void UN_C
( T alpha,
const AbstractDistMatrix<T>& APre,
AbstractDistMatrix<T>& CPre,
bool conjugate=false )
{
EL_DEBUG_CSE
const Int r = APre.Width();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,VR, STAR> A1_VR_STAR(g);
DistMatrix<T,STAR,MR > A1Trans_STAR_MR(g);
A1_MC_STAR.AlignWith( C );
A1_VR_STAR.AlignWith( C );
A1Trans_STAR_MR.AlignWith( C );
for( Int k=0; k<r; k+=bsize )
{
const Int nb = Min(bsize,r-k);
auto A1 = A( ALL, IR(k,k+nb) );
A1_VR_STAR = A1_MC_STAR = A1;
Transpose( A1_VR_STAR, A1Trans_STAR_MR, conjugate );
LocalTrrk( UPPER, alpha, A1_MC_STAR, A1Trans_STAR_MR, T(1), C );
}
}
inline void
SyrkLN
( T alpha, const DistMatrix<T>& A, T beta, DistMatrix<T>& C,
bool conjugate=false )
{
#ifndef RELEASE
PushCallStack("internal::SyrkLN");
if( A.Grid() != C.Grid() )
throw std::logic_error
("A and C must be distributed over the same grid");
if( A.Height() != C.Height() || A.Height() != C.Width() )
{
std::ostringstream msg;
msg << "Nonconformal SyrkLN:\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> AL(g), AR(g),
A0(g), A1(g), A2(g);
// Temporary distributions
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,VR, STAR> A1_VR_STAR(g);
DistMatrix<T,STAR,MR > A1Trans_STAR_MR(g);
A1_MC_STAR.AlignWith( C );
A1_VR_STAR.AlignWith( C );
A1Trans_STAR_MR.AlignWith( C );
// Start the algorithm
ScaleTrapezoid( beta, LEFT, LOWER, 0, C );
LockedPartitionRight( A, AL, AR, 0 );
while( AR.Width() > 0 )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
//--------------------------------------------------------------------//
A1_VR_STAR = A1_MC_STAR = A1;
A1Trans_STAR_MR.TransposeFrom( A1_VR_STAR, conjugate );
LocalTrrk( LOWER, alpha, A1_MC_STAR, A1Trans_STAR_MR, T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void SUMMA_NTC
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
EL_DEBUG_CSE
const Int sumDim = APre.Width();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
const bool conjugate = ( orientB == 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,VR,STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR> B1Trans_STAR_MR(g);
A1_MC_STAR.AlignWith( C );
B1_VR_STAR.AlignWith( C );
B1Trans_STAR_MR.AlignWith( C );
for( Int k=0; k<sumDim; k+=bsize )
{
const Int nb = Min(bsize,sumDim-k);
auto A1 = A( ALL, IR(k,k+nb) );
auto B1 = B( ALL, IR(k,k+nb) );
A1_MC_STAR = A1;
B1_VR_STAR = B1;
Transpose( B1_VR_STAR, B1Trans_STAR_MR, conjugate );
// C[MC,MR] += alpha A1[MC,*] (B1[MR,*])^T
LocalGemm
( NORMAL, NORMAL, alpha, A1_MC_STAR, B1Trans_STAR_MR, T(1), C );
}
}
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
Her2kLN
( 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::Her2kLN");
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() || A.Height() != C.Width() ||
B.Height() != C.Height() || B.Height() != C.Width() ||
A.Width() != B.Width() )
{
std::ostringstream msg;
msg << "Nonconformal Her2kLN:\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,MC,MR> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T,MC,MR> BL(g), BR(g),
B0(g), B1(g), B2(g);
// Temporary distributions
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,MC, STAR> B1_MC_STAR(g);
DistMatrix<T,VR, STAR> A1_VR_STAR(g);
DistMatrix<T,VR, STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR > A1Adj_STAR_MR(g);
DistMatrix<T,STAR,MR > B1Adj_STAR_MR(g);
A1_MC_STAR.AlignWith( C );
B1_MC_STAR.AlignWith( C );
A1_VR_STAR.AlignWith( C );
B1_VR_STAR.AlignWith( C );
A1Adj_STAR_MR.AlignWith( C );
B1Adj_STAR_MR.AlignWith( C );
// Start the algorithm
ScaleTrapezoid( beta, LEFT, LOWER, 0, C );
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionRight( B, BL, BR, 0 );
while( AR.Width() > 0 )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
//--------------------------------------------------------------------//
A1_VR_STAR = A1_MC_STAR = A1;
A1Adj_STAR_MR.AdjointFrom( A1_VR_STAR );
B1_VR_STAR = B1_MC_STAR = B1;
B1Adj_STAR_MR.AdjointFrom( B1_VR_STAR );
LocalTrr2k
( LOWER,
alpha, A1_MC_STAR, B1Adj_STAR_MR,
B1_MC_STAR, A1Adj_STAR_MR,
T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
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
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
}
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::Trr2kNTTT
( UpperOrLower uplo,
Orientation orientationOfB,
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::Trr2kNTTT");
if( E.Height() != E.Width() || A.Width() != C.Height() ||
A.Height() != E.Height() || C.Width() != E.Height() ||
B.Height() != E.Width() || D.Height() != E.Width() ||
A.Width() != B.Width() || C.Height() != D.Width() )
