void SUMMA_NTDot
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre,
Int blockSize=2000 )
{
EL_DEBUG_CSE
const Int m = CPre.Height();
const Int n = CPre.Width();
const Grid& g = APre.Grid();
DistMatrixReadProxy<T,T,STAR,VC> AProx( APre );
auto& A = AProx.GetLocked();
ElementalProxyCtrl BCtrl;
BCtrl.rowConstrain = true;
BCtrl.rowAlign = A.RowAlign();
DistMatrixReadProxy<T,T,STAR,VC> BProx( BPre, BCtrl );
auto& B = BProx.GetLocked();
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& C = CProx.Get();
DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);
for( Int kOuter=0; kOuter<m; kOuter+=blockSize )
{
const Int nbOuter = Min(blockSize,m-kOuter);
const Range<Int> indOuter( kOuter, kOuter+nbOuter );
auto A1 = A( indOuter, ALL );
for( Int kInner=0; kInner<n; kInner+=blockSize )
{
const Int nbInner = Min(blockSize,n-kInner);
const Range<Int> indInner( kInner, kInner+nbInner );
auto B1 = B( indInner, ALL );
auto C11 = C( indOuter, indInner );
LocalGemm( NORMAL, orientB, alpha, A1, B1, C11_STAR_STAR );
AxpyContract( T(1), C11_STAR_STAR, C11 );
}
}
}
inline void
GemmNNDot
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmNNDot");
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 GemmNNDot: \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();
if( A.Height() > B.Width() )
{
// Matrix views
DistMatrix<T> AT(g), AB(g),
A0(g), A1(g), A2(g);
DistMatrix<T> BL(g), B0(g),
BR(g), B1(g),
B2(g);
DistMatrix<T> CT(g), C0(g), C1L(g), C1R(g),
CB(g), C1(g), C10(g), C11(g), C12(g),
C2(g);
// Temporary distributions
DistMatrix<T,STAR,VC> A1_STAR_VC(g);
DistMatrix<T,VC,STAR> B1_VC_STAR(g);
DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);
// Star 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 );
A1_STAR_VC = A1;
B1_VC_STAR.AlignWith( A1_STAR_VC );
LockedPartitionRight( B, BL, BR, 0 );
PartitionRight( C1, C1L, C1R, 0 );
while( BR.Width() > 0 )
{
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
RepartitionRight
( C1L, /**/ C1R,
C10, /**/ C11, C12 );
Zeros( C11.Height(), C11.Width(), C11_STAR_STAR );
//------------------------------------------------------------//
B1_VC_STAR = B1;
LocalGemm
( NORMAL, NORMAL,
alpha, A1_STAR_VC, B1_VC_STAR, T(0), C11_STAR_STAR );
C11.SumScatterUpdate( T(1), C11_STAR_STAR );
//------------------------------------------------------------//
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( C1L, /**/ C1R,
C10, C11, /**/ C12 );
}
B1_VC_STAR.FreeAlignments();
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlidePartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
}
}
else
{
// Matrix views
DistMatrix<T> AT(g), AB(g),
A0(g), A1(g), A2(g);
DistMatrix<T> BL(g), B0(g),
BR(g), B1(g),
B2(g);
DistMatrix<T>
CL(g), CR(g), C1T(g), C01(g),
C0(g), C1(g), C2(g), C1B(g), C11(g),
C21(g);
// Temporary distributions
DistMatrix<T,STAR,VR> A1_STAR_VR(g);
DistMatrix<T,VR,STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);
// Star 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 );
B1_VR_STAR = B1;
A1_STAR_VR.AlignWith( B1_VR_STAR );
LockedPartitionDown
( A, AT,
AB, 0 );
PartitionDown
( C1, C1T,
C1B, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
RepartitionDown
( C1T, C01,
/***/ /***/
C11,
C1B, C21 );
Zeros( C11.Height(), C11.Width(), C11_STAR_STAR );
//------------------------------------------------------------//
A1_STAR_VR = A1;
LocalGemm
( NORMAL, NORMAL,
alpha, A1_STAR_VR, B1_VR_STAR, T(0), C11_STAR_STAR );
C11.SumScatterUpdate( T(1), C11_STAR_STAR );
//------------------------------------------------------------//
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlidePartitionDown
( C1T, C01,
C11,
/***/ /***/
C1B, C21 );
}
A1_STAR_VR.FreeAlignments();
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlidePartitionRight
( CL, /**/ CR,
C0, C1, /**/ C2 );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
