterFrom( Z1Adj_MC_STAR );
Z1Adj_MR_MC = Z1Adj;
Z1Adj_MR_MC.SumScatterUpdate( T(1), Z1Adj_MR_STAR );
Adjoint( Z1Adj_MR_MC.LockedLocalMatrix(), Z1Local );
Axpy( T(1), Z1Local, C1.LocalMatrix() );
//--------------------------------------------------------------------//
Z1Adj_MR_MC.FreeAlignments();
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
#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
GemmTTB
( Orientation orientationOfA,
Orientation orientationOfB,
T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmTTB");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
if( orientationOfA == NORMAL || orientationOfB == NORMAL )
throw std::logic_error
("GemmTTB expects A and B to be (Conjugate)Transposed");
if( A.Width() != C.Height() ||
B.Height() != C.Width() ||
A.Height() != B.Width() )
{
std::ostringstream msg;
msg << "Nonconformal GemmTTB: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B ~ " << B.Height() << " x " << B.Width() << "\n"
<< " C ~ " << C.Height() << " x " << C.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<T> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T> CT(g), C0(g),
CB(g), C1(g),
C2(g);
// Temporary distributions
DistMatrix<T,VR, STAR> A1_VR_STAR(g);
DistMatrix<T,STAR,MR > A1AdjOrTrans_STAR_MR(g);
DistMatrix<T,STAR,MC > D1_STAR_MC(g);
DistMatrix<T,MR, MC > D1_MR_MC(g);
DistMatrix<T> D1(g);
A1_VR_STAR.AlignWith( B );
A1AdjOrTrans_STAR_MR.AlignWith( B );
D1_STAR_MC.AlignWith( B );
// Start the algorithm
Scale( beta, C );
LockedPartitionRight( A, AL, AR, 0 );
PartitionDown
( C, CT,
CB, 0 );
while( AR.Width() > 0 )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
RepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
D1.AlignWith( C1 );
Zeros( C1.Height(), C1.Width(), D1_STAR_MC );
//--------------------------------------------------------------------//
A1_VR_STAR = A1;
if( orientationOfA == ADJOINT )
A1AdjOrTrans_STAR_MR.AdjointFrom( A1_VR_STAR );
else
A1AdjOrTrans_STAR_MR.TransposeFrom( A1_VR_STAR );
// D1[*,MC] := alpha (A1[MR,*])^[T/H] (B[MC,MR])^[T/H]
// = alpha (A1^[T/H])[*,MR] (B^[T/H])[MR,MC]
LocalGemm
( NORMAL, orientationOfB,
alpha, A1AdjOrTrans_STAR_MR, B, T(0), D1_STAR_MC );
// C1[MC,MR] += scattered & transposed D1[*,MC] summed over grid rows
D1_MR_MC.SumScatterFrom( D1_STAR_MC );
D1 = D1_MR_MC;
Axpy( T(1), D1, C1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LQ( DistMatrix<Complex<R>,MC,MR >& A,
DistMatrix<Complex<R>,MD,STAR>& t )
{
#ifndef RELEASE
PushCallStack("LQ");
if( A.Grid() != t.Grid() )
throw std::logic_error("{A,t} must be distributed over the same grid");
#endif
typedef Complex<R> C;
const Grid& g = A.Grid();
if( t.Viewing() )
{
if( !t.AlignedWithDiagonal( A ) )
throw std::logic_error("t was not aligned with A");
if( t.Height() != std::min(A.Height(),A.Width()) || t.Width() != 1 )
throw std::logic_error("t was not the appropriate shape");
}
else
{
t.AlignWithDiagonal( A );
t.ResizeTo( std::min(A.Height(),A.Width()), 1 );
}
// Matrix views
DistMatrix<C,MC,MR>
ATL(g), ATR(g), A00(g), A01(g), A02(g), ATopPan(g), ABottomPan(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<C,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
PartitionDownLeftDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2 );
ATopPan.View1x2( A11, A12 );
ABottomPan.View1x2( A21, A22 );
//--------------------------------------------------------------------//
internal::PanelLQ( ATopPan, t1 );
ApplyPackedReflectors
( RIGHT, UPPER, HORIZONTAL, FORWARD, CONJUGATED,
