void SUMMA_NTA
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
EL_DEBUG_CSE
const Int n = CPre.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,MR,STAR> B1Trans_MR_STAR(g);
DistMatrix<T,MC,STAR> D1_MC_STAR(g);
B1Trans_MR_STAR.AlignWith( A );
D1_MC_STAR.AlignWith( A );
for( Int k=0; k<n; k+=bsize )
{
const Int nb = Min(bsize,n-k);
auto B1 = B( IR(k,k+nb), ALL );
auto C1 = C( ALL, IR(k,k+nb) );
// C1[MC,*] := alpha A[MC,MR] (B1^[T/H])[MR,*]
Transpose( B1, B1Trans_MR_STAR, conjugate );
LocalGemm( NORMAL, NORMAL, alpha, A, B1Trans_MR_STAR, D1_MC_STAR );
// C1[MC,MR] += scattered result of D1[MC,*] summed over grid rows
AxpyContract( T(1), D1_MC_STAR, C1 );
}
}
inline void
GemmNNA
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmNNA");
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 GemmNNA: \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> BL(g), BR(g),
B0(g), B1(g), B2(g);
DistMatrix<T> CL(g), CR(g),
C0(g), C1(g), C2(g);
// Temporary distributions
DistMatrix<T,VR,STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR> B1Trans_STAR_MR(g);
DistMatrix<T,MC,STAR> D1_MC_STAR(g);
B1_VR_STAR.AlignWith( A );
B1Trans_STAR_MR.AlignWith( A );
D1_MC_STAR.AlignWith( A );
// Start 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 );
Zeros( C1.Height(), C1.Width(), D1_MC_STAR );
//--------------------------------------------------------------------//
B1_VR_STAR = B1;
B1Trans_STAR_MR.TransposeFrom( B1_VR_STAR );
// D1[MC,*] := alpha A[MC,MR] B1[MR,*]
LocalGemm
( NORMAL, TRANSPOSE, alpha, A, B1Trans_STAR_MR, T(0), D1_MC_STAR );
// C1[MC,MR] += scattered result of D1[MC,*] summed over grid rows
C1.SumScatterUpdate( T(1), D1_MC_STAR );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmRLNCOld
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& L,
DistMatrix<T>& X )
{
#ifndef RELEASE
PushCallStack("internal::TrmmRLNCOld");
if( L.Grid() != X.Grid() )
throw std::logic_error
("L and X must be distributed over the same grid");
if( L.Height() != L.Width() || X.Width() != L.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmRLNC: \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> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<T,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<T,MR, STAR> L21_MR_STAR(g);
DistMatrix<T,VC, STAR> X1_VC_STAR(g);
DistMatrix<T,MC, STAR> D1_MC_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionRight( X, XL, XR, 0 );
while( XR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionRight
( XL, /**/ XR,
X0, /**/ X1, X2 );
L21_MR_STAR.AlignWith( X2 );
D1_MC_STAR.AlignWith( X1 );
Zeros( X1.Height(), X1.Width(), D1_MC_STAR );
//--------------------------------------------------------------------//
X1_VC_STAR = X1;
L11_STAR_STAR = L11;
LocalTrmm
( RIGHT, LOWER, NORMAL, diag, T(1), L11_STAR_STAR, X1_VC_STAR );
X1 = X1_VC_STAR;
L21_MR_STAR = L21;
LocalGemm( NORMAL, NORMAL, T(1), X2, L21_MR_STAR, T(0), D1_MC_STAR );
X1.SumScatterUpdate( T(1), D1_MC_STAR );
//--------------------------------------------------------------------//
L21_MR_STAR.FreeAlignments();
D1_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionRight
( XL, /**/ XR,
X0, X1, /**/ X2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrmmRUNCOld
( UnitOrNonUnit diag,
T alpha, const DistMatrix<T>& U,
DistMatrix<T>& X )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrmmRUNCOld");
if( U.Grid() != X.Grid() )
throw std::logic_error
("U and X must be distributed over the same grid");
if( U.Height() != U.Width() || X.Width() != U.Height() )
{
std::ostringstream msg;
msg << "Nonconformal TrmmRUNC: \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>
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> XL(g), XR(g),
X0(g), X1(g), X2(g);
// Temporary distributions
DistMatrix<T,MR, STAR> U01_MR_STAR(g);
DistMatrix<T,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<T,VC, STAR> X1_VC_STAR(g);
DistMatrix<T,MC, STAR> D1_MC_STAR(g);
// Start the algorithm
Scale( alpha, X );
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionLeft( X, XL, XR, 0 );
while( XL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionLeft
( XL, /**/ XR,
X0, X1, /**/ X2 );
U01_MR_STAR.AlignWith( X0 );
D1_MC_STAR.AlignWith( X1 );
//--------------------------------------------------------------------//
X1_VC_STAR = X1;
U11_STAR_STAR = U11;
LocalTrmm
( RIGHT, UPPER, NORMAL, diag, T(1), U11_STAR_STAR, X1_VC_STAR );
X1 = X1_VC_STAR;
U01_MR_STAR = U01;
LocalGemm( NORMAL, NORMAL, T(1), X0, U01_MR_STAR, D1_MC_STAR );
X1.SumScatterUpdate( T(1), D1_MC_STAR );
//--------------------------------------------------------------------//
U01_MR_STAR.FreeAlignments();
D1_MC_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionLeft
( XL, /**/ XR,
X0, /**/ X1, X2 );
}
}
