std::unique_ptr<ElDistVector> ElRBFInterpolation::interpolate( const std::unique_ptr<ElDistVector> & values )
{
assert( Phi->Height() > 0 );
assert( H->Height() > 0 );
assert( values->Height() == Phi->Width() );
std::unique_ptr<ElDistVector> result( new ElDistVector( Phi->Grid() ) );
result->AlignRowsWith( *Phi );
El::DistMatrix<double> B = *values;
if ( HCopy.Width() == 0 )
{
HCopy = El::DistMatrix<double> ( *H );
El::DistMatrixReadProxy<double, double, El::MC, El::MR> AProx( HCopy );
auto & A = AProx.Get();
p = El::DistPermutation( A.Grid() );
dSub = El::DistMatrix<double, El::MD, El::STAR> ( A.Grid() );
El::LDL( A, dSub, p, false, El::LDLPivotCtrl<El::Base<double> >() );
}
El::DistMatrixReadProxy<double, double, El::MC, El::MR> AProx( HCopy );
El::DistMatrixReadWriteProxy<double, double, El::MC, El::MR> BProx( B );
auto & A = AProx.Get();
auto & B_LU = BProx.Get();
El::ldl::SolveAfter( A, dSub, p, B_LU, false );
El::Gemm( El::Orientation::NORMAL, El::Orientation::NORMAL, 1.0, *Phi, B, *result );
return result;
}
void HermitianPseudoinverse
( UpperOrLower uplo, ElementalMatrix<F>& APre, Base<F> tolerance )
{
DEBUG_CSE
typedef Base<F> Real;
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
const Grid& g = A.Grid();
// Get the EVD of A
// TODO: Use a relative eigenvalue lower-bound
DistMatrix<Real,VR,STAR> w(g);
DistMatrix<F> Z(g);
HermitianEig( uplo, A, w, Z );
if( tolerance == Real(0) )
{
// Set the tolerance equal to n ||A||_2 eps
const Int n = Z.Height();
const Real eps = limits::Epsilon<Real>();
const Real twoNorm = MaxNorm( w );
tolerance = n*twoNorm*eps;
}
// Invert above the tolerance
auto omegaMap =
[=]( Real omega ) { return ( omega < tolerance ? Real(0) : 1/omega ); };
EntrywiseMap( w, function<Real(Real)>(omegaMap) );
// Form the pseudoinverse
HermitianFromEVD( uplo, A, w, Z );
}
void Pseudoinverse( ElementalMatrix<F>& APre, Base<F> tolerance )
{
DEBUG_CSE
typedef Base<F> Real;
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
const Int m = A.Height();
const Int n = A.Width();
const Grid& g = A.Grid();
const Real eps = limits::Epsilon<Real>();
// Get the SVD of A
DistMatrix<Real,VR,STAR> s(g);
DistMatrix<F> U(g), V(g);
SVDCtrl<Real> ctrl;
ctrl.overwrite = true;
ctrl.bidiagSVDCtrl.approach = COMPACT_SVD;
// TODO(poulson): Let the user change these defaults
ctrl.bidiagSVDCtrl.tolType = RELATIVE_TO_MAX_SING_VAL_TOL;
ctrl.bidiagSVDCtrl.tol =
( tolerance == Real(0) ? Max(m,n)*eps : tolerance );
SVD( A, U, s, V, ctrl );
// Scale U with the inverted (nonzero) singular values, U := U / Sigma
DiagonalSolve( RIGHT, NORMAL, s, U );
// Form pinvA = (U Sigma V^H)^H = V (U Sigma)^H
Gemm( NORMAL, ADJOINT, F(1), V, U, A );
}
void Overwrite
( Orientation orientation,
ElementalMatrix<F>& APre,
const ElementalMatrix<F>& B,
