inline void
RLVB
( Conjugation conjugation, int offset,
const Matrix<Complex<R> >& H,
const Matrix<Complex<R> >& t,
Matrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("apply_packed_reflectors::RLVB");
if( offset > 0 || offset < -H.Height() )
throw std::logic_error("Transforms out of bounds");
if( H.Height() != A.Width() )
throw std::logic_error
("Height of transforms must equal width of target matrix");
if( t.Height() != H.DiagonalLength( offset ) )
throw std::logic_error("t must be the same length as H's offset diag");
#endif
typedef Complex<R> C;
Matrix<C>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<C> ARight;
Matrix<C>
tT, t0,
tB, t1,
t2;
Matrix<C> SInv, Z;
LockedPartitionUpDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionUp
( t, tT,
tB, 0 );
while( HBR.Height() < H.Height() && HBR.Width() < H.Width() )
{
LockedRepartitionUpDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
const int HPanHeight = H11.Height() + H21.Height();
const int HPanWidth =
std::min( H11.Width(), std::max(HPanHeight+offset,0) );
const int leftover = A.Width()-HPanHeight;
LockedView( HPan, H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
LockedRepartitionUp
( tT, t0,
t1,
/**/ /**/
tB, t2, HPanWidth );
View( ARight, A, 0, leftover, A.Height(), HPanHeight );
Zeros( ARight.Height(), HPan.Width(), Z );
Zeros( HPan.Width(), HPan.Width(), SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, LOWER, offset, HPanCopy );
SetDiagonal( LEFT, offset, HPanCopy, C(1) );
Herk( LOWER, ADJOINT, C(1), HPanCopy, C(0), SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, NORMAL, C(1), ARight, HPanCopy, C(0), Z );
Trsm( RIGHT, LOWER, NORMAL, NON_UNIT, C(1), SInv, Z );
Gemm( NORMAL, ADJOINT, C(-1), Z, HPanCopy, C(1), ARight );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
SlideLockedPartitionUp
( tT, t0,
/**/ /**/
t1,
tB, t2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
RLVB( int offset, const Matrix<R>& H, Matrix<R>& A )
{
#ifndef RELEASE
PushCallStack("apply_packed_reflectors::RLVB");
if( offset > 0 || offset < -H.Height() )
throw std::logic_error("Transforms out of bounds");
if( H.Height() != A.Width() )
throw std::logic_error
("Height of transforms must equal width of target matrix");
#endif
Matrix<R>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<R> ARight;
Matrix<R> SInv, Z;
LockedPartitionUpDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
while( HBR.Height() < H.Height() && HBR.Width() < H.Width() )
{
LockedRepartitionUpDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
const int HPanHeight = H11.Height() + H21.Height();
const int HPanWidth =
std::min( H11.Width(), std::max(HPanHeight+offset,0) );
const int leftover = A.Width()-HPanHeight;
LockedView( HPan, H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
View( ARight, A, 0, leftover, A.Height(), HPanHeight );
Zeros( ARight.Height(), HPanWidth, Z );
Zeros( HPanWidth, HPanWidth, SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, LOWER, offset, HPanCopy );
SetDiagonal( LEFT, offset, HPanCopy, R(1) );
Syrk( LOWER, TRANSPOSE, R(1), HPanCopy, R(0), SInv );
HalveMainDiagonal( SInv );
Gemm( NORMAL, NORMAL, R(1), ARight, HPanCopy, R(0), Z );
Trsm( RIGHT, LOWER, NORMAL, NON_UNIT, R(1), SInv, Z );
Gemm( NORMAL, TRANSPOSE, R(-1), Z, HPanCopy, R(1), ARight );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
RowEchelon( Matrix<F>& A, Matrix<F>& B )
{
#ifndef RELEASE
CallStackEntry entry("RowEchelon");
if( A.Height() != B.Height() )
LogicError("A and B must be the same height");
#endif
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02, APan,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
