inline void
ReformHermitianMatrix
( UpperOrLower uplo,
DistMatrix<Complex<R>,MC,MR >& A,
const DistMatrix<R, VR,STAR>& w,
const DistMatrix<Complex<R>,MC,MR >& Z,
const RealFunctor& f )
{
#ifndef RELEASE
PushCallStack("hermitian_function::ReformHermitianMatrix");
#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);
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 );
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 R omega = f(w1_STAR_STAR.GetLocalEntry(j,0));
C* buffer = Z1_VR_STAR.LocalBuffer(0,j);
for( int iLocal=0; iLocal<localHeight; ++iLocal )
buffer[iLocal] *= omega;
}
Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
internal::LocalTrrk( uplo, (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
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
internal::ApplyPackedReflectorsLUVF
( 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::ApplyPackedReflectorsLUVF");
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 > H.Height() )
throw std::logic_error("Transforms cannot extend above matrix");
if( offset < 0 )
throw std::logic_error("Transforms cannot extend below matrix");
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");
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),
HBL(g), HBR(g), H10(g), H11(g), H12(g),
H20(g), H21(g), H22(g);
DistMatrix<C,MC,MR>
AT(g), A0(g), ATop(g),
AB(g), A1(g),
A2(g);
DistMatrix<C,MD,STAR>
tT(g), t0(g),
tB(g), t1(g),
t2(g);
DistMatrix<C,MC, MR > HPanCopy(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 );
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 );
HPan_MC_STAR.AlignWith( ATop );
Z_STAR_MR.AlignWith( ATop );
Z_STAR_VR.AlignWith( ATop );
Z_STAR_MR.ResizeTo( HPan.Width(), ATop.Width() );
SInv_STAR_STAR.ResizeTo( HPan.Width(), HPan.Width() );
Zero( SInv_STAR_STAR );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, UPPER, offset, HPanCopy );
SetDiagonalToOne( RIGHT, 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, ATop, (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, ATop );
//--------------------------------------------------------------------//
HPan_MC_STAR.FreeAlignments();
Z_STAR_MR.FreeAlignments();
Z_STAR_VR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#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
Trr2kNNTT
( UpperOrLower uplo,
Orientation orientationOfC, Orientation orientationOfD,
T alpha, const DistMatrix<T,MC,MR>& A, const DistMatrix<T,MC,MR>& B,
const DistMatrix<T,MC,MR>& C, const DistMatrix<T,MC,MR>& D,
T beta, DistMatrix<T,MC,MR>& E )
{
#ifndef RELEASE
PushCallStack("internal::Trr2kNNTT");
if( E.Height() != E.Width() || A.Width() != C.Height() ||
A.Height() != E.Height() || C.Width() != E.Height() ||
B.Width() != E.Width() || D.Height() != E.Width() ||
A.Width() != B.Height() || C.Height() != D.Width() )
throw std::logic_error("Nonconformal Trr2kNNTT");
#endif
const Grid& g = E.Grid();
DistMatrix<T,MC,MR> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T,MC,MR> BT(g), B0(g),
BB(g), B1(g),
B2(g);
DistMatrix<T,MC,MR> CT(g), C0(g),
CB(g), C1(g),
C2(g);
DistMatrix<T,MC,MR> 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,STAR,MC > C1_STAR_MC(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_STAR_MC.AlignWith( E );
D1_VR_STAR.AlignWith( E );
D1AdjOrTrans_STAR_MR.AlignWith( E );
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionDown
( B, BT,
BB, 0 );
LockedPartitionDown
( C, CT,
CB, 0 );
LockedPartitionRight( D, DL, DR, 0 );
while( AL.Width() < A.Width() )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
LockedRepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
LockedRepartitionRight
( DL, /**/ DR,
D0, /**/ D1, D2 );
//--------------------------------------------------------------------//
A1_MC_STAR = A1;
C1_STAR_MC = 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, orientationOfC,
alpha, A1_MC_STAR, B1Trans_MR_STAR,
C1_STAR_MC, D1AdjOrTrans_STAR_MR,
beta, E );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( DL, /**/ DR,
D0, D1, /**/ 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
Her2kUC
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::Her2kUC");
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 Her2kUC:\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> 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, ADJOINT, TRANSPOSE, ADJOINT, 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 );
}
#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,
bool conjugate=false )
{
#ifndef RELEASE
CallStackEntry entry("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() )
LogicError
("{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 )
LogicError("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";
LogicError( 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() )
LogicError("Partial matrix distributions are misaligned");
#endif
const Grid& g = A.Grid();
const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );
// 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 = 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;
MakeTriangular( UPPER, D11 );
LocalGemv( NORMAL, alpha, D11, x1_MR_STAR, T(1), z1_MC_STAR );
SetDiagonal( D11, T(0) );
LocalGemv( orientation, alpha, D11, x1_MC_STAR, T(1), z1_MR_STAR );
LocalGemv( NORMAL, alpha, A12, x2_MR_STAR, T(1), z1_MC_STAR );
LocalGemv( orientation, alpha, A12, x1_MC_STAR, T(1), z2_MR_STAR );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( xT_MR_STAR, x0_MR_STAR,
x1_MR_STAR,
/**********/ /**********/
xB_MR_STAR, x2_MR_STAR );
}
PopBlocksizeStack();
}
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
}
inline void
SymmRUA
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
T beta, DistMatrix<T>& C,
bool conjugate=false )
{
#ifndef RELEASE
PushCallStack("internal::SymmRUA");
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();
const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );
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> B1Trans_MR_STAR(g);
DistMatrix<T,VC, STAR> B1Trans_VC_STAR(g);
DistMatrix<T,STAR,MC > B1_STAR_MC(g);
DistMatrix<T,MC, STAR> Z1Trans_MC_STAR(g);
DistMatrix<T,MR, STAR> Z1Trans_MR_STAR(g);
DistMatrix<T,MC, MR > Z1Trans(g);
DistMatrix<T,MR, MC > Z1Trans_MR_MC(g);
B1Trans_MR_STAR.AlignWith( A );
B1Trans_VC_STAR.AlignWith( A );
B1_STAR_MC.AlignWith( A );
Z1Trans_MC_STAR.AlignWith( A );
Z1Trans_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 );
Z1Trans_MR_MC.AlignWith( C1 );
Zeros( C1.Width(), C1.Height(), Z1Trans_MC_STAR );
Zeros( C1.Width(), C1.Height(), Z1Trans_MR_STAR );
//--------------------------------------------------------------------//
B1Trans_MR_STAR.TransposeFrom( B1, conjugate );
B1Trans_VC_STAR = B1Trans_MR_STAR;
B1_STAR_MC.TransposeFrom( B1Trans_VC_STAR, conjugate );
LocalSymmetricAccumulateRU
( orientation, alpha, A, B1_STAR_MC, B1Trans_MR_STAR,
Z1Trans_MC_STAR, Z1Trans_MR_STAR );
Z1Trans.SumScatterFrom( Z1Trans_MC_STAR );
Z1Trans_MR_MC = Z1Trans;
Z1Trans_MR_MC.SumScatterUpdate( T(1), Z1Trans_MR_STAR );
Transpose( Z1Trans_MR_MC.LockedMatrix(), Z1Local, conjugate );
Axpy( T(1), Z1Local, C1.Matrix() );
//--------------------------------------------------------------------//
Z1Trans_MR_MC.FreeAlignments();
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}