0, ATopPan, t1, ABottomPan );
//--------------------------------------------------------------------//
SlidePartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LU( Matrix<F>& A, Matrix<Int>& p )
{
#ifndef RELEASE
CallStackEntry entry("LU");
#endif
p.ResizeTo( Min(A.Height(),A.Width()), 1 );
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02, ABRL, ABRR,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<Int>
pT, p0,
pB, p1,
p2;
// Pivot composition
std::vector<Int> image, preimage;
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( p, pT,
pB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( pT, p0,
/**/ /**/
p1,
pB, p2 );
PartitionRight( ABR, ABRL, ABRR, A11.Width() );
const Int pivotOffset = A01.Height();
//--------------------------------------------------------------------//
lu::Panel( ABRL, p1, pivotOffset );
ComposePivots( p1, pivotOffset, image, preimage );
ApplyRowPivots( ABL, image, preimage );
ApplyRowPivots( ABRR, image, preimage );
Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 );
Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlidePartitionDown
( pT, p0,
p1,
/**/ /**/
pB, p2 );
}
}
inline void
LocalSymmetricAccumulateLU
( Orientation orientation, T alpha,
const DistMatrix<T>& A,
const DistMatrix<T,MC, STAR>& B_MC_STAR,
const DistMatrix<T,STAR,MR >& BAdjOrTrans_STAR_MR,
DistMatrix<T,MC, STAR>& Z_MC_STAR,
DistMatrix<T,MR, STAR>& Z_MR_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalSymmetricAccumulateLU");
if( A.Grid() != B_MC_STAR.Grid() ||
B_MC_STAR.Grid() != BAdjOrTrans_STAR_MR.Grid() ||
BAdjOrTrans_STAR_MR.Grid() != Z_MC_STAR.Grid() ||
Z_MC_STAR.Grid() != Z_MR_STAR.Grid() )
throw std::logic_error
("{A,B,Z} must be distributed over the same grid");
if( A.Height() != A.Width() ||
A.Height() != B_MC_STAR.Height() ||
A.Height() != BAdjOrTrans_STAR_MR.Width() ||
A.Height() != Z_MC_STAR.Height() ||
A.Height() != Z_MR_STAR.Height() ||
B_MC_STAR.Width() != BAdjOrTrans_STAR_MR.Height() ||
BAdjOrTrans_STAR_MR.Height() != Z_MC_STAR.Width() ||
Z_MC_STAR.Width() != Z_MR_STAR.Width() )
{
std::ostringstream msg;
msg << "Nonconformal LocalSymmetricAccumulateLU: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B[MC,* ] ~ " << B_MC_STAR.Height() << " x "
<< B_MC_STAR.Width() << "\n"
<< " B^H/T[* ,MR] ~ " << BAdjOrTrans_STAR_MR.Height() << " x "
<< BAdjOrTrans_STAR_MR.Width() << "\n"
<< " Z[MC,* ] ~ " << Z_MC_STAR.Height() << " x "
<< Z_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( B_MC_STAR.ColAlignment() != A.ColAlignment() ||
BAdjOrTrans_STAR_MR.RowAlignment() != A.RowAlignment() ||
Z_MC_STAR.ColAlignment() != A.ColAlignment() ||
Z_MR_STAR.ColAlignment() != A.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = A.Grid();
DistMatrix<T>
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<T> D11(g);
DistMatrix<T,MC,STAR>
BT_MC_STAR(g), B0_MC_STAR(g),
BB_MC_STAR(g), B1_MC_STAR(g),
B2_MC_STAR(g);
DistMatrix<T,STAR,MR>
BLAdjOrTrans_STAR_MR(g), BRAdjOrTrans_STAR_MR(g),
B0AdjOrTrans_STAR_MR(g), B1AdjOrTrans_STAR_MR(g),
B2AdjOrTrans_STAR_MR(g);
DistMatrix<T,MC,STAR>
ZT_MC_STAR(g), Z0_MC_STAR(g),
ZB_MC_STAR(g), Z1_MC_STAR(g),
Z2_MC_STAR(g);
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
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDown
( B_MC_STAR, BT_MC_STAR,
BB_MC_STAR, 0 );
LockedPartitionRight
( BAdjOrTrans_STAR_MR, BLAdjOrTrans_STAR_MR, BRAdjOrTrans_STAR_MR, 0 );
PartitionDown
( Z_MC_STAR, ZT_MC_STAR,
ZB_MC_STAR, 0 );
