inline void
TrmmLLNA
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLLNA");
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 TrmmLLNA: \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();
DistMatrix<T>
XL(g), XR(g),
X0(g), X1(g), X2(g);
DistMatrix<T,VR, STAR> X1_VR_STAR(g);
DistMatrix<T,STAR,MR > X1Trans_STAR_MR(g);
DistMatrix<T,MC, STAR> Z1_MC_STAR(g);
X1_VR_STAR.AlignWith( L );
X1Trans_STAR_MR.AlignWith( L );
Z1_MC_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_MC_STAR );
//--------------------------------------------------------------------//
X1_VR_STAR = X1;
X1Trans_STAR_MR.TransposeFrom( X1_VR_STAR );
LocalTrmmAccumulateLLN
( TRANSPOSE, diag, alpha, L, X1Trans_STAR_MR, Z1_MC_STAR );
X1.SumScatterFrom( Z1_MC_STAR );
//--------------------------------------------------------------------//
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
GemmNNA
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmNNA");
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 GemmNNA: \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> BL(g), BR(g),
B0(g), B1(g), B2(g);
DistMatrix<T> CL(g), CR(g),
C0(g), C1(g), C2(g);
// Temporary distributions
DistMatrix<T,VR,STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR> B1Trans_STAR_MR(g);
DistMatrix<T,MC,STAR> D1_MC_STAR(g);
B1_VR_STAR.AlignWith( A );
B1Trans_STAR_MR.AlignWith( A );
D1_MC_STAR.AlignWith( A );
// Start 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 );
Zeros( C1.Height(), C1.Width(), D1_MC_STAR );
//--------------------------------------------------------------------//
B1_VR_STAR = B1;
B1Trans_STAR_MR.TransposeFrom( B1_VR_STAR );
// D1[MC,*] := alpha A[MC,MR] B1[MR,*]
LocalGemm
( NORMAL, TRANSPOSE, alpha, A, B1Trans_STAR_MR, T(0), D1_MC_STAR );
// C1[MC,MR] += scattered result of D1[MC,*] summed over grid rows
C1.SumScatterUpdate( T(1), D1_MC_STAR );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
}
#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 );
}
}
void
ApplyPackedReflectorsRUHF
( Conjugation conjugation, int offset,
const DistMatrix<Complex<R> >& H,
const DistMatrix<Complex<R>,MD,STAR>& t,
DistMatrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsRUHF");
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 || offset > H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Width() )
throw std::logic_error
("Length of transforms must equal width 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();
DistMatrix<C>
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>
AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<C,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
DistMatrix<C,STAR,VR > HPan_STAR_VR(g);
DistMatrix<C,STAR,MR > HPan_STAR_MR(g);
DistMatrix<C,STAR,STAR> t1_STAR_STAR(g);
DistMatrix<C,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<C,STAR,MC > ZAdj_STAR_MC(g);
DistMatrix<C,STAR,VC > ZAdj_STAR_VC(g);
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionRight
( A, AL, AR, 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 HPanWidth = H11.Width() + H12.Width();
const int HPanHeight =
std::min( H11.Height(), std::max(HPanWidth-offset,0) );
LockedView( HPan, H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanHeight );
RepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
HPan_STAR_MR.AlignWith( AR );
ZAdj_STAR_MC.AlignWith( AR );
ZAdj_STAR_VC.AlignWith( AR );
Zeros( HPanHeight, AR.Height(), ZAdj_STAR_MC );
Zeros( HPanHeight, HPanHeight, SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, UPPER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
HPan_STAR_VR = HPanCopy;
Herk
( UPPER, NORMAL,
C(1), HPan_STAR_VR.LockedLocalMatrix(),
C(0), SInv_STAR_STAR.LocalMatrix() );
SInv_STAR_STAR.SumOverGrid();
t1_STAR_STAR = t1;
FixDiagonal( conjugation, t1_STAR_STAR, SInv_STAR_STAR );
