void SUMMA_NTB
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
EL_DEBUG_CSE
const Int m = CPre.Height();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
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,MR,STAR> A1Trans_MR_STAR(g);
DistMatrix<T,STAR,MC> D1_STAR_MC(g);
DistMatrix<T,MR,MC> D1_MR_MC(g);
A1Trans_MR_STAR.AlignWith( B );
D1_STAR_MC.AlignWith( B );
for( Int k=0; k<m; k+=bsize )
{
const Int nb = Min(bsize,m-k);
auto A1 = A( IR(k,k+nb), ALL );
auto C1 = C( IR(k,k+nb), ALL );
// D1[*,MC] := alpha A1[*,MR] (B[MC,MR])^T
// = alpha (A1^T)[MR,*] (B^T)[MR,MC]
Transpose( A1, A1Trans_MR_STAR );
LocalGemm( TRANSPOSE, orientB, alpha, A1Trans_MR_STAR, B, D1_STAR_MC );
// C1[MC,MR] += scattered & transposed D1[*,MC] summed over grid rows
Contract( D1_STAR_MC, D1_MR_MC );
Axpy( T(1), D1_MR_MC, C1 );
}
}
void LT_C
( T alpha,
const AbstractDistMatrix<T>& APre,
AbstractDistMatrix<T>& CPre,
bool conjugate=false )
{
EL_DEBUG_CSE
const Int r = APre.Height();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );
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,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 );
for( Int k=0; k<r; k+=bsize )
{
const Int nb = Min(bsize,r-k);
auto A1 = A( IR(k,k+nb), ALL );
Transpose( A1, A1Trans_MR_STAR );
Transpose( A1Trans_MR_STAR, A1_STAR_VR );
A1_STAR_MC = A1_STAR_VR;
LocalTrrk
( LOWER, orientation, TRANSPOSE,
alpha, A1_STAR_MC, A1Trans_MR_STAR, T(1), C );
}
}
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
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
}