PartitionDown
( Z_MR_STAR, ZT_MR_STAR,
ZB_MR_STAR, 0 );
while( ATL.Height() < A.Height() )
{
LockedRepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDown
( BT_MC_STAR, B0_MC_STAR,
/**********/ /**********/
B1_MC_STAR,
BB_MC_STAR, B2_MC_STAR );
LockedRepartitionRight
( BLAdjOrTrans_STAR_MR, /**/ BRAdjOrTrans_STAR_MR,
B0AdjOrTrans_STAR_MR, /**/ B1AdjOrTrans_STAR_MR,
B2AdjOrTrans_STAR_MR );
RepartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
/**********/ /**********/
Z1_MC_STAR,
ZB_MC_STAR, Z2_MC_STAR );
RepartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
/**********/ /**********/
Z1_MR_STAR,
ZB_MR_STAR, Z2_MR_STAR );
D11.AlignWith( A11 );
//--------------------------------------------------------------------//
D11 = A11;
MakeTrapezoidal( LEFT, UPPER, 0, D11 );
LocalGemm
( NORMAL, orientation,
alpha, D11, B1AdjOrTrans_STAR_MR, T(1), Z1_MC_STAR );
MakeTrapezoidal( LEFT, UPPER, 1, D11 );
LocalGemm
( orientation, NORMAL, alpha, D11, B1_MC_STAR, T(1), Z1_MR_STAR );
LocalGemm
( NORMAL, orientation,
alpha, A12, B2AdjOrTrans_STAR_MR, T(1), Z1_MC_STAR );
LocalGemm
( orientation, NORMAL, alpha, A12, B1_MC_STAR, T(1), Z2_MR_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDown
( BT_MC_STAR, B0_MC_STAR,
B1_MC_STAR,
/**********/ /**********/
BB_MC_STAR, B2_MC_STAR );
SlideLockedPartitionRight
( BLAdjOrTrans_STAR_MR, /**/ BRAdjOrTrans_STAR_MR,
B0AdjOrTrans_STAR_MR, B1AdjOrTrans_STAR_MR, /**/ B2AdjOrTrans_STAR_MR );
SlidePartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
Z1_MC_STAR,
/**********/ /**********/
ZB_MC_STAR, Z2_MC_STAR );
SlidePartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
Z1_MR_STAR,
/**********/ /**********/
ZB_MR_STAR, Z2_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LLVF( int offset, const Matrix<R>& H, Matrix<R>& A )
{
#ifndef RELEASE
PushCallStack("apply_packed_reflectors::LLVF");
if( offset > 0 || offset < -H.Height() )
throw std::logic_error("Transforms out of bounds");
if( H.Height() != A.Height() )
throw std::logic_error
("Height of transforms must equal height of target matrix");
#endif
Matrix<R>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<R>
AT, A0,
AB, A1,
A2;
Matrix<R> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
int HPanHeight = H11.Height() + H21.Height();
int HPanWidth = std::min( H11.Width(), std::max(HPanHeight+offset,0) );
LockedView( HPan, H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
Zeros( HPanWidth, AB.Width(), Z );
Zeros( HPanWidth, HPanWidth, SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, LOWER, offset, HPanCopy );
SetDiagonal( LEFT, offset, HPanCopy, R(1) );
Syrk( LOWER, TRANSPOSE, R(1), HPanCopy, R(0), SInv );
HalveMainDiagonal( SInv );
Gemm( TRANSPOSE, NORMAL, R(1), HPanCopy, AB, R(0), Z );
Trsm( LEFT, LOWER, NORMAL, NON_UNIT, R(1), SInv, Z );
Gemm( NORMAL, NORMAL, R(-1), HPanCopy, Z, R(1), AB );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelHouseholder( Matrix<F>& A, Matrix<F>& t )
{
#ifndef RELEASE
CallStackEntry entry("lq::PanelHouseholder");
if( t.Height() != Min(A.Height(),A.Width()) || t.Width() != 1 )
LogicError
("t must be a vector of height equal to the minimum dimension of A");
#endif
Matrix<F>
ATL, ATR, A00, a01, A02, aTopRow, ABottomPan,
ABL, ABR, a10, alpha11, a12,
A20, a21, A22;
Matrix<F>
tT, t0,
tB, tau1,
t2;
Matrix<F> z, aTopRowConj;
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22, 1 );
RepartitionDown
( tT, t0,
/**/ /****/
tau1,
tB, t2, 1 );
View1x2( aTopRow, alpha11, a12 );