HPan_STAR_MR = HPan_STAR_VR;
LocalGemm
( NORMAL, ADJOINT, C(1), HPan_STAR_MR, AR, C(0), ZAdj_STAR_MC );
ZAdj_STAR_VC.SumScatterFrom( ZAdj_STAR_MC );
LocalTrsm
( LEFT, UPPER, ADJOINT, NON_UNIT, C(1), SInv_STAR_STAR, ZAdj_STAR_VC );
ZAdj_STAR_MC = ZAdj_STAR_VC;
LocalGemm
( ADJOINT, NORMAL, C(-1), ZAdj_STAR_MC, HPan_STAR_MR, C(1), AR );
//--------------------------------------------------------------------//
HPan_STAR_MR.FreeAlignments();
ZAdj_STAR_MC.FreeAlignments();
ZAdj_STAR_VC.FreeAlignments();
SlidePartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
ApplyPackedReflectorsRUHF
( int offset, const Matrix<R>& H, Matrix<R>& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsRUHF");
if( offset < 0 || offset > H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Width() )
throw std::logic_error
("Length of transforms must equal width of target matrix");
#endif
Matrix<R>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<R>
AL, AR,
A0, A1, A2;
Matrix<R> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
PartitionRight( A, AL, AR, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/**************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
RepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
const int HPanWidth = H11.Width() + H12.Width();
const int HPanHeight =
std::min( H11.Height(), std::max(HPanWidth-offset,0) );
LockedView( HPan, H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
Zeros( AR.Height(), HPanHeight, Z );
Zeros( HPanHeight, HPanHeight, SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, UPPER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
Syrk( UPPER, NORMAL, R(1), HPanCopy, R(0), SInv );
HalveMainDiagonal( SInv );
Gemm( NORMAL, TRANSPOSE, R(1), AR, HPanCopy, R(0), Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, R(1), SInv, Z );
Gemm( NORMAL, NORMAL, R(-1), Z, HPanCopy, R(1), AR );
//--------------------------------------------------------------------//
SlidePartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void
ApplyPackedReflectorsRUHF
( Conjugation conjugation, int offset,
const Matrix<Complex<R> >& H,
const Matrix<Complex<R> >& t,
Matrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsRUHF");
if( offset < 0 || offset > H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Width() )
throw std::logic_error
("Length of transforms must equal width 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, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<C>
AL, AR,
A0, A1, A2;
Matrix<C>
tT, t0,
tB, t1,
t2;
Matrix<C> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionRight
( A, AL, AR, 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 HPanWidth = H11.Width() + H12.Width();
const int HPanHeight =
std::min( H11.Height(), std::max(HPanWidth-offset,0) );
LockedView( HPan, H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanHeight );
RepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
Zeros( AR.Height(), HPanHeight, Z );
Zeros( HPanHeight, HPanHeight, SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, UPPER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
Herk( UPPER, NORMAL, C(1), HPanCopy, C(0), SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, ADJOINT, C(1), AR, HPanCopy, C(0), Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, C(1), SInv, Z );
Gemm( NORMAL, NORMAL, C(-1), Z, HPanCopy, C(1), AR );
//--------------------------------------------------------------------//
SlidePartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
internal::GemmTNA
( Orientation orientationOfA,
T alpha, const DistMatrix<T,MC,MR>& A,
const DistMatrix<T,MC,MR>& B,
T beta, DistMatrix<T,MC,MR>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmTNA");
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 )
