void L1DistanceMatrix(direction_t dirA, direction_t dirB, T alpha,
const El::ElementalMatrix<T> &APre, const El::ElementalMatrix<T> &BPre,
T beta, El::ElementalMatrix<T> &CPre) {
if (dirA == base::COLUMNS && dirB == base::COLUMNS) {
// Use a SUMMA-like routine, with C as stationary
// Basically an adaptation of Elementals TN case for stationary C.
const El::Int m = CPre.Height();
const El::Int n = CPre.Width();
const El::Int sumDim = BPre.Height();
const El::Int bsize = El::Blocksize();
const El::Grid& g = APre.Grid();
El::DistMatrixReadProxy<T, T, El::MC, El::MR> AProx(APre);
El::DistMatrixReadProxy<T, T, El::MC, El::MR> BProx(BPre);
El::DistMatrixReadWriteProxy<T, T, El::MC, El::MR> CProx(CPre);
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
El::DistMatrix<T, El::STAR, El::MC> A1_STAR_MC(g);
El::DistMatrix<T, El::STAR, El::MR> B1_STAR_MR(g);
A1_STAR_MC.AlignWith(C);
B1_STAR_MR.AlignWith(C);
El::Scale(beta, C);
for(El::Int k = 0; k < sumDim; k += bsize) {
const El::Int nb = std::min(bsize,sumDim-k);
auto A1 = A(El::IR(k,k+nb), El::IR(0,m));
auto B1 = B(El::IR(k,k+nb), El::IR(0,n));
A1_STAR_MC = A1;
B1_STAR_MR = B1;
L1DistanceMatrix(base::COLUMNS, base::COLUMNS, alpha,
A1_STAR_MC.LockedMatrix(), B1_STAR_MR.LockedMatrix(),
T(1.0), C.Matrix());
}
}
// TODO the rest of the cases.
}
void SUMMA_TNC
( Orientation orientA,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
DEBUG_CSE
const Int sumDim = BPre.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,STAR,MC> A1_STAR_MC(g);
DistMatrix<T,MR,STAR> B1Trans_MR_STAR(g);
A1_STAR_MC.AlignWith( C );
B1Trans_MR_STAR.AlignWith( C );
for( Int k=0; k<sumDim; k+=bsize )
{
const Int nb = Min(bsize,sumDim-k);
auto A1 = A( IR(k,k+nb), ALL );
auto B1 = B( IR(k,k+nb), ALL );
// C[MC,MR] += alpha (A1[*,MC])^T B1[*,MR]
// = alpha (A1^T)[MC,*] B1[*,MR]
A1_STAR_MC = A1;
Transpose( B1, B1Trans_MR_STAR );
LocalGemm
( orientA, TRANSPOSE, alpha, A1_STAR_MC, B1Trans_MR_STAR, T(1), C );
}
}
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 );
}
}
void SymmetricL1DistanceMatrix(El::UpperOrLower uplo, direction_t dir, T alpha,
const El::ElementalMatrix<T> &APre, T beta, El::ElementalMatrix<T> &CPre) {
if (dir == base::COLUMNS) {
const El::Int r = APre.Height();
const El::Int bsize = El::Blocksize();
const El::Grid& g = APre.Grid();
El::DistMatrixReadProxy<T, T, El::MC, El::MR> AProx(APre);
El::DistMatrixReadWriteProxy<T, T, El::MC, El::MR> CProx(CPre);
auto& A = AProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
El::DistMatrix<T, El::STAR, El::MR> A1_STAR_MR(g);
El::DistMatrix<T, El::STAR, El::MC> A1_STAR_MC(g);
A1_STAR_MC.AlignWith(C);
A1_STAR_MR.AlignWith(C);
El::ScaleTrapezoid(beta, uplo, C);
for(El::Int k = 0; k < r; k += bsize) {
const El::Int nb = std::min(bsize, r - k);
auto A1 = A(El::IR(k, k + nb), El::ALL);
A1_STAR_MC = A1;
A1_STAR_MR = A1;
internal::L1DistanceMatrixTU(uplo, base::COLUMNS, base::COLUMNS,
alpha, A1_STAR_MC, A1_STAR_MR,
T(1.0), C);
}
}
// TODO the rest of the cases.