View1x2( ABottomPan, a21, A22 );
//--------------------------------------------------------------------//
// Compute the Householder reflector
const F tau = Reflector( alpha11, a12 );
tau1.Set( 0, 0, tau );
// Apply the Householder reflector
const F alpha = alpha11.Get(0,0);
alpha11.Set(0,0,1);
Conjugate( aTopRow, aTopRowConj );
Zeros( z, ABottomPan.Height(), 1 );
Gemv( NORMAL, F(1), ABottomPan, aTopRowConj, F(0), z );
Ger( -Conj(tau), z, aTopRowConj, ABottomPan );
alpha11.Set(0,0,alpha);
//--------------------------------------------------------------------//
SlidePartitionDown
( tT, t0,
tau1,
/**/ /****/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
}
inline void
LU( Matrix<F>& A, Matrix<int>& p )
{
#ifndef RELEASE
CallStackEntry entry("LU");
if( p.Viewing() &&
(p.Height() != std::min(A.Height(),A.Width()) || p.Width() != 1) )
throw std::logic_error
("p must be a vector of the same height as the min dimension of A");
#endif
if( !p.Viewing() )
p.ResizeTo( std::min(A.Height(),A.Width()), 1 );
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02, ABRL, ABRR,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<int>
pT, p0,
pB, p1,
p2;
// Pivot composition
std::vector<int> image, preimage;
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( p, pT,
pB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( pT, p0,
/**/ /**/
p1,
pB, p2 );
PartitionRight( ABR, ABRL, ABRR, A11.Width() );
const int pivotOffset = A01.Height();
//--------------------------------------------------------------------//
lu::Panel( ABRL, p1, pivotOffset );
ComposePivots( p1, pivotOffset, image, preimage );
ApplyRowPivots( ABL, image, preimage );
ApplyRowPivots( ABRR, image, preimage );
Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 );
Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlidePartitionDown
( pT, p0,
p1,
/**/ /**/
pB, p2 );
}
}
inline void
ApplyPackedReflectorsLUVF
( Conjugation conjugation, int offset,
const Matrix<Complex<R> >& H,
const Matrix<Complex<R> >& t,
Matrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsLUVF");
if( offset < 0 || offset > H.Height() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Height() )
throw std::logic_error
("Width of transforms must equal height of target matrix");
if( t.Height() != H.DiagonalLength( offset ) )
throw std::logic_error("t must be the same length as H's offset diag");
#endif
typedef Complex<R> C;
Matrix<C>
HTL, HTR, H00, H01, H02, HPan,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<C>
AT, A0, ATop,
AB, A1,
A2;
Matrix<C>
tT, t0,
tB, t1,
t2;
Matrix<C> HPanCopy;
Matrix<C> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
const int HPanHeight = H01.Height() + H11.Height();
const int HPanOffset =
std::min( H11.Width(), std::max(offset-H00.Width(),0) );
const int HPanWidth = H11.Width()-HPanOffset;
HPan.LockedView( H, 0, H00.Width()+HPanOffset, HPanHeight, HPanWidth );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanWidth );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
ATop.View2x1( A0,
A1 );
Zeros( HPan.Width(), ATop.Width(), Z );
Zeros( HPan.Width(), HPan.Width(), SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, UPPER, offset, HPanCopy );
SetDiagonalToOne( RIGHT, offset, HPanCopy );
Herk( LOWER, ADJOINT, C(1), HPanCopy, C(0), SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( ADJOINT, NORMAL, C(1), HPanCopy, ATop, C(0), Z );
Trsm( LEFT, LOWER, NORMAL, NON_UNIT, C(1), SInv, Z );