throw std::logic_error("GemmTNA assumes A is (Conjugate)Transposed");
if( A.Width() != C.Height() ||
B.Width() != C.Width() ||
A.Height() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal GemmTNA: \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,MC,MR> BL(g), BR(g),
B0(g), B1(g), B2(g);
DistMatrix<T,MC,MR> CL(g), CR(g),
C0(g), C1(g), C2(g);
// Temporary distributions
DistMatrix<T,MC,STAR> B1_MC_STAR(g);
DistMatrix<T,MR,STAR> D1_MR_STAR(g);
DistMatrix<T,MR,MC > D1_MR_MC(g);
DistMatrix<T,MC,MR > D1(g);
// Start the algorithm
Scal( 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_MC_STAR.AlignWith( A );
D1_MR_STAR.AlignWith( A );
D1_MR_STAR.ResizeTo( C1.Height(), C1.Width() );
D1.AlignWith( C1 );
//--------------------------------------------------------------------//
B1_MC_STAR = B1; // B1[MC,*] <- B1[MC,MR]
// D1[MR,*] := alpha (A1[MC,MR])^T B1[MC,*]
// = alpha (A1^T)[MR,MC] B1[MC,*]
internal::LocalGemm
( orientationOfA, NORMAL, alpha, A, B1_MC_STAR, (T)0, D1_MR_STAR );
// C1[MC,MR] += scattered & transposed D1[MR,*] summed over grid cols
D1_MR_MC.SumScatterFrom( D1_MR_STAR );
D1 = D1_MR_MC;
Axpy( (T)1, D1, C1 );
//--------------------------------------------------------------------//
B1_MC_STAR.FreeAlignments();
D1_MR_STAR.FreeAlignments();
D1.FreeAlignments();
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
SymmLLA
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::SymmLLA");
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>
BL(g), BR(g),
B0(g), B1(g), B2(g);
DistMatrix<T>
CL(g), CR(g),
C0(g), C1(g), C2(g);
DistMatrix<T,MC,STAR> B1_MC_STAR(g);
DistMatrix<T,VR,STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR> B1Trans_STAR_MR(g);
DistMatrix<T> Z1(g);
DistMatrix<T,MC,STAR> Z1_MC_STAR(g);
DistMatrix<T,MR,STAR> Z1_MR_STAR(g);
DistMatrix<T,MR,MC > Z1_MR_MC(g);
B1_MC_STAR.AlignWith( A );
B1_VR_STAR.AlignWith( A );
B1Trans_STAR_MR.AlignWith( A );
Z1_MC_STAR.AlignWith( A );
Z1_MR_STAR.AlignWith( A );
Scale( beta, C );
LockedPartitionRight
( B, BL, BR, 0 );
PartitionRight
( C, CL, CR, 0 );
while( CL.Width() < C.Width() )
{
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
RepartitionRight
( CL, /**/ CR,
C0, /**/ C1, C2 );
Z1.AlignWith( C1 );
Zeros( C1.Height(), C1.Width(), Z1_MC_STAR );
Zeros( C1.Height(), C1.Width(), Z1_MR_STAR );
//--------------------------------------------------------------------//
B1_MC_STAR = B1;
B1_VR_STAR = B1_MC_STAR;
B1Trans_STAR_MR.TransposeFrom( B1_VR_STAR );
LocalSymmetricAccumulateLL
( TRANSPOSE,
alpha, A, B1_MC_STAR, B1Trans_STAR_MR, Z1_MC_STAR, Z1_MR_STAR );
Z1_MR_MC.SumScatterFrom( Z1_MR_STAR );
Z1 = Z1_MR_MC;
Z1.SumScatterUpdate( T(1), Z1_MC_STAR );
Axpy( T(1), Z1, C1 );
//--------------------------------------------------------------------//
Z1.FreeAlignments();
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void
RLVF
( Conjugation conjugation, Int offset,
const DistMatrix<F>& H, const DistMatrix<F,MD,STAR>& t, DistMatrix<F>& A )
{
#ifndef RELEASE
CallStackEntry cse("apply_packed_reflectors::RLVF");
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>
AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<F,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
DistMatrix<F,VC, STAR> HPan_VC_STAR(g);
DistMatrix<F,MR, STAR> HPan_MR_STAR(g);
DistMatrix<F,STAR,STAR> t1_STAR_STAR(g);
DistMatrix<F,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<F,STAR,MC > ZAdj_STAR_MC(g);
DistMatrix<F,STAR,VC > ZAdj_STAR_VC(g);
LockedPartitionDownOffsetDiagonal
( offset,
H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionRight( A, AL, AR, 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 );
RepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2, H11.Height() );
LockedView2x1( HPan, H11, H21 );
HPan_MR_STAR.AlignWith( AR );