ElementalMatrix<F>& X )
{
DEBUG_ONLY(CSE cse("ls::Overwrite"))
DistMatrixReadProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
DistMatrix<F,MD,STAR> t(A.Grid());
DistMatrix<Base<F>,MD,STAR> d(A.Grid());
const Int m = A.Height();
const Int n = A.Width();
if( m >= n )
{
QR( A, t, d );
qr::SolveAfter( orientation, A, t, d, B, X );
}
else
{
LQ( A, t, d );
lq::SolveAfter( orientation, A, t, d, B, X );
}
}
void HermitianSign
( UpperOrLower uplo,
AbstractDistMatrix<Field>& APre,
const HermitianEigCtrl<Field>& ctrl )
{
EL_DEBUG_CSE
DistMatrixReadWriteProxy<Field,Field,MC,MR> AProx( APre );
auto& A = AProx.Get();
// Get the EVD of A
typedef Base<Field> Real;
const Grid& g = A.Grid();
DistMatrix<Real,VR,STAR> w(g);
DistMatrix<Field> Q(g);
auto ctrlMod( ctrl );
ctrlMod.tridiagEigCtrl.sort = UNSORTED;
HermitianEig( uplo, A, w, Q, ctrlMod );
const Int numLocalEigs = w.LocalHeight();
for( Int iLoc=0; iLoc<numLocalEigs; ++iLoc )
{
const Real omega = w.GetLocal(iLoc,0);
if( omega >= 0 )
w.SetLocal(iLoc,0,Real(1));
else
w.SetLocal(iLoc,0,Real(-1));
}
// Reform the Hermitian matrix with the modified eigenvalues
HermitianFromEVD( uplo, A, w, Q );
}
void NormalUniformSpectrum
( ElementalMatrix<Complex<Real>>& APre, Int n,
Complex<Real> center, Real radius )
{
EL_DEBUG_CSE
typedef Complex<Real> C;
DistMatrixWriteProxy<C,C,MC,MR> AProx( APre );
auto& A = AProx.Get();
const Grid& grid = A.Grid();
A.Resize( n, n );
// Form d and D
vector<C> d( n );
if( grid.Rank() == 0 )
for( Int j=0; j<n; ++j )
d[j] = SampleBall<C>( center, radius );
mpi::Broadcast( d.data(), n, 0, grid.Comm() );
Diagonal( A, d );
// Apply a Haar matrix from both sides
DistMatrix<C> Q(grid);
DistMatrix<C,MD,STAR> t(grid);
DistMatrix<Real,MD,STAR> s(grid);
ImplicitHaar( Q, t, s, n );
// Copy the result into the correct distribution
qr::ApplyQ( LEFT, NORMAL, Q, t, s, A );
qr::ApplyQ( RIGHT, ADJOINT, Q, t, s, A );
}
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 );
}
}
void NMF
( const ElementalMatrix<Real>& APre,
ElementalMatrix<Real>& XPre,
ElementalMatrix<Real>& YPre,
const NMFCtrl<Real>& ctrl )
{
DEBUG_ONLY(CSE cse("NMF"))
DistMatrixReadProxy<Real,Real,MC,MR>
AProx( APre );
DistMatrixReadWriteProxy<Real,Real,MC,MR>
XProx( XPre );
DistMatrixWriteProxy<Real,Real,MC,MR>
YProx( YPre );
auto& A = AProx.GetLocked();
auto& X = XProx.Get();
auto& Y = YProx.Get();
DistMatrix<Real> AAdj(A.Grid()), XAdj(A.Grid()), YAdj(A.Grid());
Adjoint( A, AAdj );
for( Int iter=0; iter<ctrl.maxIter; ++iter )
{
NNLS( X, A, YAdj, ctrl.nnlsCtrl );
Adjoint( YAdj, Y );
NNLS( Y, AAdj, XAdj, ctrl.nnlsCtrl );
Adjoint( XAdj, X );
}
}
void Explicit( AbstractDistMatrix<F>& L, AbstractDistMatrix<F>& APre )
{
EL_DEBUG_CSE
const Grid& g = APre.Grid();
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