Gemm( NORMAL, NORMAL, C(-1), HPanCopy, Z, C(1), ATop );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LU( DistMatrix<F>& A, DistMatrix<int,VC,STAR>& p )
{
#ifndef RELEASE
CallStackEntry entry("LU");
if( A.Grid() != p.Grid() )
throw std::logic_error("{A,p} must be distributed over the same grid");
if( p.Viewing() &&
(std::min(A.Height(),A.Width()) != p.Height() || p.Width() != 1) )
throw std::logic_error
("p must be a vector of the same height as the min dimension of A.");
#endif
const Grid& g = A.Grid();
if( !p.Viewing() )
p.ResizeTo( std::min(A.Height(),A.Width()), 1 );
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g), AB(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<int,VC,STAR>
pT(g), p0(g),
pB(g), p1(g),
p2(g);
// Temporary distributions
DistMatrix<F, STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F, MC, STAR> A21_MC_STAR(g);
DistMatrix<F, STAR,VR > A12_STAR_VR(g);
DistMatrix<F, STAR,MR > A12_STAR_MR(g);
DistMatrix<int,STAR,STAR> p1_STAR_STAR(g);
// Pivot composition
std::vector<int> image, preimage;
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( p, pT,
pB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( pT, p0,
/**/ /**/
p1,
pB, p2 );
View1x2( AB, ABL, ABR );
const int pivotOffset = A01.Height();
A12_STAR_VR.AlignWith( A22 );
A12_STAR_MR.AlignWith( A22 );
A21_MC_STAR.AlignWith( A22 );
A11_STAR_STAR.ResizeTo( A11.Height(), A11.Width() );
p1_STAR_STAR.ResizeTo( p1.Height(), 1 );
//--------------------------------------------------------------------//
A21_MC_STAR = A21;
A11_STAR_STAR = A11;
lu::Panel( A11_STAR_STAR, A21_MC_STAR, p1_STAR_STAR, pivotOffset );
ComposePivots( p1_STAR_STAR, pivotOffset, image, preimage );
ApplyRowPivots( AB, image, preimage );
// Perhaps we should give up perfectly distributing this operation since
// it's total contribution is only O(n^2)
A12_STAR_VR = A12;
LocalTrsm
( LEFT, LOWER, NORMAL, UNIT, F(1), A11_STAR_STAR, A12_STAR_VR );
A12_STAR_MR = A12_STAR_VR;
LocalGemm( NORMAL, NORMAL, F(-1), A21_MC_STAR, A12_STAR_MR, F(1), A22 );
A11 = A11_STAR_STAR;
A12 = A12_STAR_MR;
A21 = A21_MC_STAR;
p1 = p1_STAR_STAR;
//--------------------------------------------------------------------//
A12_STAR_VR.FreeAlignments();
A12_STAR_MR.FreeAlignments();
A21_MC_STAR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlidePartitionDown
( pT, p0,
p1,
/**/ /**/
pB, p2 );
}
}
inline void
SymmLLC
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::SymmLLC");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<T>
ATL(g), ATR(g), A00(g), A01(g), A02(g), AColPan(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g), ARowPan(g),
A20(g), A21(g), A22(g);
DistMatrix<T>
BT(g), B0(g),
BB(g), B1(g),
B2(g);
DistMatrix<T>
CT(g), C0(g), CAbove(g),
CB(g), C1(g), CBelow(g),
C2(g);
// Temporary distributions
DistMatrix<T,MC, STAR> AColPan_MC_STAR(g);
DistMatrix<T,STAR,MC > ARowPan_STAR_MC(g);
DistMatrix<T,MR, STAR> B1Trans_MR_STAR(g);
B1Trans_MR_STAR.AlignWith( C );
// Start the algorithm
Scale( beta, C );
LockedPartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDown
( B, BT,
BB, 0 );
PartitionDown
( C, CT,
CB, 0 );
while( CB.Height() > 0 )
{
LockedRepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
RepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
LockedView1x2( ARowPan, A10, A11 );
LockedView2x1
( AColPan, A11,
A21 );
View2x1
( CAbove, C0,
C1 );
View2x1
( CBelow, C1,
C2 );
AColPan_MC_STAR.AlignWith( CBelow );
ARowPan_STAR_MC.AlignWith( CAbove );
