inline void
HermitianFromEVD
( UpperOrLower uplo,
Matrix<F>& A,
const Matrix<BASE(F)>& w,
const Matrix<F>& Z )
{
#ifndef RELEASE
CallStackEntry entry("HermitianFromEVD");
#endif
typedef BASE(F) R;
Matrix<F> ZL, ZR,
Z0, Z1, Z2;
Matrix<R> wT, w0,
wB, w1,
w2;
Matrix<F> Z1Copy, Y1;
A.ResizeTo( Z.Height(), Z.Height() );
if( uplo == LOWER )
MakeTrapezoidal( UPPER, A, 1 );
else
MakeTrapezoidal( LOWER, A, -1 );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
//--------------------------------------------------------------------//
Y1 = Z1Copy = Z1;
DiagonalScale( RIGHT, NORMAL, w1, Y1 );
Trrk( uplo, NORMAL, ADJOINT, F(1), Z1Copy, Y1, F(1), A );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
}
inline void
Syr2kUT
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
T beta, DistMatrix<T>& C,
bool conjugate=false )
{
#ifndef RELEASE
CallStackEntry entry("internal::Syr2kUT");
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.Width() != C.Height() ||
A.Width() != C.Width() ||
B.Width() != C.Height() ||
B.Width() != C.Width() ||
A.Height() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal Syr2kUT:\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();
const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );
// Matrix views
DistMatrix<T> AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<T> BT(g), B0(g),
BB(g), B1(g),
B2(g);
// Temporary distributions
DistMatrix<T,MR, STAR> A1Trans_MR_STAR(g);
DistMatrix<T,MR, STAR> B1Trans_MR_STAR(g);
DistMatrix<T,STAR,VR > A1_STAR_VR(g);
DistMatrix<T,STAR,VR > B1_STAR_VR(g);
DistMatrix<T,STAR,MC > A1_STAR_MC(g);
DistMatrix<T,STAR,MC > B1_STAR_MC(g);
A1Trans_MR_STAR.AlignWith( C );
B1Trans_MR_STAR.AlignWith( C );
A1_STAR_MC.AlignWith( C );
B1_STAR_MC.AlignWith( C );
// Start the algorithm
ScaleTrapezoid( beta, LEFT, UPPER, 0, C );
LockedPartitionDown
( A, AT,
AB, 0 );
LockedPartitionDown
( B, BT,
BB, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
//--------------------------------------------------------------------//
A1Trans_MR_STAR.TransposeFrom( A1 );
A1_STAR_VR.TransposeFrom( A1Trans_MR_STAR );
A1_STAR_MC = A1_STAR_VR;
B1Trans_MR_STAR.TransposeFrom( B1 );
B1_STAR_VR.TransposeFrom( B1Trans_MR_STAR );
B1_STAR_MC = B1_STAR_VR;
LocalTrr2k
( UPPER, orientation, TRANSPOSE, orientation, TRANSPOSE,
alpha, A1_STAR_MC, B1Trans_MR_STAR,
B1_STAR_MC, A1Trans_MR_STAR,
T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
}
}
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
RUVF
( Conjugation conjugation, Int offset,
const Matrix<F>& H, const Matrix<F>& t, Matrix<F>& A )
{
#ifndef RELEASE
CallStackEntry cse("apply_packed_reflectors::RUVF");
// 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> ALeft;
Matrix<F>
tT, t0,
tB, t1,
t2;
Matrix<F> SInv, Z;
LockedPartitionDownOffsetDiagonal
( offset,
H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 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 );
LockedView2x1( HPan, H01, H11 );
View( ALeft, A, 0, 0, A.Height(), HPan.Height() );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( UPPER, HPanCopy, 0, RIGHT );
SetDiagonal( HPanCopy, F(1), 0, RIGHT );
Herk( UPPER, ADJOINT, F(1), HPanCopy, SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, NORMAL, F(1), ALeft, HPanCopy, Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, F(1), SInv, Z );
Gemm( NORMAL, ADJOINT, F(-1), Z, HPanCopy, F(1), ALeft );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
}
}
void Trr2kNNNT
( UpperOrLower uplo,
Orientation orientationOfD,
T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
const DistMatrix<T>& C, const DistMatrix<T>& D,
T beta, DistMatrix<T>& E )
{
#ifndef RELEASE
PushCallStack("internal::Trr2kNNNT");
if( E.Height() != E.Width() || A.Width() != C.Width() ||
A.Height() != E.Height() || C.Height() != E.Height() ||
B.Width() != E.Width() || D.Height() != E.Width() ||
A.Width() != B.Height() || C.Width() != D.Width() )
throw std::logic_error("Nonconformal Trr2kNNNT");
#endif
const Grid& g = E.Grid();
DistMatrix<T> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T> BT(g), B0(g),
BB(g), B1(g),
B2(g);