DistMatrix<F,MD,STAR> householderScalars(g);
DistMatrix<Base<F>,MD,STAR> signature(g);
LQ( A, householderScalars, signature );
const Int m = A.Height();
const Int n = A.Width();
const Int minDim = Min(m,n);
auto AL = A( IR(0,m), IR(0,minDim) );
Copy( AL, L );
MakeTrapezoidal( LOWER, L );
// TODO: Replace this with an in-place expansion of Q
DistMatrix<F> Q(g);
Identity( Q, A.Height(), A.Width() );
lq::ApplyQ( RIGHT, NORMAL, A, householderScalars, signature, Q );
Copy( Q, APre );
}
SafeProduct<Base<F>> AfterCholesky
( UpperOrLower uplo, const ElementalMatrix<F>& APre )
{
DEBUG_CSE
DistMatrixReadProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.GetLocked();
typedef Base<F> Real;
const Int n = A.Height();
const Grid& g = A.Grid();
DistMatrix<F,MD,STAR> d(g);
GetDiagonal( A, d );
Real localKappa = 0;
if( d.Participating() )
{
const Real scale = Real(n)/Real(2);
const Int nLocalDiag = d.LocalHeight();
for( Int iLoc=0; iLoc<nLocalDiag; ++iLoc )
{
const Real delta = RealPart(d.GetLocal(iLoc,0));
localKappa += Log(delta)/scale;
}
}
SafeProduct<Real> det( n );
det.kappa = mpi::AllReduce( localKappa, g.VCComm() );
det.rho = Real(1);
return det;
}
void Overwrite
( Orientation orientation,
ElementalMatrix<F>& APre,
const ElementalMatrix<F>& B,
ElementalMatrix<F>& X )
{
DEBUG_CSE
DistMatrixReadProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
DistMatrix<F,MD,STAR> phase(A.Grid());
DistMatrix<Base<F>,MD,STAR> signature(A.Grid());
const Int m = A.Height();
const Int n = A.Width();
if( m >= n )
{
QR( A, phase, signature );
qr::SolveAfter( orientation, A, phase, signature, B, X );
}
else
{
LQ( A, phase, signature );
lq::SolveAfter( orientation, A, phase, signature, B, X );
}
}
void ExplicitUnitary
( ElementalMatrix<F>& APre, bool thinQR, const QRCtrl<Base<F>>& ctrl )
{
DEBUG_CSE
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
const Grid& g = A.Grid();
DistMatrix<F,MD,STAR> householderScalars(g);
DistMatrix<Base<F>,MD,STAR> signature(g);
if( ctrl.colPiv )
{
DistPermutation Omega(g);
QR( A, householderScalars, signature, Omega, ctrl );
}
else
QR( A, householderScalars, signature );
if( thinQR )
{
A.Resize( A.Height(), householderScalars.Height() );
ExpandPackedReflectors
( LOWER, VERTICAL, CONJUGATED, 0, A, householderScalars );
DiagonalScale( RIGHT, NORMAL, signature, A );
}
else
{
auto ACopy = A;
// TODO: Use an extension of ExpandPackedReflectors to make this faster
Identity( A, A.Height(), A.Height() );
qr::ApplyQ( LEFT, NORMAL, ACopy, householderScalars, signature, A );
}
}
void ApplyQ
( LeftOrRight side,
Orientation orientation,
const AbstractDistMatrix<F>& APre,
const AbstractDistMatrix<F>& householderScalars,
const AbstractDistMatrix<Base<F>>& signature,
AbstractDistMatrix<F>& BPre )
{
EL_DEBUG_CSE
