int QDWH
( DistMatrix<F>& A,
typename Base<F>::type lowerBound,
typename Base<F>::type upperBound )
{
#ifndef RELEASE
PushCallStack("QDWH");
#endif
typedef typename Base<F>::type R;
const Grid& g = A.Grid();
const int height = A.Height();
const int width = A.Width();
const R oneHalf = R(1)/R(2);
const R oneThird = R(1)/R(3);
if( height < width )
throw std::logic_error("Height cannot be less than width");
const R epsilon = lapack::MachineEpsilon<R>();
const R tol = 5*epsilon;
const R cubeRootTol = Pow(tol,oneThird);
// Form the first iterate
Scale( 1/upperBound, A );
int numIts=0;
R frobNormADiff;
DistMatrix<F> ALast( g );
DistMatrix<F> Q( height+width, width, g );
DistMatrix<F> QT(g), QB(g);
PartitionDown( Q, QT,
QB, height );
DistMatrix<F> C( g );
DistMatrix<F> ATemp( g );
do
{
++numIts;
ALast = A;
R L2;
Complex<R> dd, sqd;
if( Abs(1-lowerBound) < tol )
{
L2 = 1;
dd = 0;
sqd = 1;
}
else
{
L2 = lowerBound*lowerBound;
dd = Pow( 4*(1-L2)/(L2*L2), oneThird );
sqd = Sqrt( 1+dd );
}
const Complex<R> arg = 8 - 4*dd + 8*(2-L2)/(L2*sqd);
const R a = (sqd + Sqrt( arg )/2).real;
const R b = (a-1)*(a-1)/4;
const R c = a+b-1;
const Complex<R> alpha = a-b/c;
const Complex<R> beta = b/c;
lowerBound = lowerBound*(a+b*L2)/(1+c*L2);
if( c > 100 )
{
//
// The standard QR-based algorithm
//
QT = A;
Scale( Sqrt(c), QT );
MakeIdentity( QB );
ExplicitQR( Q );
Gemm( NORMAL, ADJOINT, alpha/Sqrt(c), QT, QB, beta, A );
}
else
{
//
// Use faster Cholesky-based algorithm since A is well-conditioned
//
Identity( width, width, C );
Herk( LOWER, ADJOINT, F(c), A, F(1), C );
Cholesky( LOWER, C );
ATemp = A;
Trsm( RIGHT, LOWER, ADJOINT, NON_UNIT, F(1), C, ATemp );
Trsm( RIGHT, LOWER, NORMAL, NON_UNIT, F(1), C, ATemp );
Scale( beta, A );
Axpy( alpha, ATemp, A );
}
Axpy( F(-1), A, ALast );
frobNormADiff = Norm( ALast, FROBENIUS_NORM );
}
while( frobNormADiff > cubeRootTol || Abs(1-lowerBound) > tol );
#ifndef RELEASE
PopCallStack();
#endif
return numIts;
}
QDWHInfo
QDWHInner
( AbstractDistMatrix<F>& APre, Base<F> sMinUpper, const QDWHCtrl& ctrl )
{
EL_DEBUG_CSE
DistMatrixReadWriteProxy<F,F,MC,MR> AProx( APre );
auto& A = AProx.Get();
typedef Base<F> Real;
typedef Complex<Real> Cpx;
const Int m = A.Height();
const Int n = A.Width();
const Real oneThird = Real(1)/Real(3);
if( m < n )
LogicError("Height cannot be less than width");
QDWHInfo info;
QRCtrl<Base<F>> qrCtrl;
qrCtrl.colPiv = ctrl.colPiv;
const Real eps = limits::Epsilon<Real>();
const Real tol = 5*eps;
const Real cubeRootTol = Pow(tol,oneThird);
Real L = sMinUpper / Sqrt(Real(n));
const Grid& g = A.Grid();
DistMatrix<F> ALast(g), ATemp(g), C(g);
DistMatrix<F> Q( m+n, n, g );
auto QT = Q( IR(0,m ), ALL );
auto QB = Q( IR(m,END), ALL );
Real frobNormADiff;
while( info.numIts < ctrl.maxIts )
{
ALast = A;