DistMatrix<T> CL(g), CR(g),
C0(g), C1(g), C2(g);
DistMatrix<T> DL(g), DR(g),
D0(g), D1(g), D2(g);
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,MR, STAR> B1Trans_MR_STAR(g);
DistMatrix<T,MC, STAR> C1_MC_STAR(g);
DistMatrix<T,VR, STAR> D1_VR_STAR(g);
DistMatrix<T,STAR,MR > D1AdjOrTrans_STAR_MR(g);
A1_MC_STAR.AlignWith( E );
B1Trans_MR_STAR.AlignWith( E );
C1_MC_STAR.AlignWith( E );
D1_VR_STAR.AlignWith( E );
D1AdjOrTrans_STAR_MR.AlignWith( E );
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionDown
( B, BT,
BB, 0 );
LockedPartitionRight( C, CL, CR, 0 );
LockedPartitionRight( D, DL, DR, 0 );
while( AL.Width() < A.Width() )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
LockedRepartitionRight
( CL, /**/ CR,
C0, /**/ C1, C2 );
LockedRepartitionRight
( CL, /**/ CR,
C0, /**/ C1, C2 );
//--------------------------------------------------------------------//
A1_MC_STAR = A1;
C1_MC_STAR = C1;
B1Trans_MR_STAR.TransposeFrom( B1 );
D1_VR_STAR = D1;
if( orientationOfD == ADJOINT )
D1AdjOrTrans_STAR_MR.AdjointFrom( D1_VR_STAR );
else
D1AdjOrTrans_STAR_MR.TransposeFrom( D1_VR_STAR );
LocalTrr2k
( uplo, TRANSPOSE,
alpha, A1_MC_STAR, B1Trans_MR_STAR,
C1_MC_STAR, D1AdjOrTrans_STAR_MR,
beta, E );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( DL, /**/ DR,
D0, D1, /**/ D2 );
SlideLockedPartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
GemmTTC
( Orientation orientationOfA,
Orientation orientationOfB,
T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmTTC");
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
("GemmTTC 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 GemmTTC: \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() );
}
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<T> AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<T> BL(g), BR(g),
B0(g), B1(g), B2(g);
// Temporary distributions
DistMatrix<T,STAR,MC > A1_STAR_MC(g);
DistMatrix<T,VR, STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR > B1AdjOrTrans_STAR_MR(g);
A1_STAR_MC.AlignWith( C );
B1_VR_STAR.AlignWith( C );
B1AdjOrTrans_STAR_MR.AlignWith( C );
// Start the algorithm
Scale( beta, C );
LockedPartitionDown
( A, AT,
AB, 0 );
LockedPartitionRight( B, BL, BR, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
//--------------------------------------------------------------------//
A1_STAR_MC = A1;
B1_VR_STAR = B1;
if( orientationOfB == ADJOINT )
B1AdjOrTrans_STAR_MR.AdjointFrom( B1_VR_STAR );
else
B1AdjOrTrans_STAR_MR.TransposeFrom( B1_VR_STAR );
// C[MC,MR] += alpha (A1[*,MC])^[T/H] (B1[MR,*])^[T/H]
// = alpha (A1^[T/H])[MC,*] (B1^[T/H])[*,MR]
LocalGemm
( orientationOfA, NORMAL,
alpha, A1_STAR_MC, B1AdjOrTrans_STAR_MR, T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
RLHF
( Conjugation conjugation, int offset,
const DistMatrix<Complex<R> >& H,
const DistMatrix<Complex<R>,MD,STAR>& t,
DistMatrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("apply_packed_reflectors::RLHF");
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
("Width 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> ALeft(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 );
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 = H10.Width() + H11.Width();
const int HPanOffset =
std::min( H11.Height(), std::max(-offset-H00.Height(),0) );
const int HPanHeight = H11.Height()-HPanOffset;
LockedView
( HPan, H, H00.Height()+HPanOffset, 0, HPanHeight, HPanWidth );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanHeight );
View( ALeft, A, 0, 0, A.Height(), HPanWidth );
HPan_STAR_MR.AlignWith( ALeft );
ZAdj_STAR_MC.AlignWith( ALeft );
ZAdj_STAR_VC.AlignWith( ALeft );
Zeros( HPan.Height(), ALeft.Height(), ZAdj_STAR_MC );
Zeros( HPan.Height(), HPan.Height(), SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, LOWER, offset, HPanCopy );
SetDiagonal( RIGHT, offset, HPanCopy, C(1) );
HPan_STAR_VR = HPanCopy;
Herk
( UPPER, NORMAL,
C(1), HPan_STAR_VR.LockedMatrix(),
C(0), SInv_STAR_STAR.Matrix() );
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, ALeft, 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), ALeft );