BT, B0,
BB, B1,
B2;
Matrix<Int> p1;
// Pivot composition
std::vector<Int> image, preimage;
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( B, BT,
BB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( BT, B0,
/**/ /**/
B1,
BB, B2 );
View2x1
( APan, A12,
A22 );
//--------------------------------------------------------------------//
lu::Panel( APan, p1, A00.Height() );
ComposePivots( p1, A00.Height(), image, preimage );
ApplyRowPivots( BB, image, preimage );
Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 );
Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, B1 );
Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 );
Gemm( NORMAL, NORMAL, F(-1), A21, B1, F(1), B2 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlidePartitionDown
( BT, B0,
B1,
/**/ /**/
BB, B2 );
}
}
QDWHInfo QDWHInner( Matrix<F>& A, Base<F> sMinUpper, const QDWHCtrl& ctrl )
{
EL_DEBUG_CSE
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));
Real frobNormADiff;
Matrix<F> ALast, ATemp, C;
Matrix<F> Q( m+n, n );
auto QT = Q( IR(0,m ), ALL );
auto QB = Q( IR(m,END), ALL );
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;
}
inline void
LUHF
( Conjugation conjugation, Int offset,
const Matrix<F>& H, const Matrix<F>& t, Matrix<F>& A )
{
#ifndef RELEASE
CallStackEntry cse("apply_packed_reflectors::LUHF");
// TODO: Proper dimension checks
if( t.Height() != H.DiagonalLength(offset) )
LogicError("t must be the same length as H's offset diag");
#endif
Matrix<F>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<F>
AT, A0,
AB, A1,
A2;
Matrix<F>
tT, t0,
tB, t1,
t2;
Matrix<F> SInv, Z;
LockedPartitionDownOffsetDiagonal
( offset,
H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionDown
( t, tT,
tB, 0 );
PartitionDown
( A, AT,
AB, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
LockedRepartitionDown
( tT, t0,
/**/ /**/
t1,
tB, t2 );
RepartitionDown
( AT, A0,
/**/ /**/
A1,
AB, A2, H11.Height() );
LockedView1x2( HPan, H11, H12 );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTriangular( UPPER, HPanCopy );
SetDiagonal( HPanCopy, F(1) );
Herk( LOWER, NORMAL, F(1), HPanCopy, SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, NORMAL, F(1), HPanCopy, AB, Z );
Trsm( LEFT, LOWER, NORMAL, NON_UNIT, F(1), SInv, Z );
Gemm( ADJOINT, NORMAL, F(-1), HPanCopy, Z, F(1), AB );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
SlideLockedPartitionDown
( tT, t0,
t1,
/**/ /**/
tB, t2 );
SlidePartitionDown
( AT, A0,
A1,
/**/ /**/
AB, A2 );
}
}
inline void
TwoSidedTrsmUVar4( UnitOrNonUnit diag, Matrix<F>& A, const Matrix<F>& U )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrsmUVar4");
if( A.Height() != A.Width() )
throw std::logic_error("A must be square");
if( U.Height() != U.Width() )
throw std::logic_error("Triangular matrices must be square");
if( A.Height() != U.Height() )
throw std::logic_error("A and U must be the same size");
#endif
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
UTL, UTR, U00, U01, U02,
UBL, UBR, U10, U11, U12,
U20, U21, U22;
// Temporary products
Matrix<F> Y12;
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
//--------------------------------------------------------------------//
// A01 := A01 inv(U11)
Trsm( RIGHT, UPPER, NORMAL, diag, F(1), U11, A01 );
// A11 := inv(U11)' A11 inv(U11)
TwoSidedTrsmUUnb( diag, A11, U11 );
// A02 := A02 - A01 U12
Gemm( NORMAL, NORMAL, F(-1), A01, U12, F(1), A02 );
// Y12 := A11 U12
Zeros( A12.Height(), A12.Width(), Y12 );
Hemm( LEFT, UPPER, F(1), A11, U12, F(0), Y12 );
// A12 := inv(U11)' A12