ZAdj_STAR_MC.AlignWith( AR );
ZAdj_STAR_VC.AlignWith( AR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTriangular( LOWER, HPanCopy );
SetDiagonal( HPanCopy, F(1) );
HPan_VC_STAR = HPanCopy;
Zeros( SInv_STAR_STAR, HPan.Width(), HPan.Width() );
Herk
( UPPER, ADJOINT,
F(1), HPan_VC_STAR.LockedMatrix(),
F(0), SInv_STAR_STAR.Matrix() );
SInv_STAR_STAR.SumOverGrid();
t1_STAR_STAR = t1;
FixDiagonal( conjugation, t1_STAR_STAR, SInv_STAR_STAR );
HPan_MR_STAR = HPan_VC_STAR;
LocalGemm( ADJOINT, ADJOINT, F(1), HPan_MR_STAR, AR, ZAdj_STAR_MC );
ZAdj_STAR_VC.SumScatterFrom( ZAdj_STAR_MC );
LocalTrsm
( LEFT, UPPER, ADJOINT, NON_UNIT, F(1), SInv_STAR_STAR, ZAdj_STAR_VC );
ZAdj_STAR_MC = ZAdj_STAR_VC;
LocalGemm
( ADJOINT, ADJOINT, F(-1), ZAdj_STAR_MC, HPan_MR_STAR, F(1), AR );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
}
}
inline void
TrsmRUN
( UnitOrNonUnit diag, F alpha, const DistMatrix<F>& U, DistMatrix<F>& X,
bool checkIfSingular )
{
#ifndef RELEASE
PushCallStack("internal::TrsmRUN");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<F> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,MR > U12_STAR_MR(g);
DistMatrix<F,VC, STAR> X1_VC_STAR(g);
DistMatrix<F,STAR,MC > X1Trans_STAR_MC(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
X1_VC_STAR.AlignWith( X2 );
X1Trans_STAR_MC.AlignWith( X2 );
U12_STAR_MR.AlignWith( X2 );
//--------------------------------------------------------------------//
U11_STAR_STAR = U11;
X1_VC_STAR = X1;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, X1_VC_STAR,
checkIfSingular );
X1Trans_STAR_MC.TransposeFrom( X1_VC_STAR );
X1.TransposeFrom( X1Trans_STAR_MC );
U12_STAR_MR = U12;
// X2[MC,MR] -= X1[MC,* ] U12[* ,MR]
// = X1^T[* ,MC] U12[* ,MR]
LocalGemm
( TRANSPOSE, NORMAL, F(-1), X1Trans_STAR_MC, U12_STAR_MR, F(1), X2 );
//--------------------------------------------------------------------//
X1_VC_STAR.FreeAlignments();
X1Trans_STAR_MC.FreeAlignments();
U12_STAR_MR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
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);
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
// TODO: Think of how to make this safer with respect to alignment issues
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 );
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 );
//--------------------------------------------------------------------//
const int currentRow = a01.Height();
// Store the index/value of the pivot candidate in A
F pivot = alpha11.GetLocal(0,0);
int pivotRow = currentRow;
for( int i=0; i<a21.Height(); ++i )
{
F value = a21.GetLocal(i,0);
if( FastAbs(value) > FastAbs(pivot) )
{
pivot = value;
pivotRow = currentRow + i + 1;
}
}
// Update the pivot candidate to include local data from B
for( int i=0; i<B.LocalHeight(); ++i )
{
F value = b1.GetLocal(i,0);
if( FastAbs(value) > FastAbs(pivot) )
{
pivot = value;
pivotRow = A.Height() + colShift + i*r;
}
}
// Fill the send buffer with:
// [ pivotValue | pivot row data | pivotRow ]
if( pivotRow < A.Height() )
{
sendBufFloat[0] = A.GetLocal(pivotRow,a10.Width());
const int ALDim = A.LocalLDim();
const F* ABuffer = A.LocalBuffer(pivotRow,0);
for( int j=0; j<width; ++j )
sendBufFloat[j+1] = ABuffer[j*ALDim];
}
else
{
const int localRow = ((pivotRow-A.Height())-colShift)/r;
sendBufFloat[0] = b1.GetLocal(localRow,0);
const int BLDim = B.LocalLDim();
const F* BBuffer = B.LocalBuffer(localRow,0);
for( int j=0; j<width; ++j )
sendBufFloat[j+1] = BBuffer[j*BLDim];
}
*sendBufInt = pivotRow;
// Communicate to establish the pivot information
mpi::AllReduce
( &sendData[0], &recvData[0], numBytes, PivotOp<F>(), g.ColComm() );
// Update the pivot vector
pivotRow = *recvBufInt;
p.SetLocal(currentRow,0,pivotRow+pivotOffset);
// Copy the current row into the pivot row