//--------------------------------------------------------------------//
HPan_STAR_MR.FreeAlignments();
ZAdj_STAR_MC.FreeAlignments();
ZAdj_STAR_VC.FreeAlignments();
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LocalSymvColAccumulateU
( T alpha,
const DistMatrix<T>& A,
const DistMatrix<T,MC,STAR>& x_MC_STAR,
const DistMatrix<T,MR,STAR>& x_MR_STAR,
DistMatrix<T,MC,STAR>& z_MC_STAR,
DistMatrix<T,MR,STAR>& z_MR_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalSymvColAccumulateU");
if( A.Grid() != x_MC_STAR.Grid() ||
x_MC_STAR.Grid() != x_MR_STAR.Grid() ||
x_MR_STAR.Grid() != z_MC_STAR.Grid() ||
z_MC_STAR.Grid() != z_MR_STAR.Grid() )
throw std::logic_error
("{A,x,z} must be distributed over the same grid");
if( x_MC_STAR.Width() != 1 || x_MR_STAR.Width() != 1 ||
z_MC_STAR.Width() != 1 || z_MR_STAR.Width() != 1 )
throw std::logic_error("Expected x and z to be column vectors");
if( A.Height() != A.Width() ||
A.Height() != x_MC_STAR.Height() ||
A.Height() != x_MR_STAR.Height() ||
A.Height() != z_MC_STAR.Height() ||
A.Height() != z_MR_STAR.Height() )
{
std::ostringstream msg;
msg << "Nonconformal LocalSymvColAccumulateU: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " x[MC,* ] ~ " << x_MC_STAR.Height() << " x "
<< x_MC_STAR.Width() << "\n"
<< " x[MR,* ] ~ " << x_MR_STAR.Height() << " x "
<< x_MR_STAR.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() );
}
if( x_MC_STAR.ColAlignment() != A.ColAlignment() ||
x_MR_STAR.ColAlignment() != 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();
// Matrix views
DistMatrix<T> A11(g), A12(g);
DistMatrix<T> D11(g);
DistMatrix<T,MC,STAR> x1_MC_STAR(g);
DistMatrix<T,MR,STAR>
xT_MR_STAR(g), x0_MR_STAR(g),
xB_MR_STAR(g), x1_MR_STAR(g),
x2_MR_STAR(g);
DistMatrix<T,MC,STAR> z1_MC_STAR(g);
DistMatrix<T,MR,STAR> z1_MR_STAR(g),
z2_MR_STAR(g);
// We want our local gemvs to be of width blocksize, so we will
// temporarily change to max(r,c) times the current blocksize
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*LocalSymvBlocksize<T>() );
LockedPartitionDown
( x_MR_STAR, xT_MR_STAR,
xB_MR_STAR, 0 );
while( xT_MR_STAR.Height() < x_MR_STAR.Height() )
{
LockedRepartitionDown
( xT_MR_STAR, x0_MR_STAR,
/**********/ /**********/
x1_MR_STAR,
xB_MR_STAR, x2_MR_STAR );
const int n0 = x0_MR_STAR.Height();
const int n1 = x1_MR_STAR.Height();
const int n2 = x2_MR_STAR.Height();
LockedView( A11, A, n0, n0, n1, n1 );
LockedView( A12, A, n0, n0+n1, n1, n2 );
LockedView( x1_MC_STAR, x_MC_STAR, n0, 0, n1, 1 );
View( z1_MC_STAR, z_MC_STAR, n0, 0, n1, 1 );
View( z1_MR_STAR, z_MR_STAR, n0, 0, n1, 1 );
View( z2_MR_STAR, z_MR_STAR, n0+n1, 0, n2, 1 );
D11.AlignWith( A11 );
//--------------------------------------------------------------------//
// TODO: These diagonal block updates can be greatly improved
D11 = A11;
MakeTrapezoidal( LEFT, UPPER, 0, D11 );
Gemv
( NORMAL,
alpha, D11.LockedLocalMatrix(),
x1_MR_STAR.LockedLocalMatrix(),
T(1), z1_MC_STAR.LocalMatrix() );
MakeTrapezoidal( LEFT, UPPER, 1, D11 );
Gemv
( TRANSPOSE,
alpha, D11.LockedLocalMatrix(),
x1_MC_STAR.LockedLocalMatrix(),
T(1), z1_MR_STAR.LocalMatrix() );
Gemv
( NORMAL,
alpha, A12.LockedLocalMatrix(),
x2_MR_STAR.LockedLocalMatrix(),
T(1), z1_MC_STAR.LocalMatrix() );
Gemv
( TRANSPOSE,
alpha, A12.LockedLocalMatrix(),
x1_MC_STAR.LockedLocalMatrix(),
T(1), z2_MR_STAR.LocalMatrix() );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDown
( xT_MR_STAR, x0_MR_STAR,
x1_MR_STAR,
/**********/ /**********/
xB_MR_STAR, x2_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
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::GemmTNC
( 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::GemmTNC");
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("GemmTNC assumes A is (Conjugate)Transposed");
if( A.Width() != C.Height() ||
B.Width() != C.Width() ||
A.Height() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal GemmTNC: \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> AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<T,MC,MR> BT(g), B0(g),
BB(g), B1(g),
B2(g);
// Temporary distributions
DistMatrix<T,STAR,MC> A1_STAR_MC(g);