//--------------------------------------------------------------------//
AColPan_MC_STAR = AColPan;
ARowPan_STAR_MC = ARowPan;
MakeTrapezoidal( LEFT, LOWER, 0, AColPan_MC_STAR );
MakeTrapezoidal( RIGHT, LOWER, -1, ARowPan_STAR_MC );
B1Trans_MR_STAR.TransposeFrom( B1 );
LocalGemm
( NORMAL, TRANSPOSE,
alpha, AColPan_MC_STAR, B1Trans_MR_STAR, T(1), CBelow );
LocalGemm
( TRANSPOSE, TRANSPOSE,
alpha, ARowPan_STAR_MC, B1Trans_MR_STAR, T(1), CAbove );
//--------------------------------------------------------------------//
AColPan_MC_STAR.FreeAlignments();
ARowPan_STAR_MC.FreeAlignments();
SlideLockedPartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void LSquare
( DistMatrix<Complex<R> >& A,
DistMatrix<Complex<R>,STAR,STAR>& t )
{
#ifndef RELEASE
CallStackEntry entry("hermitian_tridiag::LSquare");
if( A.Grid() != t.Grid() )
throw std::logic_error("{A,t} must be distributed over the same grid");
#endif
const Grid& g = A.Grid();
#ifndef RELEASE
if( g.Height() != g.Width() )
throw std::logic_error("The process grid must be square");
if( A.Height() != A.Width() )
throw std::logic_error("A must be square");
if( t.Viewing() )
throw std::logic_error("t must not be a view");
#endif
typedef Complex<R> C;
DistMatrix<C,MD,STAR> tDiag(g);
tDiag.AlignWithDiagonal( A, -1 );
tDiag.ResizeTo( A.Height()-1, 1 );
// Matrix views
DistMatrix<C>
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<C,MD,STAR> tT(g), t0(g),
tB(g), t1(g),
t2(g);
// Temporary distributions
DistMatrix<C> WPan(g);
DistMatrix<C,STAR,STAR> t1_STAR_STAR(g);
DistMatrix<C,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<C,MC, STAR> APan_MC_STAR(g), A11_MC_STAR(g),
A21_MC_STAR(g);
DistMatrix<C,MR, STAR> APan_MR_STAR(g), A11_MR_STAR(g),
A21_MR_STAR(g);
DistMatrix<C,MC, STAR> WPan_MC_STAR(g), W11_MC_STAR(g),
W21_MC_STAR(g);
DistMatrix<C,MR, STAR> WPan_MR_STAR(g), W11_MR_STAR(g),
W21_MR_STAR(g);
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( tDiag, tT,
tB, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2 );
if( A22.Height() > 0 )
{
WPan.AlignWith( A11 );
APan_MC_STAR.AlignWith( A11 );
WPan_MC_STAR.AlignWith( A11 );
APan_MR_STAR.AlignWith( A11 );
WPan_MR_STAR.AlignWith( A11 );
//----------------------------------------------------------------//
WPan.ResizeTo( ABR.Height(), A11.Width() );
APan_MC_STAR.ResizeTo( ABR.Height(), A11.Width() );
WPan_MC_STAR.ResizeTo( ABR.Height(), A11.Width() );
APan_MR_STAR.ResizeTo( ABR.Height(), A11.Width() );
WPan_MR_STAR.ResizeTo( ABR.Height(), A11.Width() );
hermitian_tridiag::PanelLSquare
( ABR, WPan, t1,
APan_MC_STAR, APan_MR_STAR, WPan_MC_STAR, WPan_MR_STAR );
PartitionDown
( APan_MC_STAR, A11_MC_STAR,
A21_MC_STAR, A11.Height() );
PartitionDown
( APan_MR_STAR, A11_MR_STAR,
A21_MR_STAR, A11.Height() );
PartitionDown
( WPan_MC_STAR, W11_MC_STAR,
W21_MC_STAR, A11.Height() );
PartitionDown
( WPan_MR_STAR, W11_MR_STAR,
W21_MR_STAR, A11.Height() );
LocalTrr2k
( LOWER, ADJOINT, ADJOINT,
C(-1), A21_MC_STAR, W21_MR_STAR,
W21_MC_STAR, A21_MR_STAR,
C(1), A22 );
//----------------------------------------------------------------//
WPan_MR_STAR.FreeAlignments();
APan_MR_STAR.FreeAlignments();
WPan_MC_STAR.FreeAlignments();
APan_MC_STAR.FreeAlignments();
WPan.FreeAlignments();
}
else
{
A11_STAR_STAR = A11;
t1_STAR_STAR.ResizeTo( t1.Height(), 1 );
HermitianTridiag
( LOWER, A11_STAR_STAR.Matrix(), t1_STAR_STAR.Matrix() );
A11 = A11_STAR_STAR;
t1 = t1_STAR_STAR;
}
SlidePartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