if( pivotRow < A.Height() )
{
const int ALDim = A.LocalLDim();
F* ASetBuffer = A.LocalBuffer(pivotRow,0);
const F* AGetBuffer = A.LocalBuffer(currentRow,0);
for( int j=0; j<width; ++j )
ASetBuffer[j*ALDim] = AGetBuffer[j*ALDim];
}
else
{
const int ownerRank = (colAlignment+(pivotRow-A.Height())) % r;
if( g.Row() == ownerRank )
{
const int localRow = ((pivotRow-A.Height())-colShift) / r;
const int ALDim = A.LocalLDim();
const int BLDim = B.LocalLDim();
F* BBuffer = B.LocalBuffer(localRow,0);
const F* ABuffer = A.LocalBuffer(currentRow,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(currentRow,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
const F alpha = alpha11.GetLocal(0,0);
if( alpha == F(0) )
throw SingularMatrixException();
const F alpha11Inv = F(1) / alpha;
Scale( alpha11Inv, a21.LocalMatrix() );
Scale( 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 );
}
PopBlocksizeStack();
#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
HemmRUC
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::HemmRUC");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
throw std::logic_error("{A,B,C} must be distributed on 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> BL(g), BR(g),
B0(g), B1(g), B2(g);
DistMatrix<T> CL(g), CR(g),
C0(g), C1(g), C2(g),
CLeft(g), CRight(g);
// Temporary distributions
DistMatrix<T,MC,STAR> B1_MC_STAR(g);
DistMatrix<T,VR, STAR> AColPan_VR_STAR(g);
DistMatrix<T,STAR,MR > AColPanAdj_STAR_MR(g);
DistMatrix<T,MR, STAR> ARowPanAdj_MR_STAR(g);
B1_MC_STAR.AlignWith( C );
// Start the algorithm
Scale( beta, C );
LockedPartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionRight( B, BL, BR, 0 );
PartitionRight( C, CL, CR, 0 );
while( CR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
RepartitionRight
( CL, /**/ CR,
C0, /**/ C1, C2 );
ARowPan.LockedView1x2( A11, A12 );
AColPan.LockedView2x1
( A01,
A11 );
CLeft.View1x2( C0, C1 );
CRight.View1x2( C1, C2 );
AColPan_VR_STAR.AlignWith( CLeft );
AColPanAdj_STAR_MR.AlignWith( CLeft );
ARowPanAdj_MR_STAR.AlignWith( CRight );
//--------------------------------------------------------------------//
B1_MC_STAR = B1;
AColPan_VR_STAR = AColPan;
AColPanAdj_STAR_MR.AdjointFrom( AColPan_VR_STAR );
ARowPanAdj_MR_STAR.AdjointFrom( ARowPan );
MakeTrapezoidal( LEFT, LOWER, 0, ARowPanAdj_MR_STAR );
MakeTrapezoidal( RIGHT, LOWER, -1, AColPanAdj_STAR_MR );
LocalGemm
( NORMAL, ADJOINT,
alpha, B1_MC_STAR, ARowPanAdj_MR_STAR, T(1), CRight );
LocalGemm
( NORMAL, NORMAL,
alpha, B1_MC_STAR, AColPanAdj_STAR_MR, T(1), CLeft );
//--------------------------------------------------------------------//
AColPan_VR_STAR.FreeAlignments();
AColPanAdj_STAR_MR.FreeAlignments();
ARowPanAdj_MR_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ 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
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
GemmTTA
( Orientation orientationOfA,
Orientation orientationOfB,
T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmTTA");
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
("GemmTTA 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 GemmTTA: \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> BT(g), B0(g),
BB(g), B1(g),
B2(g);
DistMatrix<T> CL(g), CR(g),
C0(g), C1(g), C2(g);
// Temporary distributions
DistMatrix<T,STAR,MC > B1_STAR_MC(g);
DistMatrix<T,MR, STAR> D1_MR_STAR(g);
DistMatrix<T,MR, MC > D1_MR_MC(g);
DistMatrix<T> D1(g);
B1_STAR_MC.AlignWith( A );
D1_MR_STAR.AlignWith( A );
// Start the algorithm
Scale( beta, C );
LockedPartitionDown
( B, BT,
BB, 0 );
PartitionRight( C, CL, CR, 0 );
while( BB.Height() > 0 )
{
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
RepartitionRight
( CL, /**/ CR,
C0, /**/ C1, C2 );
D1.AlignWith( C1 );
Zeros( C1.Height(), C1.Width(), D1_MR_STAR );