DistMatrix<T,STAR,MR> B1_STAR_MR(g);
// Start the algorithm
Scal( beta, C );
LockedPartitionDown
( A, AT,
AB, 0 );
LockedPartitionDown
( B, BT,
BB, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
A1_STAR_MC.AlignWith( C );
B1_STAR_MR.AlignWith( C );
//--------------------------------------------------------------------//
A1_STAR_MC = A1; // A1[*,MC] <- A1[MC,MR]
B1_STAR_MR = B1; // B1[*,MR] <- B1[MC,MR]
// C[MC,MR] += alpha (A1[*,MC])^T B1[*,MR]
// = alpha (A1^T)[MC,*] B1[*,MR]
internal::LocalGemm
( orientationOfA, NORMAL, alpha, A1_STAR_MC, B1_STAR_MR, (T)1, C );
//--------------------------------------------------------------------//
A1_STAR_MC.FreeAlignments();
B1_STAR_MR.FreeAlignments();
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
HerkLC( T alpha, const DistMatrix<T>& A, T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::HerkLC");
if( A.Grid() != C.Grid() )
throw std::logic_error
("A and C must be distributed over the same grid");
if( A.Width() != C.Height() || A.Width() != C.Width() )
{
std::ostringstream msg;
msg << "Nonconformal HerkLC:\n"
<< " A ~ " << A.Height() << " x " << A.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);
// Temporary distributions
DistMatrix<T,MR, STAR> A1Trans_MR_STAR(g);
DistMatrix<T,STAR,VR > A1_STAR_VR(g);
DistMatrix<T,STAR,MC > A1_STAR_MC(g);
A1Trans_MR_STAR.AlignWith( C );
A1_STAR_MC.AlignWith( C );
// Start the algorithm
ScaleTrapezoid( beta, LEFT, LOWER, 0, C );
LockedPartitionDown
( A, AT,
AB, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
//--------------------------------------------------------------------//
A1Trans_MR_STAR.TransposeFrom( A1 );
A1_STAR_VR.TransposeFrom( A1Trans_MR_STAR );
A1_STAR_MC = A1_STAR_VR;
LocalTrrk
( LOWER, ADJOINT, TRANSPOSE,
alpha, A1_STAR_MC, A1Trans_MR_STAR, T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
ReformHermitianMatrix
( UpperOrLower uplo,
DistMatrix<R,MC,MR>& A,
const DistMatrix<R,VR,STAR>& w,
const DistMatrix<R,MC,MR>& Z,
const RealFunctor& f )
{
#ifndef RELEASE
PushCallStack("hermitian_function::ReformHermitianMatrix");
#endif
const Grid& g = A.Grid();
DistMatrix<R,MC,MR> ZL(g), ZR(g),
Z0(g), Z1(g), Z2(g);
DistMatrix<R,VR,STAR> wT(g), w0(g),
wB(g), w1(g),
w2(g);
DistMatrix<R,MC, STAR> Z1_MC_STAR(g);
DistMatrix<R,VR, STAR> Z1_VR_STAR(g);
DistMatrix<R,STAR,MR > Z1Trans_STAR_MR(g);
DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);
if( uplo == LOWER )
MakeTrapezoidal( LEFT, UPPER, 1, A );
else
MakeTrapezoidal( LEFT, LOWER, -1, A );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
Z1_MC_STAR.AlignWith( A );
Z1_VR_STAR.AlignWith( A );
Z1Trans_STAR_MR.AlignWith( A );
//--------------------------------------------------------------------//
Z1_MC_STAR = Z1;
Z1_VR_STAR = Z1_MC_STAR;
w1_STAR_STAR = w1;
// Scale Z1[VR,* ] with the modified eigenvalues
const int width = Z1_VR_STAR.Width();
const int localHeight = Z1_VR_STAR.LocalHeight();
for( int j=0; j<width; ++j )
{
const R omega = f(w1_STAR_STAR.GetLocalEntry(j,0));
R* buffer = Z1_VR_STAR.LocalBuffer(0,j);
for( int iLocal=0; iLocal<localHeight; ++iLocal )
buffer[iLocal] *= omega;
}
Z1Trans_STAR_MR.TransposeFrom( Z1_VR_STAR );
internal::LocalTrrk( uplo, (R)1, Z1_MC_STAR, Z1Trans_STAR_MR, (R)1, A );
//--------------------------------------------------------------------//
Z1Trans_STAR_MR.FreeAlignments();
Z1_VR_STAR.FreeAlignments();
Z1_MC_STAR.FreeAlignments();
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
ReformNormalMatrix
( DistMatrix<Complex<R>,MC,MR >& A,
const DistMatrix<R, VR,STAR>& w,
const DistMatrix<Complex<R>,MC,MR >& Z,
const ComplexFunctor& f )
{
#ifndef RELEASE
PushCallStack("hermitian_function::ReformNormalMatrix");
#endif
const Grid& g = A.Grid();
typedef Complex<R> C;
DistMatrix<C,MC,MR> ZL(g), ZR(g),
Z0(g), Z1(g), Z2(g);
DistMatrix<R,VR,STAR> wT(g), w0(g),
wB(g), w1(g),
w2(g);
DistMatrix<C,MC, STAR> Z1_MC_STAR(g);
DistMatrix<C,VR, STAR> Z1_VR_STAR(g);
DistMatrix<C,STAR,MR > Z1Adj_STAR_MR(g);
DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);
Zero( A );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