const bool normal = (orientation==NORMAL);
const bool onLeft = (side==LEFT);
const bool applyDFirst = normal==onLeft;
const Int minDim = Min(APre.Height(),APre.Width());
const ForwardOrBackward direction = ( normal==onLeft ? BACKWARD : FORWARD );
const Conjugation conjugation = ( normal ? CONJUGATED : UNCONJUGATED );
DistMatrixReadProxy<F,F,MC,MR> AProx( APre );
DistMatrixReadWriteProxy<F,F,MC,MR> BProx( BPre );
auto& A = AProx.GetLocked();
auto& B = BProx.Get();
const Int m = B.Height();
const Int n = B.Width();
if( applyDFirst )
{
if( onLeft )
{
auto BTop = B( IR(0,minDim), IR(0,n) );
DiagonalScale( side, orientation, signature, BTop );
}
else
{
auto BLeft = B( IR(0,m), IR(0,minDim) );
DiagonalScale( side, orientation, signature, BLeft );
}
}
ApplyPackedReflectors
( side, LOWER, VERTICAL, direction, conjugation, 0,
A, householderScalars, B );
if( !applyDFirst )
{
if( onLeft )
{
auto BTop = B( IR(0,minDim), IR(0,n) );
DiagonalScale( side, orientation, signature, BTop );
}
else
{
auto BLeft = B( IR(0,m), IR(0,minDim) );
DiagonalScale( side, orientation, signature, BLeft );
}
}
}
void Sign( ElementalMatrix<F>& APre, const SignCtrl<Base<F>> ctrl )
{
DEBUG_CSE
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
sign::Newton( A, ctrl );
}
void Sign( AbstractDistMatrix<Field>& APre, const SignCtrl<Base<Field>> ctrl )
{
EL_DEBUG_CSE
DistMatrixReadWriteProxy<Field,Field,MC,MR> AProx( APre );
auto& A = AProx.Get();
sign::Newton( A, ctrl );
}
SVDInfo GolubReinsch
( AbstractDistMatrix<Field>& APre,
AbstractDistMatrix<Base<Field>>& s,
const SVDCtrl<Base<Field>>& ctrl )
{
EL_DEBUG_CSE
DistMatrixReadWriteProxy<Field,Field,MC,MR> AProx( APre );
auto& A = AProx.Get();
return GolubReinsch( A, s, ctrl );
}
void RuizEquil
( AbstractDistMatrix<Field>& APre,
AbstractDistMatrix<Base<Field>>& dRowPre,
AbstractDistMatrix<Base<Field>>& dColPre,
bool progress )
{
EL_DEBUG_CSE
typedef Base<Field> Real;
ElementalProxyCtrl control;
control.colConstrain = true;
control.rowConstrain = true;
control.colAlign = 0;
control.rowAlign = 0;
DistMatrixReadWriteProxy<Field,Field,MC,MR> AProx( APre, control );
DistMatrixWriteProxy<Real,Real,MC,STAR> dRowProx( dRowPre, control );
DistMatrixWriteProxy<Real,Real,MR,STAR> dColProx( dColPre, control );
auto& A = AProx.Get();
auto& dRow = dRowProx.Get();
auto& dCol = dColProx.Get();
const Int m = A.Height();
const Int n = A.Width();
Ones( dRow, m, 1 );
Ones( dCol, n, 1 );
// TODO(poulson): Expose these as control parameters
// For now, simply hard-code the number of iterations
const Int maxIter = 4;
DistMatrix<Real,MC,STAR> rowScale(A.Grid());
DistMatrix<Real,MR,STAR> colScale(A.Grid());
const Int indent = PushIndent();
for( Int iter=0; iter<maxIter; ++iter )
{
// Rescale the columns
// -------------------
ColumnMaxNorms( A, colScale );