}
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
GemmTTC
( Orientation orientationOfA,
Orientation orientationOfB,
T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
PushCallStack("internal::GemmTTC");
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 || orientationOfB == NORMAL )
throw std::logic_error
("GemmTTC expects A and B to be (Conjugate)Transposed");
if( A.Width() != C.Height() ||
B.Height() != C.Width() ||
A.Height() != B.Width() )
{
std::ostringstream msg;
msg << "Nonconformal GemmTTC: \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> AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<T> BL(g), BR(g),
B0(g), B1(g), B2(g);
// Temporary distributions
DistMatrix<T,STAR,MC > A1_STAR_MC(g);
DistMatrix<T,VR, STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR > B1AdjOrTrans_STAR_MR(g);
A1_STAR_MC.AlignWith( C );
B1_VR_STAR.AlignWith( C );
B1AdjOrTrans_STAR_MR.AlignWith( C );
// Start the algorithm
Scale( beta, C );
LockedPartitionDown
( A, AT,
AB, 0 );
LockedPartitionRight( B, BL, BR, 0 );
while( AB.Height() > 0 )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
//--------------------------------------------------------------------//
A1_STAR_MC = A1;
B1_VR_STAR = B1;
if( orientationOfB == ADJOINT )
B1AdjOrTrans_STAR_MR.AdjointFrom( B1_VR_STAR );
else
B1AdjOrTrans_STAR_MR.TransposeFrom( B1_VR_STAR );
// C[MC,MR] += alpha (A1[*,MC])^[T/H] (B1[MR,*])^[T/H]
// = alpha (A1^[T/H])[MC,*] (B1^[T/H])[*,MR]
LocalGemm
( orientationOfA, NORMAL,
alpha, A1_STAR_MC, B1AdjOrTrans_STAR_MR, T(1), C );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
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
}
inline void
Trr2kTTTT
( UpperOrLower uplo,
Orientation orientationOfA, Orientation orientationOfB,
Orientation orientationOfC, 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::Trr2kTTTT");
if( E.Height() != E.Width() || A.Height() != C.Height() ||
A.Width() != E.Height() || C.Width() != E.Height() ||
B.Height() != E.Width() || D.Height() != E.Width() ||
A.Height() != B.Width() || C.Height() != D.Width() )
throw std::logic_error("Nonconformal Trr2kTTTT");
#endif
const Grid& g = E.Grid();
DistMatrix<T> AT(g), A0(g),
AB(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> DL(g), DR(g),
D0(g), D1(g), D2(g);
DistMatrix<T,STAR,MC > A1_STAR_MC(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);
A1_STAR_MC.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 );
LockedPartitionDown
( A, AT,
AB, 0 );
LockedPartitionRight( B, BL, BR, 0 );
LockedPartitionDown
( C, CT,
CB, 0 );
LockedPartitionRight( D, DL, DR, 0 );
while( AT.Height() < A.Height() )
{
LockedRepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2 );
LockedRepartitionRight
( BL, /**/ BR,
B0, /**/ B1, B2 );
LockedRepartitionDown
( CT, C0,
/**/ /**/
C1,
CB, C2 );
LockedRepartitionRight
( DL, /**/ DR,
D0, /**/ D1, D2 );
//--------------------------------------------------------------------//
A1_STAR_MC = 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 );
LocalTrr2k
( uplo, orientationOfA, orientationOfC,
alpha, A1_STAR_MC, B1AdjOrTrans_STAR_MR,
C1_STAR_MC, D1AdjOrTrans_STAR_MR,
beta, E );
//--------------------------------------------------------------------//
SlideLockedPartitionRight
( DL, /**/ DR,
D0, D1, /**/ D2 );
SlideLockedPartitionDown
( CT, C0,
C1,
/**/ /**/
CB, C2 );
SlideLockedPartitionRight
( BL, /**/ BR,
B0, B1, /**/ B2 );
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
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
}
inline void
internal::GemmTNC
( 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::GemmTNC");
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("GemmTNC assumes A is (Conjugate)Transposed");
if( A.Width() != C.Height() ||
B.Width() != C.Width() ||
A.Height() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal GemmTNC: \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,MC,MR> AT(g), A0(g),
AB(g), A1(g),
A2(g);
DistMatrix<T,MC,MR> BT(g), B0(g),
BB(g), B1(g),
B2(g);
// Temporary distributions
DistMatrix<T,STAR,MC> A1_STAR_MC(g);
DistMatrix<T,STAR,MR> B1_STAR_MR(g);
// Start the algorithm
Scal( beta, 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 );
A1_STAR_MC.AlignWith( C );
B1_STAR_MR.AlignWith( C );
//--------------------------------------------------------------------//
A1_STAR_MC = A1; // A1[*,MC] <- A1[MC,MR]
B1_STAR_MR = B1; // B1[*,MR] <- B1[MC,MR]
// C[MC,MR] += alpha (A1[*,MC])^T B1[*,MR]
// = alpha (A1^T)[MC,*] B1[*,MR]
internal::LocalGemm
( orientationOfA, NORMAL, alpha, A1_STAR_MC, B1_STAR_MR, (T)1, C );
//--------------------------------------------------------------------//
A1_STAR_MC.FreeAlignments();
B1_STAR_MR.FreeAlignments();
SlideLockedPartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
SlideLockedPartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}