Z1_MC_STAR.AlignWith( A );
Z1_VR_STAR.AlignWith( A );
Z1Adj_STAR_MR.AlignWith( A );
//--------------------------------------------------------------------//
Z1_MC_STAR = Z1;
Z1_VR_STAR = Z1_MC_STAR;
w1_STAR_STAR = w1;
// Scale Z1[VR,* ] with the modified eigenvalues
const int width = Z1_VR_STAR.Width();
const int localHeight = Z1_VR_STAR.LocalHeight();
for( int j=0; j<width; ++j )
{
const C conjOmega = Conj(f(w1_STAR_STAR.GetLocalEntry(j,0)));
C* buffer = Z1_VR_STAR.LocalBuffer(0,j);
for( int iLocal=0; iLocal<localHeight; ++iLocal )
buffer[iLocal] *= conjOmega;
}
Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
internal::LocalGemm
( NORMAL, NORMAL, (C)1, Z1_MC_STAR, Z1Adj_STAR_MR, (C)1, A );
//--------------------------------------------------------------------//
Z1Adj_STAR_MR.FreeAlignments();
Z1_VR_STAR.FreeAlignments();
Z1_MC_STAR.FreeAlignments();
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
HermitianFromEVD
( UpperOrLower uplo,
DistMatrix<F>& A,
const DistMatrix<BASE(F),VR,STAR>& w,
const DistMatrix<F>& Z )
{
#ifndef RELEASE
CallStackEntry entry("HermitianFromEVD");
#endif
const Grid& g = A.Grid();
typedef BASE(F) R;
DistMatrix<F> ZL(g), ZR(g),
Z0(g), Z1(g), Z2(g);
DistMatrix<R,VR,STAR> wT(g), w0(g),
wB(g), w1(g),
w2(g);
DistMatrix<F,MC, STAR> Z1_MC_STAR(g);
DistMatrix<F,VR, STAR> Z1_VR_STAR(g);
DistMatrix<F,STAR,MR > Z1Adj_STAR_MR(g);
DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);
A.ResizeTo( Z.Height(), Z.Height() );
if( uplo == LOWER )
MakeTrapezoidal( UPPER, A, 1 );
else
MakeTrapezoidal( LOWER, A, -1 );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
Z1_MC_STAR.AlignWith( A );
Z1_VR_STAR.AlignWith( A );
Z1Adj_STAR_MR.AlignWith( A );
//--------------------------------------------------------------------//
Z1_MC_STAR = Z1;
Z1_VR_STAR = Z1_MC_STAR;
w1_STAR_STAR = w1;
DiagonalScale( RIGHT, NORMAL, w1_STAR_STAR, Z1_VR_STAR );
Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
LocalTrrk( uplo, F(1), Z1_MC_STAR, Z1Adj_STAR_MR, F(1), A );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
}
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
}
inline void
Trr2kNNTN
( UpperOrLower uplo,
Orientation orientationOfC,
T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
const DistMatrix<T>& C, const DistMatrix<T>& D,
T beta, DistMatrix<T>& E )
{
#ifndef RELEASE
PushCallStack("internal::Trr2kNNTN");
if( E.Height() != E.Width() || A.Width() != C.Height() ||
A.Height() != E.Height() || C.Width() != E.Height() ||
B.Width() != E.Width() || D.Width() != E.Width() ||
A.Width() != B.Height() || C.Height() != D.Height() )
throw std::logic_error("Nonconformal Trr2kNNTN");
#endif
const Grid& g = E.Grid();
DistMatrix<T> AL(g), AR(g),
A0(g), A1(g), A2(g);
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> DT(g), D0(g),
DB(g), D1(g),
D2(g);
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,MR, STAR> B1Trans_MR_STAR(g);
DistMatrix<T,STAR,MC > C1_STAR_MC(g);
DistMatrix<T,MR, STAR> D1Trans_MR_STAR(g);
A1_MC_STAR.AlignWith( E );
B1Trans_MR_STAR.AlignWith( E );
C1_STAR_MC.AlignWith( E );
D1Trans_MR_STAR.AlignWith( E );
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionDown
( B, BT,
BB, 0 );
LockedPartitionDown
( C, CT,
CB, 0 );
LockedPartitionDown
( D, DT,
DB, 0 );
while( AL.Width() < A.Width() )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
LockedRepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
LockedRepartitionDown
( DT, D0,
/**/ /**/
D1,
DB, D2 );
//--------------------------------------------------------------------//
A1_MC_STAR = A1;
C1_STAR_MC = C1;
B1Trans_MR_STAR.TransposeFrom( B1 );
D1Trans_MR_STAR.TransposeFrom( D1 );
LocalTrr2k
( uplo, TRANSPOSE, orientationOfC, TRANSPOSE,
alpha, A1_MC_STAR, B1Trans_MR_STAR,
C1_STAR_MC, D1Trans_MR_STAR,
beta, E );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( DT, D0,
D1,
/**/ /**/
DB, D2 );
SlideLockedPartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
}
#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
}
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
RLHF
( Conjugation conjugation, int offset,