//--------------------------------------------------------------------//
B1_STAR_MC = B1; // B1[*,MC] <- B1[MC,MR]
// D1[MR,*] := alpha (A[MC,MR])^T (B1[*,MC])^T
// = alpha (A^T)[MR,MC] (B1^T)[MC,*]
LocalGemm
( orientationOfA, orientationOfB,
alpha, A, B1_STAR_MC, T(0), D1_MR_STAR );
// C1[MC,MR] += scattered & transposed D1[MR,*] summed over grid cols
D1_MR_MC.SumScatterFrom( D1_MR_STAR );
D1 = D1_MR_MC;
Axpy( T(1), D1, C1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void
RLVF
( Conjugation conjugation, Int offset,
const Matrix<F>& H, const Matrix<F>& t, Matrix<F>& A )
{
#ifndef RELEASE
CallStackEntry cse("apply_packed_reflectors::RLVF");
// 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>
AL, AR,
A0, 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 );
PartitionRight( A, AL, AR, 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 );
RepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2, H11.Height() );
LockedView2x1( HPan, H11, H21 );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTriangular( LOWER, HPanCopy );
SetDiagonal( HPanCopy, F(1) );
Herk( UPPER, ADJOINT, F(1), HPanCopy, SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, NORMAL, F(1), AR, HPanCopy, Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, F(1), SInv, Z );
Gemm( NORMAL, ADJOINT, F(-1), Z, HPanCopy, F(1), AR );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
}
}
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
ApplyPackedReflectorsRUHF
( int offset,
const DistMatrix<R>& H,
DistMatrix<R>& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsRUHF");
if( H.Grid() != A.Grid() )
throw std::logic_error("{H,A} must be distributed over the same grid");
if( offset < 0 || offset > H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Width() )
throw std::logic_error
("Length of transforms must equal width of target matrix");
#endif
const Grid& g = H.Grid();
DistMatrix<R>
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>
AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<R,STAR,VR > HPan_STAR_VR(g);
DistMatrix<R,STAR,MR > HPan_STAR_MR(g);
DistMatrix<R,STAR,STAR> SInv_STAR_STAR(g);
DistMatrix<R,STAR,MC > ZTrans_STAR_MC(g);
DistMatrix<R,STAR,VC > ZTrans_STAR_VC(g);
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
PartitionRight( A, AL, AR, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/**************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
RepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
const int HPanWidth = H11.Width() + H12.Width();
const int HPanHeight =
std::min( H11.Height(), std::max(HPanWidth-offset,0) );
LockedView( HPan, H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
HPan_STAR_MR.AlignWith( AR );
ZTrans_STAR_MC.AlignWith( AR );
ZTrans_STAR_VC.AlignWith( AR );
Zeros( HPanHeight, AR.Height(), ZTrans_STAR_MC );
Zeros( HPanHeight, HPanHeight, SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, UPPER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
HPan_STAR_VR = HPanCopy;
Syrk
( UPPER, NORMAL,
R(1), HPan_STAR_VR.LockedLocalMatrix(),
R(0), SInv_STAR_STAR.LocalMatrix() );
SInv_STAR_STAR.SumOverGrid();
HalveMainDiagonal( SInv_STAR_STAR );
HPan_STAR_MR = HPan_STAR_VR;
LocalGemm
( NORMAL, TRANSPOSE, R(1), HPan_STAR_MR, AR, R(0), ZTrans_STAR_MC );
ZTrans_STAR_VC.SumScatterFrom( ZTrans_STAR_MC );
LocalTrsm
( LEFT, UPPER, TRANSPOSE, NON_UNIT,
R(1), SInv_STAR_STAR, ZTrans_STAR_VC );
ZTrans_STAR_MC = ZTrans_STAR_VC;
LocalGemm
( TRANSPOSE, NORMAL, R(-1), ZTrans_STAR_MC, HPan_STAR_MR, R(1), AR );
//--------------------------------------------------------------------//
HPan_STAR_MR.FreeAlignments();
ZTrans_STAR_MC.FreeAlignments();
ZTrans_STAR_VC.FreeAlignments();
SlidePartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}