EntrywiseMap( colScale, MakeFunction(DampScaling<Real>) );
DiagonalScale( LEFT, NORMAL, colScale, dCol );
DiagonalSolve( RIGHT, NORMAL, colScale, A );
// Rescale the rows
// ----------------
RowMaxNorms( A, rowScale );
EntrywiseMap( rowScale, MakeFunction(DampScaling<Real>) );
DiagonalScale( LEFT, NORMAL, rowScale, dRow );
DiagonalSolve( LEFT, NORMAL, rowScale, A );
}
SetIndent( indent );
}
void HermitianSign
( UpperOrLower uplo,
ElementalMatrix<F>& APre,
ElementalMatrix<F>& NPre,
const HermitianEigCtrl<F>& ctrl )
{
DEBUG_CSE
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
DistMatrixWriteProxy<F,F,MC,MR> NProx( NPre );
auto& A = AProx.Get();
auto& N = NProx.Get();
// Get the EVD of A
typedef Base<F> Real;
const Grid& g = A.Grid();
DistMatrix<Real,VR,STAR> w(g);
DistMatrix<F> Q(g);
auto ctrlMod( ctrl );
ctrlMod.tridiagEigCtrl.sort = UNSORTED;
HermitianEig( uplo, A, w, Q, ctrlMod );
const Int n = A.Height();
const Int numLocalEigs = w.LocalHeight();
DistMatrix<Real,VR,STAR> wSgn(g), wAbs(g);
wSgn.AlignWith( w );
wAbs.AlignWith( w );
wSgn.Resize( n, 1 );
wAbs.Resize( n, 1 );
for( Int iLoc=0; iLoc<numLocalEigs; ++iLoc )
{
const Real omega = w.GetLocal(iLoc,0);
if( omega >= 0 )
{
wSgn.SetLocal(iLoc,0,Real(1));
wAbs.SetLocal(iLoc,0,omega);
}
else
{
wSgn.SetLocal(iLoc,0,Real(-1));
wAbs.SetLocal(iLoc,0,-omega);
}
}
// Form the Hermitian matrix with the modified eigenvalues
HermitianFromEVD( uplo, A, wSgn, Q );
HermitianFromEVD( uplo, N, wAbs, Q );
}
void GolubReinsch
( AbstractDistMatrix<Field>& APre,
AbstractDistMatrix<Field>& UPre,
AbstractDistMatrix<Base<Field>>& s,
AbstractDistMatrix<Field>& VPre,
const SVDCtrl<Base<Field>>& ctrl )
{
EL_DEBUG_CSE
DistMatrixReadWriteProxy<Field,Field,MC,MR> AProx( APre );
DistMatrixWriteProxy<Field,Field,MC,MR> UProx( UPre );
DistMatrixWriteProxy<Field,Field,MC,MR> VProx( VPre );
auto& A = AProx.Get();
auto& U = UProx.Get();
auto& V = VProx.Get();
return GolubReinsch( A, U, s, V, ctrl );
}
void Sign
( AbstractDistMatrix<Field>& APre,
AbstractDistMatrix<Field>& NPre,
const SignCtrl<Base<Field>> ctrl )
{
EL_DEBUG_CSE
DistMatrixReadWriteProxy<Field,Field,MC,MR> AProx( APre );
DistMatrixWriteProxy<Field,Field,MC,MR> NProx( NPre );
auto& A = AProx.Get();
auto& N = NProx.Get();
DistMatrix<Field> ACopy( A );
sign::Newton( A, ctrl );
Gemm( NORMAL, NORMAL, Field(1), A, ACopy, N );
}
void Sign
( ElementalMatrix<F>& APre,
ElementalMatrix<F>& NPre,
const SignCtrl<Base<F>> ctrl )
{
DEBUG_CSE
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
DistMatrixWriteProxy<F,F,MC,MR> NProx( NPre );
auto& A = AProx.Get();
auto& N = NProx.Get();
DistMatrix<F> ACopy( A );
sign::Newton( A, ctrl );
Gemm( NORMAL, NORMAL, F(1), A, ACopy, N );
}
void
Householder
( AbstractDistMatrix<F>& APre,