Trsm( LEFT, UPPER, ADJOINT, diag, F(1), U11, A12 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
// A22 := A22 - (A12' U12 + U12' A12)
Her2k( UPPER, ADJOINT, F(-1), A12, U12, F(1), A22 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/**********************************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
TrtrsmLLN
( UnitOrNonUnit diag, F alpha, const Matrix<F>& L, Matrix<F>& X,
bool checkIfSingular=true )
{
#ifndef RELEASE
CallStackEntry entry("internal::TrtrsmLLN");
#endif
// Matrix views
Matrix<F>
LTL, LTR, L00, L01, L02,
LBL, LBR, L10, L11, L12,
L20, L21, L22;
Matrix<F>
XTL, XTR, X00, X01, X02,
XBL, XBR, X10, X11, X12,
X20, X21, X22;
Matrix<F> Z11;
// Start the algorithm
ScaleTrapezoid( alpha, LOWER, X );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionDownDiagonal
( X, XTL, XTR,
XBL, XBR, 0 );
while( XBR.Height() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionDownDiagonal
( XTL, /**/ XTR, X00, /**/ X01, X02,
/*************/ /******************/
/**/ X10, /**/ X11, X12,
XBL, /**/ XBR, X20, /**/ X21, X22 );
//--------------------------------------------------------------------//
Trsm( LEFT, LOWER, NORMAL, diag, F(1), L11, X10, checkIfSingular );
TrtrsmLLNUnb( diag, F(1), L11, X11 );
Gemm( NORMAL, NORMAL, F(-1), L21, X10, F(1), X20 );
Z11 = X11;
MakeTriangular( LOWER, Z11 );
Gemm( NORMAL, NORMAL, F(-1), L21, Z11, F(1), X21 );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionDownDiagonal
( XTL, /**/ XTR, X00, X01, /**/ X02,
/**/ X10, X11, /**/ X12,
/*************/ /******************/
XBL, /**/ XBR, X20, X21, /**/ X22 );
}
}
void TransformColumns
( const Matrix<F>& V,
DistMatrix<F,MC,MR,BLOCK>& A )
{
DEBUG_CSE
const Int width = A.Width();
const Grid& grid = A.Grid();
const Int blockWidth = A.BlockWidth();
const Int firstBlockWidth = blockWidth - A.RowCut();
if( width <= firstBlockWidth || grid.Width() == 1 )
{
if( grid.Col() == A.ColOwner(0) )
{
// This process row can locally update its portion of A
TransformColumns( V, A.Matrix() );
}
}
else if( width <= firstBlockWidth + blockWidth )
{
const int firstCol = A.ColOwner( 0 );
const int secondCol = A.ColOwner( firstBlockWidth );
if( grid.Col() == firstCol )
{
//
// Replace A with
//
// | ALeft, ARight | | VLeft, VRight |,
//
// where ALeft is owned by this process column and ARight by the
// next.
//
// Partition space for the combined matrix
Matrix<F> ACombine( A.LocalHeight(), width );
auto ALeft = ACombine( ALL, IR(0,firstBlockWidth) );
auto ARight = ACombine( ALL, IR(firstBlockWidth,END) );
// Copy our portion into the combined matrix
ALeft = A.LockedMatrix();
// Exchange the data
El::SendRecv( ALeft, ARight, A.RowComm(), secondCol, secondCol );
// Form our portion of the result
auto VLeft = V( ALL, IR(0,firstBlockWidth) );
Gemm( NORMAL, NORMAL, F(1), ACombine, VLeft, A.Matrix() );
}
else if( grid.Col() == secondCol )
{
//
// Replace A with
//
// | ALeft, ARight | | VLeft, VRight |,
//
// where ALeft is owned by the previous process column and ARight
// by this one.
//
// Partition space for the combined matrix
Matrix<F> ACombine( A.LocalHeight(), width );
auto ALeft = ACombine( ALL, IR(0,firstBlockWidth) );
auto ARight = ACombine( ALL, IR(firstBlockWidth,END) );
// Copy our portion into the combined matrix
ARight = A.LockedMatrix();
// Exchange the data
El::SendRecv( ARight, ALeft, A.RowComm(), firstCol, firstCol );
// Form our portion of the result
auto VRight = V( ALL, IR(firstBlockWidth,END) );
Gemm( NORMAL, NORMAL, F(1), ACombine, VRight, A.Matrix() );
}
}
else
{
// Fall back to the entire process column interacting.
// TODO(poulson): Only form the subset of the result that we need.