throw std::logic_error("Nonconformal Trr2kNTTT");
#endif
const Grid& g = E.Grid();
DistMatrix<T,MC,MR> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T,MC,MR> BL(g), BR(g),
B0(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,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);
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionRight( B, BL, BR, 0 );
LockedPartitionDown
( C, CT,
CB, 0 );
LockedPartitionRight( D, DL, DR, 0 );
while( AL.Width() < A.Width() )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
LockedRepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
LockedRepartitionRight
( DL, /**/ DR,
D0, /**/ D1, D2 );
A1_MC_STAR.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 );
//--------------------------------------------------------------------//
A1_MC_STAR = 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 );
internal::LocalTrr2k
( uplo, orientationOfC,
alpha, A1_MC_STAR, B1AdjOrTrans_STAR_MR,
C1_STAR_MC, D1AdjOrTrans_STAR_MR,
beta, E );
//--------------------------------------------------------------------//
D1AdjOrTrans_STAR_MR.FreeAlignments();
D1_VR_STAR.FreeAlignments();
C1_STAR_MC.FreeAlignments();
B1AdjOrTrans_STAR_MR.FreeAlignments();
B1_VR_STAR.FreeAlignments();
A1_MC_STAR.FreeAlignments();
SlideLockedPartitionRight
( DL, /**/ DR,
D0, D1, /**/ D2 );
SlideLockedPartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
Syr2kUN
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
T beta, DistMatrix<T>& C,
bool conjugate=false )
{
#ifndef RELEASE
CallStackEntry entry("internal::Syr2kUN");
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() || A.Height() != C.Width() ||
B.Height() != C.Height() || B.Height() != C.Width() ||
A.Width() != B.Width() )
{
std::ostringstream msg;
msg << "Nonconformal Syr2kUN:\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 = C.Grid();
// Matrix views
DistMatrix<T> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T> BL(g), BR(g),
B0(g), B1(g), B2(g);
// Temporary distributions
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,MC, STAR> B1_MC_STAR(g);
DistMatrix<T,VR, STAR> A1_VR_STAR(g);
DistMatrix<T,VR, STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR > A1Trans_STAR_MR(g);
DistMatrix<T,STAR,MR > B1Trans_STAR_MR(g);
A1_MC_STAR.AlignWith( C );
B1_MC_STAR.AlignWith( C );
A1_VR_STAR.AlignWith( C );
B1_VR_STAR.AlignWith( C );
A1Trans_STAR_MR.AlignWith( C );
B1Trans_STAR_MR.AlignWith( C );
// Start the algorithm
ScaleTrapezoid( beta, LEFT, UPPER, 0, C );
LockedPartitionRight( A, AL, AR, 0 );
LockedPartitionRight( B, BL, BR, 0 );
while( AR.Width() > 0 )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
//--------------------------------------------------------------------//
A1_VR_STAR = A1_MC_STAR = A1;
A1Trans_STAR_MR.TransposeFrom( A1_VR_STAR, conjugate );
B1_VR_STAR = B1_MC_STAR = B1;
B1Trans_STAR_MR.TransposeFrom( B1_VR_STAR, conjugate );
LocalTrr2k
( UPPER,
alpha, A1_MC_STAR, B1Trans_STAR_MR,
B1_MC_STAR, A1Trans_STAR_MR,
T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
}
}
inline void
internal::GemmTNB
( 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::GemmTNB");
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("GemmTNB assumes A is (Conjugate)Transposed");
if( A.Width() != C.Height() ||
B.Width() != C.Width() ||
A.Height() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal GemmTNB: \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,MC,MR> AL(g), AR(g),
A0(g), A1(g), A2(g);
DistMatrix<T,MC,MR> CT(g), C0(g),
CB(g), C1(g),
C2(g);
// Temporary distributions
DistMatrix<T,MC,STAR> A1_MC_STAR(g);
DistMatrix<T,STAR,MR> D1_STAR_MR(g);
// Start the algorithm
Scal( beta, C );
LockedPartitionRight( A, AL, AR, 0 );
PartitionDown
( C, CT,
CB, 0 );
while( AR.Width() > 0 )
{
LockedRepartitionRight
( AL, /**/ AR,
A0, /**/ A1, A2 );
RepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
A1_MC_STAR.AlignWith( B );
D1_STAR_MR.AlignWith( B );
D1_STAR_MR.ResizeTo( C1.Height(), C1.Width() );
//--------------------------------------------------------------------//
A1_MC_STAR = A1; // A1[MC,*] <- A1[MC,MR]
// D1[*,MR] := alpha (A1[MC,*])^T B[MC,MR]
// = alpha (A1^T)[*,MC] B[MC,MR]
internal::LocalGemm
( orientationOfA, NORMAL, alpha, A1_MC_STAR, B, (T)0, D1_STAR_MR );
// C1[MC,MR] += scattered result of D1[*,MR] summed over grid cols
C1.SumScatterUpdate( (T)1, D1_STAR_MR );
//--------------------------------------------------------------------//
A1_MC_STAR.FreeAlignments();
D1_STAR_MR.FreeAlignments();
SlideLockedPartitionRight
( AL, /**/ AR,
A0, A1, /**/ A2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}