const Matrix<Complex<R> >& H,
const Matrix<Complex<R> >& t,
Matrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("apply_packed_reflectors::RLHF");
if( offset > 0 || offset < -H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Width() )
throw std::logic_error
("Width 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> ALeft;
Matrix<C>
tT, t0,
tB, t1,
t2;
Matrix<C> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 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 = H10.Width() + H11.Width();
const int HPanOffset =
std::min( H11.Height(), std::max(-offset-H00.Height(),0) );
const int HPanHeight = H11.Height()-HPanOffset;
LockedView
( HPan, H, H00.Height()+HPanOffset, 0, HPanHeight, HPanWidth );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2, HPanHeight );
View( ALeft, A, 0, 0, A.Height(), HPanWidth );
Zeros( ALeft.Height(), HPan.Height(), Z );
Zeros( HPan.Height(), HPan.Height(), SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, LOWER, offset, HPanCopy );
SetDiagonal( RIGHT, offset, HPanCopy, C(1) );
Herk( UPPER, NORMAL, C(1), HPanCopy, C(0), SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, ADJOINT, C(1), ALeft, HPanCopy, C(0), Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, C(1), SInv, Z );
Gemm( NORMAL, NORMAL, C(-1), Z, HPanCopy, C(1), ALeft );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
}
#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
RUVF
( Conjugation conjugation, Int offset,
const DistMatrix<F>& H, const DistMatrix<F,MD,STAR>& t, DistMatrix<F>& A )
{
#ifndef RELEASE
CallStackEntry cse("apply_packed_reflectors::RUVF");
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> ALeft(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 );
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 );
LockedView2x1( HPan, H01, H11 );
View( ALeft, A, 0, 0, A.Height(), HPan.Height() );
HPan_MR_STAR.AlignWith( ALeft );
ZAdj_STAR_MC.AlignWith( ALeft );
ZAdj_STAR_VC.AlignWith( ALeft );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( UPPER, HPanCopy, 0, RIGHT );
SetDiagonal( HPanCopy, F(1), 0, RIGHT );
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, ALeft, 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), ALeft );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
}
}
inline void
Trr2kTTTT
( UpperOrLower uplo,
Orientation orientationOfA, Orientation orientationOfB,
Orientation orientationOfC, Orientation orientationOfD,
T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
const DistMatrix<T>& C, const DistMatrix<T>& D,
T beta, DistMatrix<T>& E )
{
#ifndef RELEASE
PushCallStack("internal::Trr2kTTTT");
if( E.Height() != E.Width() || A.Height() != C.Height() ||
A.Width() != E.Height() || C.Width() != E.Height() ||
B.Height() != E.Width() || D.Height() != E.Width() ||
A.Height() != B.Width() || C.Height() != D.Width() )
throw std::logic_error("Nonconformal Trr2kTTTT");
#endif
const Grid& g = E.Grid();
DistMatrix<T> AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<T> BL(g), BR(g),
B0(g), B1(g), B2(g);
DistMatrix<T> CT(g), C0(g),
CB(g), C1(g),
C2(g);
DistMatrix<T> DL(g), DR(g),
D0(g), D1(g), D2(g);
DistMatrix<T,STAR,MC > A1_STAR_MC(g);
DistMatrix<T,VR, STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR > B1AdjOrTrans_STAR_MR(g);
DistMatrix<T,STAR,MC > C1_STAR_MC(g);
DistMatrix<T,VR, STAR> D1_VR_STAR(g);
DistMatrix<T,STAR,MR > D1AdjOrTrans_STAR_MR(g);
A1_STAR_MC.AlignWith( E );
B1_VR_STAR.AlignWith( E );
B1AdjOrTrans_STAR_MR.AlignWith( E );
C1_STAR_MC.AlignWith( E );
D1_VR_STAR.AlignWith( E );
D1AdjOrTrans_STAR_MR.AlignWith( E );
LockedPartitionDown
( A, AT,
AB, 0 );
LockedPartitionRight( B, BL, BR, 0 );
LockedPartitionDown
( C, CT,
CB, 0 );
LockedPartitionRight( D, DL, DR, 0 );
while( AT.Height() < A.Height() )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
LockedRepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
LockedRepartitionRight
( DL, /**/ DR,
D0, /**/ D1, D2 );
//--------------------------------------------------------------------//
A1_STAR_MC = A1;
C1_STAR_MC = C1;
B1_VR_STAR = B1;
D1_VR_STAR = D1;
if( orientationOfB == ADJOINT )
