void SUMMA_TNB
( Orientation orientA,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
DEBUG_CSE
const Int m = CPre.Height();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
const bool conjugate = ( orientA == ADJOINT );
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MC,STAR> A1_MC_STAR(g);
DistMatrix<T,MR,STAR> D1Trans_MR_STAR(g);
A1_MC_STAR.AlignWith( B );
D1Trans_MR_STAR.AlignWith( B );
for( Int k=0; k<m; k+=bsize )
{
const Int nb = Min(bsize,m-k);
auto A1 = A( ALL, IR(k,k+nb) );
auto C1 = C( IR(k,k+nb), ALL );
// D1[*,MR] := alpha (A1[MC,*])^[T/H] B[MC,MR]
// = alpha (A1^[T/H])[*,MC] B[MC,MR]
A1_MC_STAR = A1; // A1[MC,*] <- A1[MC,MR]
LocalGemm( orientA, NORMAL, T(1), B, A1_MC_STAR, D1Trans_MR_STAR );
TransposeAxpyContract( alpha, D1Trans_MR_STAR, C1, conjugate );
}
}
inline void
TrmmLLNCOld
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmLLNCOld");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( L.Height() != L.Width() || L.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLLNC: \n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,STAR,MC > L10_STAR_MC(g);
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> D1Trans_MR_STAR(g);
DistMatrix<T,MR, MC > D1Trans_MR_MC(g);
DistMatrix<T,MC, MR > D1(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionUp
( X, XT,
XB, 0 );
while( XT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
RepartitionUp
( XT, X0,
X1,
/**/ /**/
XB, X2 );
L10_STAR_MC.AlignWith( X0 );
D1Trans_MR_STAR.AlignWith( X1 );
D1Trans_MR_MC.AlignWith( X1 );
D1.AlignWith( X1 );
//--------------------------------------------------------------------//
L11_STAR_STAR = L11;
X1_STAR_VR = X1;
LocalTrmm( LEFT, LOWER, NORMAL, diag, T(1), L11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
L10_STAR_MC = L10;
LocalGemm
( TRANSPOSE, TRANSPOSE, T(1), X0, L10_STAR_MC, D1Trans_MR_STAR );
D1Trans_MR_MC.SumScatterFrom( D1Trans_MR_STAR );
Zeros( D1, X1.Height(), X1.Width() );
Transpose( D1Trans_MR_MC.Matrix(), D1.Matrix() );
Axpy( T(1), D1, X1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
D1Trans_MR_MC.FreeAlignments();
D1Trans_MR_STAR.FreeAlignments();
L10_STAR_MC.FreeAlignments();
SlideLockedPartitionUpDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
SlidePartitionUp
( XT, X0,
/**/ /**/
X1,
XB, X2 );
}
}
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
}
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
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
internal::TrmmLUNC
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T,MC,MR>& U,
DistMatrix<T,MC,MR>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmLUNC");
if( U.Grid() != X.Grid() )
throw std::logic_error
("U and X must be distributed over the same grid");
if( U.Height() != U.Width() || U.Width() != X.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmLUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X ~ " << X.Height() << " x " << X.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T,MC,MR>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<T,MC,MR> XT(g), X0(g),
XB(g), X1(g),
X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<T,STAR,MC > U12_STAR_MC(g);
DistMatrix<T,STAR,VR > X1_STAR_VR(g);
DistMatrix<T,MR, STAR> D1Trans_MR_STAR(g);
DistMatrix<T,MR, MC > D1Trans_MR_MC(g);
DistMatrix<T,MC, MR > D1(g);
// Start the algorithm
Scal( alpha, X );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionDown
( X, XT,
XB, 0 );
while( XB.Height() > 0 )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
RepartitionDown
( XT, X0,
/**/ /**/
X1,
XB, X2 );
U12_STAR_MC.AlignWith( X2 );
D1Trans_MR_STAR.AlignWith( X1 );
D1Trans_MR_MC.AlignWith( X1 );
D1.AlignWith( X1 );
D1Trans_MR_STAR.ResizeTo( X1.Width(), X1.Height() );
D1.ResizeTo( X1.Height(), X1.Width() );
//--------------------------------------------------------------------//
X1_STAR_VR = X1;
U11_STAR_STAR = U11;
internal::LocalTrmm
( LEFT, UPPER, NORMAL, diag, (T)1, U11_STAR_STAR, X1_STAR_VR );
X1 = X1_STAR_VR;
U12_STAR_MC = U12;
internal::LocalGemm
( TRANSPOSE, TRANSPOSE, (T)1, X2, U12_STAR_MC, (T)0, D1Trans_MR_STAR );
D1Trans_MR_MC.SumScatterFrom( D1Trans_MR_STAR );
Transpose( D1Trans_MR_MC.LocalMatrix(), D1.LocalMatrix() );
Axpy( (T)1, D1, X1 );
//--------------------------------------------------------------------//
D1.FreeAlignments();
D1Trans_MR_MC.FreeAlignments();
D1Trans_MR_STAR.FreeAlignments();
U12_STAR_MC.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlidePartitionDown
( XT, X0,
X1,
/**/ /**/
XB, X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