// Redistribute from matrix-diagonal form to fully replicated
t = tDiag;
}
inline void
internal::ApplyPackedReflectorsLLVF
( int offset,
const DistMatrix<R,MC,MR>& H,
DistMatrix<R,MC,MR>& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsLLVF");
if( H.Grid() != A.Grid() )
throw std::logic_error("{H,A} must be distributed over the same grid");
if( offset > 0 )
throw std::logic_error("Transforms cannot extend above matrix");
if( offset < -H.Height() )
throw std::logic_error("Transforms cannot extend below matrix");
if( H.Height() != A.Height() )
throw std::logic_error
("Height of transforms must equal height of target matrix");
#endif
const Grid& g = H.Grid();
// Matrix views
DistMatrix<R,MC,MR>
HTL(g), HTR(g), H00(g), H01(g), H02(g), HPan(g), HPanCopy(g),
HBL(g), HBR(g), H10(g), H11(g), H12(g),
H20(g), H21(g), H22(g);
DistMatrix<R,MC,MR>
AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<R,VC, STAR> HPan_VC_STAR(g);
DistMatrix<R,MC, STAR> HPan_MC_STAR(g);
DistMatrix<R,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<R,STAR,MR > Z_STAR_MR(g);
DistMatrix<R,STAR,VR > Z_STAR_VR(g);
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
int HPanHeight = H11.Height() + H21.Height();
int HPanWidth = std::min( H11.Width(), std::max(HPanHeight+offset,0) );
HPan.LockedView( H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
HPan_MC_STAR.AlignWith( AB );
Z_STAR_MR.AlignWith( AB );
Z_STAR_VR.AlignWith( AB );
Z_STAR_MR.ResizeTo( HPanWidth, AB.Width() );
SInv_STAR_STAR.ResizeTo( HPanWidth, HPanWidth );
Zero( SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, LOWER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
HPan_VC_STAR = HPanCopy;
Syrk
( UPPER, TRANSPOSE,
(R)1, HPan_VC_STAR.LockedLocalMatrix(),
(R)0, SInv_STAR_STAR.LocalMatrix() );
SInv_STAR_STAR.SumOverGrid();
HalveMainDiagonal( SInv_STAR_STAR );
HPan_MC_STAR = HPanCopy;
internal::LocalGemm
( TRANSPOSE, NORMAL,
(R)1, HPan_MC_STAR, AB, (R)0, Z_STAR_MR );
Z_STAR_VR.SumScatterFrom( Z_STAR_MR );
internal::LocalTrsm
( LEFT, UPPER, TRANSPOSE, NON_UNIT,
(R)1, SInv_STAR_STAR, Z_STAR_VR );
Z_STAR_MR = Z_STAR_VR;
internal::LocalGemm
( NORMAL, NORMAL, (R)-1, HPan_MC_STAR, Z_STAR_MR, (R)1, AB );
//--------------------------------------------------------------------//
HPan_MC_STAR.FreeAlignments();
Z_STAR_MR.FreeAlignments();
Z_STAR_VR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
GemmNNDot
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmNNDot");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
if( A.Height() != C.Height() ||
B.Width() != C.Width() ||
A.Width() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal GemmNNDot: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B ~ " << B.Height() << " x " << B.Width() << "\n"
<< " C ~ " << C.Height() << " x " << C.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = A.Grid();
if( A.Height() > B.Width() )
{
// Matrix views
DistMatrix<T> AT(g), AB(g),
A0(g), A1(g), A2(g);
DistMatrix<T> BL(g), B0(g),
BR(g), B1(g),
B2(g);
DistMatrix<T> CT(g), C0(g), C1L(g), C1R(g),
CB(g), C1(g), C10(g), C11(g), C12(g),
C2(g);
// Temporary distributions
DistMatrix<T,STAR,VC> A1_STAR_VC(g);
DistMatrix<T,VC,STAR> B1_VC_STAR(g);
DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);
// Star the algorithm
Scale( beta, C );
LockedPartitionDown
( A, AT,
AB, 0 );
PartitionDown
( C, CT,
CB, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
RepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
A1_STAR_VC = A1;