Real L2;
Cpx dd, sqd;
if( Abs(1-L) < tol )
{
L2 = 1;
dd = 0;
sqd = 1;
}
else
{
L2 = L*L;
dd = Pow( 4*(1-L2)/(L2*L2), oneThird );
sqd = Sqrt( Real(1)+dd );
}
const Cpx arg = Real(8) - Real(4)*dd + Real(8)*(2-L2)/(L2*sqd);
const Real a = (sqd + Sqrt(arg)/Real(2)).real();
const Real b = (a-1)*(a-1)/4;
const Real c = a+b-1;
const Real alpha = a-b/c;
const Real beta = b/c;
L = L*(a+b*L2)/(1+c*L2);
if( c > 100 )
{
//
// The standard QR-based algorithm
//
QT = A;
QT *= Sqrt(c);
MakeIdentity( QB );
qr::ExplicitUnitary( Q, true, qrCtrl );
Gemm( NORMAL, ADJOINT, F(alpha/Sqrt(c)), QT, QB, F(beta), A );
++info.numQRIts;
}
else
{
//
// Use faster Cholesky-based algorithm since A is well-conditioned
//
Identity( C, n, n );
Herk( LOWER, ADJOINT, c, A, Real(1), C );
Cholesky( LOWER, C );
ATemp = A;
Trsm( RIGHT, LOWER, ADJOINT, NON_UNIT, F(1), C, ATemp );
Trsm( RIGHT, LOWER, NORMAL, NON_UNIT, F(1), C, ATemp );
A *= beta;
Axpy( alpha, ATemp, A );
++info.numCholIts;
}
++info.numIts;
ALast -= A;
frobNormADiff = FrobeniusNorm( ALast );
if( frobNormADiff <= cubeRootTol && Abs(1-L) <= tol )
break;
}
return info;
}
int Halley
( DistMatrix<F>& A, typename Base<F>::type upperBound )
{
#ifndef RELEASE
PushCallStack("Halley");
#endif
typedef typename Base<F>::type R;
const Grid& g = A.Grid();
const int height = A.Height();
const int width = A.Width();
const R oneHalf = R(1)/R(2);
const R oneThird = R(1)/R(3);
if( height < width )
throw std::logic_error("Height cannot be less than width");
const R epsilon = lapack::MachineEpsilon<R>();
const R tol = 5*epsilon;
const R cubeRootTol = Pow(tol,oneThird);
const R a = 3;
const R b = 1;
const R c = 3;
// Form the first iterate
Scale( 1/upperBound, A );
int numIts=0;
R frobNormADiff;
DistMatrix<F> ALast( g );
DistMatrix<F> Q( height+width, width, g );
DistMatrix<F> QT(g), QB(g);
PartitionDown( Q, QT,
QB, height );
DistMatrix<F> C( g );
DistMatrix<F> ATemp( g );
do
{
if( numIts > 100 )
throw std::runtime_error("Halley iteration did not converge");
++numIts;
ALast = A;
// TODO: Come up with a test for when we can use the Cholesky approach
if( true )
{
//
// The standard QR-based algorithm
//
QT = A;
Scale( Sqrt(c), QT );
MakeIdentity( QB );
ExplicitQR( Q );
Gemm( NORMAL, ADJOINT, F(a-b/c)/Sqrt(c), QT, QB, F(b/c), A );
}
else
{
//
// Use faster Cholesky-based algorithm since A is well-conditioned
//
Identity( width, width, C );
Herk( LOWER, ADJOINT, F(c), A, F(1), C );
Cholesky( LOWER, C );
ATemp = A;
Trsm( RIGHT, LOWER, ADJOINT, NON_UNIT, F(1), C, ATemp );
Trsm( RIGHT, LOWER, NORMAL, NON_UNIT, F(1), C, ATemp );
Scale( b/c, A );
Axpy( a-b/c, ATemp, A );
}
Axpy( F(-1), A, ALast );
frobNormADiff = Norm( ALast, FROBENIUS_NORM );
}
while( frobNormADiff > cubeRootTol );
#ifndef RELEASE
PopCallStack();
#endif
return numIts;
}