AbstractDistMatrix<F>& householderScalarsPre,
AbstractDistMatrix<Base<F>>& signaturePre )
{
DEBUG_CSE
DEBUG_ONLY(AssertSameGrids( APre, householderScalarsPre, signaturePre ))
const Int m = APre.Height();
const Int n = APre.Width();
const Int minDim = Min(m,n);
const Int iOff = m-minDim;
const Int jOff = n-minDim;
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
DistMatrixWriteProxy<F,F,MD,STAR>
householderScalarsProx( householderScalarsPre );
DistMatrixWriteProxy<Base<F>,Base<F>,MD,STAR> signatureProx( signaturePre );
auto& A = AProx.Get();
auto& householderScalars = householderScalarsProx.Get();
auto& signature = signatureProx.Get();
householderScalars.Resize( minDim, 1 );
signature.Resize( minDim, 1 );
const Int bsize = Blocksize();
const Int kLast = LastOffset( minDim, bsize );
for( Int k=kLast; k>=0; k-=bsize )
{
const Int nb = Min(bsize,minDim-k);
const Int ki = k + iOff;
const Int kj = k + jOff;
const Range<Int> ind0Vert( 0, ki ),
ind1( k, k+nb ),
ind1Vert( ki, ki+nb ),
indL( 0, kj+nb );
auto A0L = A( ind0Vert, indL );
auto A1L = A( ind1Vert, indL );
auto householderScalars1 = householderScalars( ind1, ALL );
auto sig1 = signature( ind1, ALL );
PanelHouseholder( A1L, householderScalars1, sig1 );
ApplyQ( RIGHT, ADJOINT, A1L, householderScalars1, sig1, A0L );
}
}
void HermitianFromEVD
( UpperOrLower uplo,
AbstractDistMatrix<F>& APre,
const AbstractDistMatrix<Base<F>>& wPre,
const AbstractDistMatrix<F>& ZPre )
{
DEBUG_CSE
typedef Base<F> Real;
DistMatrixWriteProxy<F,F,MC,MR> AProx( APre );
DistMatrixReadProxy<Real,Real,VR,STAR> wProx( wPre );
DistMatrixReadProxy<F,F,MC,MR> ZProx( ZPre );
auto& A = AProx.Get();
auto& w = wProx.GetLocked();
auto& Z = ZProx.GetLocked();
const Grid& g = A.Grid();
DistMatrix<F,MC, STAR> Z1_MC_STAR(g);
DistMatrix<F,VR, STAR> Z1_VR_STAR(g);
DistMatrix<F,STAR,MR > Z1Adj_STAR_MR(g);
const Int m = Z.Height();
const Int n = Z.Width();
A.Resize( m, m );
if( uplo == LOWER )
MakeTrapezoidal( UPPER, A, 1 );
else
MakeTrapezoidal( LOWER, A, -1 );
const Int bsize = Blocksize();
for( Int k=0; k<n; k+=bsize )
{
const Int nb = Min(bsize,n-k);
auto Z1 = Z( ALL, IR(k,k+nb) );
auto w1 = w( IR(k,k+nb), ALL );
Z1_MC_STAR.AlignWith( A );
Z1_MC_STAR = Z1;
Z1_VR_STAR.AlignWith( A );
Z1_VR_STAR = Z1_MC_STAR;
DiagonalScale( RIGHT, NORMAL, w1, Z1_VR_STAR );
Z1Adj_STAR_MR.AlignWith( A );
Adjoint( Z1_VR_STAR, Z1Adj_STAR_MR );
LocalTrrk( uplo, F(1), Z1_MC_STAR, Z1Adj_STAR_MR, F(1), A );
}
}
void LT_Dot
( T alpha,
const AbstractDistMatrix<T>& APre,
AbstractDistMatrix<T>& CPre,
const bool conjugate,
Int blockSize=2000 )
{
EL_DEBUG_CSE
const Int n = CPre.Height();
const Grid& g = APre.Grid();
const Orientation orient = ( conjugate ? ADJOINT : TRANSPOSE );
DistMatrixReadProxy<T,T,VC,STAR> AProx( APre );
auto& A = AProx.GetLocked();