DistMatrix<F,MC,STAR,BLOCK> A_MC_STAR( A );
Matrix<F> ALocCopy( A_MC_STAR.Matrix() );
Gemm( NORMAL, NORMAL, F(1), ALocCopy, V, A_MC_STAR.Matrix() );
A = A_MC_STAR;
}
}
inline void
TwoSidedTrsmUVar2( UnitOrNonUnit diag, Matrix<F>& A, const Matrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar2");
if( A.Height() != A.Width() )
LogicError("A must be square");
if( U.Height() != U.Width() )
LogicError("Triangular matrices must be square");
if( A.Height() != U.Height() )
LogicError("A and U must be the same size");
#endif
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
UTL, UTR, U00, U01, U02,
UBL, UBR, U10, U11, U12,
U20, U21, U22;
// Temporary products
Matrix<F> Y01;
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
//--------------------------------------------------------------------//
// Y01 := A00 U01
Zeros( Y01, A01.Height(), A01.Width() );
Hemm( LEFT, UPPER, F(1), A00, U01, F(0), Y01 );
// A01 := A01 - 1/2 Y01
Axpy( F(-1)/F(2), Y01, A01 );
// A11 := A11 - (U01' A01 + A01' U01)
Her2k( UPPER, ADJOINT, F(-1), U01, A01, F(1), A11 );
// A11 := inv(U11)' A11 inv(U11)
TwoSidedTrsmUUnb( diag, A11, U11 );
// A12 := A12 - A02' U01
Gemm( ADJOINT, NORMAL, F(-1), A02, U01, F(1), A12 );
// A12 := inv(U11)' A12
Trsm( LEFT, UPPER, ADJOINT, diag, F(1), U11, A12 );
// A01 := A01 - 1/2 Y01
Axpy( F(-1)/F(2), Y01, A01 );
// A01 := A01 inv(U11)
Trsm( RIGHT, UPPER, NORMAL, diag, F(1), U11, A01 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
}
}
inline void
ApplyPackedReflectorsRLHF
( int offset, const Matrix<R>& H, Matrix<R>& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsRLHF");
if( offset > 0 || offset < -H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Width() )
throw std::logic_error
("Width of transforms must equal width of target matrix");
#endif
Matrix<R>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<R> ALeft;
Matrix<R> SInv, Z;
LockedPartitionDownDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
{
LockedRepartitionDownDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
const int HPanWidth = H10.Width() + H11.Width();
const int HPanOffset =
std::min( H11.Height(), std::max(-offset-H00.Height(),0) );
const int HPanHeight = H11.Height()-HPanOffset;
HPan.LockedView( H, H00.Height()+HPanOffset, 0, HPanHeight, HPanWidth );
ALeft.View( A, 0, 0, A.Height(), HPanWidth );
Zeros( ALeft.Height(), HPan.Height(), Z );
Zeros( HPan.Height(), HPan.Height(), SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( RIGHT, LOWER, offset, HPanCopy );
SetDiagonalToOne( RIGHT, offset, HPanCopy );
Syrk( UPPER, NORMAL, R(1), HPanCopy, R(0), SInv );
HalveMainDiagonal( SInv );
Gemm( NORMAL, TRANSPOSE, R(1), ALeft, HPanCopy, R(0), Z );
Trsm( RIGHT, UPPER, NORMAL, NON_UNIT, R(1), SInv, Z );
Gemm( NORMAL, NORMAL, R(-1), Z, HPanCopy, R(1), ALeft );
//--------------------------------------------------------------------//
SlideLockedPartitionDownDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
void Tikhonov
( Orientation orientation,
const ElementalMatrix<F>& APre,
const ElementalMatrix<F>& BPre,
const ElementalMatrix<F>& G,
ElementalMatrix<F>& XPre,
TikhonovAlg alg )
{
DEBUG_CSE
DistMatrixReadProxy<F,F,MC,MR>
AProx( APre ),
BProx( BPre );
DistMatrixWriteProxy<F,F,MC,MR>
XProx( XPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& X = XProx.Get();
const bool normal = ( orientation==NORMAL );
const Int m = ( normal ? A.Height() : A.Width() );
const Int n = ( normal ? A.Width() : A.Height() );
if( G.Width() != n )
LogicError("Tikhonov matrix was the wrong width");
if( orientation == TRANSPOSE && IsComplex<F>::value )
LogicError("Transpose version of complex Tikhonov not yet supported");
if( m >= n )
{
DistMatrix<F> Z(A.Grid());
if( alg == TIKHONOV_CHOLESKY )
{
if( orientation == NORMAL )
Herk( LOWER, ADJOINT, Base<F>(1), A, Z );