B1AdjOrTrans_STAR_MR.AdjointFrom( B1_VR_STAR );
else
B1AdjOrTrans_STAR_MR.TransposeFrom( B1_VR_STAR );
if( orientationOfD == ADJOINT )
D1AdjOrTrans_STAR_MR.AdjointFrom( D1_VR_STAR );
else
D1AdjOrTrans_STAR_MR.TransposeFrom( D1_VR_STAR );
LocalTrr2k
( uplo, orientationOfA, orientationOfC,
alpha, A1_STAR_MC, B1AdjOrTrans_STAR_MR,
C1_STAR_MC, D1AdjOrTrans_STAR_MR,
beta, E );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( DL, /**/ DR,
D0, D1, /**/ D2 );
SlideLockedPartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
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
}
void Trr2kNTTN
( UpperOrLower uplo,
Orientation orientationOfB, Orientation orientationOfC,
T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
const DistMatrix<T>& C, const DistMatrix<T>& D,
T beta, DistMatrix<T>& E )
{
#ifndef RELEASE
CallStackEntry entry("internal::Trr2kNTTN");
if( E.Height() != E.Width() || A.Width() != C.Height() ||
A.Height() != E.Height() || C.Width() != E.Height() ||
B.Height() != E.Width() || D.Width() != E.Width() ||
A.Width() != B.Width() || C.Height() != D.Height() )
LogicError("Nonconformal Trr2kNTTN");
#endif
const Grid& g = E.Grid();
DistMatrix<T> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T> BL(g), BR(g),
B0(g), B1(g), B2(g);
DistMatrix<T> CT(g), C0(g),
CB(g), C1(g),
C2(g);
DistMatrix<T> DT(g), D0(g),
DB(g), D1(g),
D2(g);
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,VR, STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR > B1AdjOrTrans_STAR_MR(g);
DistMatrix<T,STAR,MC > C1_STAR_MC(g);
DistMatrix<T,MR, STAR> D1Trans_MR_STAR(g);
A1_MC_STAR.AlignWith( E );
B1_VR_STAR.AlignWith( E );
B1AdjOrTrans_STAR_MR.AlignWith( E );
C1_STAR_MC.AlignWith( E );
D1Trans_MR_STAR.AlignWith( E );
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionRight( B, BL, BR, 0 );
LockedPartitionDown
( C, CT,
CB, 0 );
LockedPartitionDown
( D, DT,
DB, 0 );
while( AL.Width() < A.Width() )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
LockedRepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
LockedRepartitionDown
( DT, D0,
/**/ /**/
D1,
DB, D2 );
//--------------------------------------------------------------------//
A1_MC_STAR = A1;
C1_STAR_MC = C1;
B1_VR_STAR = B1;
if( orientationOfB == ADJOINT )
B1AdjOrTrans_STAR_MR.AdjointFrom( B1_VR_STAR );
else
B1AdjOrTrans_STAR_MR.TransposeFrom( B1_VR_STAR );
D1Trans_MR_STAR.TransposeFrom( D1 );
LocalTrr2k
( uplo, orientationOfC, TRANSPOSE,
alpha, A1_MC_STAR, B1AdjOrTrans_STAR_MR,
C1_STAR_MC, D1Trans_MR_STAR,
beta, E );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlideLockedPartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
SlideLockedPartitionDown
( DT, D0,
D1,
/**/ /**/
DB, D2 );
}
}
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
GemmNNC
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmNNC");
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 GemmNNC: \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> BT(g), B0(g),
BB(g), B1(g),
B2(g);
// Temporary distributions
DistMatrix<T,MC,STAR> A1_MC_STAR(g);
DistMatrix<T,MR,STAR> B1Trans_MR_STAR(g);
A1_MC_STAR.AlignWith( C );
B1Trans_MR_STAR.AlignWith( C );
// Start the algorithm
Scale( beta, C );
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionDown
( B, BT,
BB, 0 );
while( AR.Width() > 0 )
{
LockedRepartitionRight( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionDown( BT, B0,
/**/ /**/
B1,
BB, B2 );
//--------------------------------------------------------------------//
A1_MC_STAR = A1;
B1Trans_MR_STAR.TransposeFrom( B1 );
// C[MC,MR] += alpha A1[MC,*] (B1^T[MR,*])^T
// = alpha A1[MC,*] B1[*,MR]
LocalGemm
( NORMAL, TRANSPOSE, alpha, A1_MC_STAR, B1Trans_MR_STAR, T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionRight( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionDown( BT, B0,
B1,
/**/ /**/
BB, B2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
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
}
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
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
LocalSymmetricAccumulateRU
( Orientation orientation, T alpha,