B1_VC_STAR.AlignWith( A1_STAR_VC );
LockedPartitionRight( B, BL, BR, 0 );
PartitionRight( C1, C1L, C1R, 0 );
while( BR.Width() > 0 )
{
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
RepartitionRight
( C1L, /**/ C1R,
C10, /**/ C11, C12 );
Zeros( C11.Height(), C11.Width(), C11_STAR_STAR );
//------------------------------------------------------------//
B1_VC_STAR = B1;
LocalGemm
( NORMAL, NORMAL,
alpha, A1_STAR_VC, B1_VC_STAR, T(0), C11_STAR_STAR );
C11.SumScatterUpdate( T(1), C11_STAR_STAR );
//------------------------------------------------------------//
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( C1L, /**/ C1R,
C10, C11, /**/ C12 );
}
B1_VC_STAR.FreeAlignments();
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
}
else
{
// Matrix views
DistMatrix<T> AT(g), AB(g),
A0(g), A1(g), A2(g);
DistMatrix<T> BL(g), B0(g),
BR(g), B1(g),
B2(g);
DistMatrix<T>
CL(g), CR(g), C1T(g), C01(g),
C0(g), C1(g), C2(g), C1B(g), C11(g),
C21(g);
// Temporary distributions
DistMatrix<T,STAR,VR> A1_STAR_VR(g);
DistMatrix<T,VR,STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);
// Star the algorithm
Scale( beta, C );
LockedPartitionRight( B, BL, BR, 0 );
PartitionRight( C, CL, CR, 0 );
while( BR.Width() > 0 )
{
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
RepartitionRight
( CL, /**/ CR,
C0, /**/ C1, C2 );
B1_VR_STAR = B1;
A1_STAR_VR.AlignWith( B1_VR_STAR );
LockedPartitionDown
( A, AT,
AB, 0 );
PartitionDown
( C1, C1T,
C1B, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
RepartitionDown
( C1T, C01,
/***/ /***/
C11,
C1B, C21 );
Zeros( C11.Height(), C11.Width(), C11_STAR_STAR );
//------------------------------------------------------------//
A1_STAR_VR = A1;
LocalGemm
( NORMAL, NORMAL,
alpha, A1_STAR_VR, B1_VR_STAR, T(0), C11_STAR_STAR );
C11.SumScatterUpdate( T(1), C11_STAR_STAR );
//------------------------------------------------------------//
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlidePartitionDown
( C1T, C01,
C11,
/***/ /***/
C1B, C21 );
}
A1_STAR_VR.FreeAlignments();
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
}
}
#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
GemmNNB
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmNNB");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
if( A.Height() != C.Height() ||
B.Width() != C.Width() ||
A.Width() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal GemmNNB: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B ~ " << B.Height() << " x " << B.Width() << "\n"
<< " C ~ " << C.Height() << " x " << C.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<T> AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<T> CT(g), C0(g),
CB(g), C1(g),
C2(g);
// Temporary distributions
DistMatrix<T,STAR,MC> A1_STAR_MC(g);
DistMatrix<T,MR,STAR> D1Trans_MR_STAR(g);
A1_STAR_MC.AlignWith( B );
D1Trans_MR_STAR.AlignWith( B );
// Start the algorithm
Scale( beta, C );
LockedPartitionDown
( A, AT,
AB, 0 );
PartitionDown
( C, CT,
CB, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
RepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
Zeros( C1.Width(), C1.Height(), D1Trans_MR_STAR );
//--------------------------------------------------------------------//
A1_STAR_MC = A1; // A1[*,MC] <- A1[MC,MR]
// D1^T[MR,* ] := alpha B^T[MR,MC] A1^T[MC,* ]
LocalGemm
( TRANSPOSE, TRANSPOSE, alpha, B, A1_STAR_MC, T(0), D1Trans_MR_STAR );
C1.TransposeSumScatterUpdate( T(1), D1Trans_MR_STAR );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