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& C = CProx.Get();
DistMatrix<T,STAR,STAR> Z( blockSize, blockSize, g );
Zero( Z );
for( Int kOuter=0; kOuter<n; kOuter+=blockSize )
{
const Int nbOuter = Min(blockSize,n-kOuter);
const Range<Int> indOuter( kOuter, kOuter+nbOuter );
auto A1 = A( ALL, indOuter );
auto C11 = C( indOuter, indOuter );
Z.Resize( nbOuter, nbOuter );
Syrk( LOWER, TRANSPOSE, alpha, A1.Matrix(), Z.Matrix(), conjugate );
AxpyContract( T(1), Z, C11 );
for( Int kInner=kOuter+nbOuter; kInner<n; kInner+=blockSize )
{
const Int nbInner = Min(blockSize,n-kInner);
const Range<Int> indInner( kInner, kInner+nbInner );
auto A2 = A( ALL, indInner );
auto C21 = C( indInner, indOuter );
LocalGemm( orient, NORMAL, alpha, A1, A2, Z );
AxpyContract( T(1), Z, C21 );
}
}
}
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 );
}
}
}
void Explicit
( ElementalMatrix<F>& APre,
ElementalMatrix<F>& R,
bool thinQR,
const QRCtrl<Base<F>>& ctrl )
{
DEBUG_ONLY(CSE cse("qr::Explicit"))
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
const Grid& g = A.Grid();
DistMatrix<F,MD,STAR> t(g);
DistMatrix<Base<F>,MD,STAR> d(g);
if( ctrl.colPiv )
{
DistPermutation Omega(g);
QR( A, t, d, Omega, ctrl );
}
else
QR( A, t, d );
const Int m = A.Height();
const Int n = A.Width();
const Int numIts = t.Height();
auto AT = A( IR(0,numIts), IR(0,n) );
Copy( AT, R );
MakeTrapezoidal( UPPER, R );
if( thinQR )
{
A.Resize( m, numIts );
ExpandPackedReflectors( LOWER, VERTICAL, CONJUGATED, 0, A, t );
DiagonalScale( RIGHT, NORMAL, d, A );
}
else
{
auto ACopy = A;
// TODO: Use an extension of ExpandPackedReflectors to make this faster
Identity( A, A.Height(), A.Height() );
qr::ApplyQ( LEFT, NORMAL, ACopy, t, d, A );
}
}
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_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 ExplicitUnitary( AbstractDistMatrix<F>& APre )
{
EL_DEBUG_CSE
const Grid& g = APre.Grid();
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
DistMatrix<F,MD,STAR> householderScalars(g);
DistMatrix<Base<F>,MD,STAR> signature(g);
LQ( A, householderScalars, signature );
// TODO: Replace this with an in-place expansion of Q
DistMatrix<F> Q(g);
Q.AlignWith( A );
Identity( Q, A.Height(), A.Width() );
lq::ApplyQ( RIGHT, NORMAL, A, householderScalars, signature, Q );
Copy( Q, APre );
}
Int
ADMM
( const ElementalMatrix<Real>& APre,
const ElementalMatrix<Real>& B,
ElementalMatrix<Real>& X,
const ADMMCtrl<Real>& ctrl )
{
DEBUG_CSE
const Real maxReal = limits::Max<Real>();
DistMatrixReadProxy<Real,Real,MC,MR> AProx( APre );
auto& A = AProx.GetLocked();
DistMatrix<Real> Q(A.Grid()), C(A.Grid());
Herk( LOWER, ADJOINT, Real(1), A, Q );
Gemm( ADJOINT, NORMAL, Real(-1), A, B, C );
return qp::box::ADMM( Q, C, Real(0), maxReal, X, ctrl );
}