const DistMatrix<T,MC, MR >& A,
const DistMatrix<T,STAR,MC >& B_STAR_MC,
const DistMatrix<T,MR, STAR>& BTrans_MR_STAR,
DistMatrix<T,MC, STAR>& ZTrans_MC_STAR,
DistMatrix<T,MR, STAR>& ZTrans_MR_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalSymmetricAccumulateRU");
if( A.Grid() != B_STAR_MC.Grid() ||
B_STAR_MC.Grid() != BTrans_MR_STAR.Grid() ||
BTrans_MR_STAR.Grid() != ZTrans_MC_STAR.Grid() ||
ZTrans_MC_STAR.Grid() != ZTrans_MR_STAR.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
if( A.Height() != A.Width() ||
A.Height() != B_STAR_MC.Width() ||
A.Height() != BTrans_MR_STAR.Height() ||
A.Height() != ZTrans_MC_STAR.Height() ||
A.Height() != ZTrans_MR_STAR.Height() ||
B_STAR_MC.Height() != BTrans_MR_STAR.Width() ||
BTrans_MR_STAR.Width() != ZTrans_MC_STAR.Width() ||
ZTrans_MC_STAR.Width() != ZTrans_MR_STAR.Width() )
{
std::ostringstream msg;
msg << "Nonconformal LocalSymmetricAccumulateRU: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B[* ,MC] ~ " << B_STAR_MC.Height() << " x "
<< B_STAR_MC.Width() << "\n"
<< " B^H/T[MR,* ] ~ " << BTrans_MR_STAR.Height() << " x "
<< BTrans_MR_STAR.Width() << "\n"
<< " Z^H/T[MC,* ] ~ " << ZTrans_MC_STAR.Height() << " x "
<< ZTrans_MC_STAR.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( B_STAR_MC.RowAlignment() != A.ColAlignment() ||
BTrans_MR_STAR.ColAlignment() != A.RowAlignment() ||
ZTrans_MC_STAR.ColAlignment() != A.ColAlignment() ||
ZTrans_MR_STAR.ColAlignment() != A.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = A.Grid();
// Matrix views
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,STAR,MC>
BL_STAR_MC(g), BR_STAR_MC(g),
B0_STAR_MC(g), B1_STAR_MC(g), B2_STAR_MC(g);
DistMatrix<T,MR,STAR>
BTTrans_MR_STAR(g), B0Trans_MR_STAR(g),
BBTrans_MR_STAR(g), B1Trans_MR_STAR(g),
B2Trans_MR_STAR(g);
DistMatrix<T,MC,STAR>
ZTTrans_MC_STAR(g), Z0Trans_MC_STAR(g),
ZBTrans_MC_STAR(g), Z1Trans_MC_STAR(g),
Z2Trans_MC_STAR(g);
DistMatrix<T,MR,STAR>
ZBTrans_MR_STAR(g), Z0Trans_MR_STAR(g),
ZTTrans_MR_STAR(g), Z1Trans_MR_STAR(g),
Z2Trans_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionRight( B_STAR_MC, BL_STAR_MC, BR_STAR_MC, 0 );
LockedPartitionDown
( BTrans_MR_STAR, BTTrans_MR_STAR,
BBTrans_MR_STAR, 0 );
PartitionDown
( ZTrans_MC_STAR, ZTTrans_MC_STAR,
ZBTrans_MC_STAR, 0 );
PartitionDown
( ZTrans_MR_STAR, ZTTrans_MR_STAR,
ZBTrans_MR_STAR, 0 );
while( ATL.Height() < A.Height() )
{
LockedRepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionRight
( BL_STAR_MC, /**/ BR_STAR_MC,
B0_STAR_MC, /**/ B1_STAR_MC, B2_STAR_MC );
LockedRepartitionDown
( BTTrans_MR_STAR, B0Trans_MR_STAR,
/***************/ /***************/
B1Trans_MR_STAR,
BBTrans_MR_STAR, B2Trans_MR_STAR );
RepartitionDown
( ZTTrans_MC_STAR, Z0Trans_MC_STAR,
/***************/ /***************/
Z1Trans_MC_STAR,
ZBTrans_MC_STAR, Z2Trans_MC_STAR );
RepartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
/***************/ /***************/
Z1Trans_MR_STAR,
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
D11.AlignWith( A11 );
//--------------------------------------------------------------------//
D11 = A11;
MakeTriangular( UPPER, D11 );
LocalGemm
( orientation, orientation,
alpha, D11, B1_STAR_MC, T(1), Z1Trans_MR_STAR );
SetDiagonal( D11, T(0) );
LocalGemm
( NORMAL, NORMAL, alpha, D11, B1Trans_MR_STAR, T(1), Z1Trans_MC_STAR );
LocalGemm
( orientation, orientation,
alpha, A12, B1_STAR_MC, T(1), Z2Trans_MR_STAR );
LocalGemm
( NORMAL, NORMAL, alpha, A12, B2Trans_MR_STAR, T(1), Z1Trans_MC_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionRight
( BL_STAR_MC, /**/ BR_STAR_MC,
B0_STAR_MC, B1_STAR_MC, /**/ B2_STAR_MC );
SlideLockedPartitionDown
( BTTrans_MR_STAR, B0Trans_MR_STAR,
B1Trans_MR_STAR,
/***************/ /***************/
BBTrans_MR_STAR, B2Trans_MR_STAR );
SlidePartitionDown
( ZTTrans_MC_STAR, Z0Trans_MC_STAR,
Z1Trans_MC_STAR,
/***************/ /***************/
ZBTrans_MC_STAR, Z2Trans_MC_STAR );
SlidePartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
Z1Trans_MR_STAR,
/***************/ /***************/
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}