else
Herk( LOWER, NORMAL, Base<F>(1), A, Z );
Herk( LOWER, ADJOINT, Base<F>(1), G, Base<F>(1), Z );
Cholesky( LOWER, Z );
}
else
{
const Int mG = G.Height();
Zeros( Z, m+mG, n );
auto ZT = Z( IR(0,m), IR(0,n) );
auto ZB = Z( IR(m,m+mG), IR(0,n) );
if( orientation == NORMAL )
ZT = A;
else
Adjoint( A, ZT );
ZB = G;
qr::ExplicitTriang( Z );
}
if( orientation == NORMAL )
Gemm( ADJOINT, NORMAL, F(1), A, B, X );
else
Gemm( NORMAL, NORMAL, F(1), A, B, X );
cholesky::SolveAfter( LOWER, NORMAL, Z, X );
}
else
{
LogicError("This case not yet supported");
}
}
inline void
Cannon_NN
( T alpha, const DistMatrix<T>& A,
const DistMatrix<T>& B,
T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
CallStackEntry entry("gemm::Cannon_NN");
if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
LogicError("{A,B,C} must have the same grid");
if( A.Height() != C.Height() ||
B.Width() != C.Width() ||
A.Width() != B.Height() )
{
std::ostringstream msg;
msg << "Nonconformal matrices: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B ~ " << B.Height() << " x " << B.Width() << "\n"
<< " C ~ " << C.Height() << " x " << C.Width() << "\n";
LogicError( msg.str() );
}
#endif
const Grid& g = A.Grid();
if( g.Height() != g.Width() )
LogicError("Process grid must be square for Cannon's");
if( C.ColAlignment() != A.ColAlignment() ||
C.RowAlignment() != B.RowAlignment() )
LogicError("C is not properly aligned");
const Int row = g.Row();
const Int col = g.Col();
const Int pSqrt = g.Height();
mpi::Comm rowComm = g.RowComm();
mpi::Comm colComm = g.ColComm();
if( A.Width() % pSqrt != 0 )
LogicError("For now, width(A) must be integer multiple of sqrt(p)");
// Begin by scaling our local portion of C
Scale( beta, C );
// Load the initial A and B packages (may want to transpose B...)
const Int localHeightA = A.LocalHeight();
const Int localHeightB = B.LocalHeight();
const Int localWidthA = A.LocalWidth();
const Int localWidthB = B.LocalWidth();
Matrix<T> pkgA(localHeightA,localWidthA,localHeightA),
pkgB(localHeightB,localWidthB,localHeightB);
for( Int jLoc=0; jLoc<localWidthA; ++jLoc )
MemCopy
( pkgA.Buffer(0,jLoc), A.LockedBuffer(0,jLoc), localHeightA );
for( Int jLoc=0; jLoc<localWidthB; ++jLoc )
MemCopy
( pkgB.Buffer(0,jLoc), B.LockedBuffer(0,jLoc), localHeightB );
// Perform the initial circular shifts so that our A and B packages align
const Int rowShiftA = A.RowShift();
const Int colShiftB = B.ColShift();
const Int leftInitA = (col+pSqrt-colShiftB) % pSqrt;
const Int rightInitA = (col+colShiftB) % pSqrt;
const Int aboveInitB = (row+pSqrt-rowShiftA) % pSqrt;
const Int belowInitB = (row+rowShiftA) % pSqrt;
const Int pkgSizeA = localHeightA*localWidthA;
const Int pkgSizeB = localHeightB*localWidthB;
mpi::SendRecv( pkgA.Buffer(), pkgSizeA, leftInitA, rightInitA, rowComm );
mpi::SendRecv( pkgB.Buffer(), pkgSizeB, aboveInitB, belowInitB, colComm );
// Now begin the data flow
const Int aboveRow = (row+pSqrt-1) % pSqrt;
const Int belowRow = (row+1) % pSqrt;
const Int leftCol = (col+pSqrt-1) % pSqrt;
const Int rightCol = (col+1) % pSqrt;
for( Int q=0; q<pSqrt; ++q )
{
Gemm( NORMAL, NORMAL, alpha, pkgA, pkgB, T(1), C.Matrix() );
if( q != pSqrt-1 )
{
mpi::SendRecv
( pkgA.Buffer(), pkgSizeA, leftCol, rightCol, rowComm );
mpi::SendRecv
( pkgB.Buffer(), pkgSizeB, aboveRow, belowRow, colComm );
}
}
}
int QDWH
( Matrix<F>& A,
typename Base<F>::type lowerBound,
typename Base<F>::type upperBound )
{
#ifndef RELEASE
PushCallStack("QDWH");
#endif
typedef typename Base<F>::type R;
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;
Matrix<F> ALast;
Matrix<F> Q( height+width, width );
Matrix<F> QT, QB;
PartitionDown( Q, QT,
QB, height );
Matrix<F> C;
Matrix<F> ATemp;
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;
}
inline void
ApplyPackedReflectorsLUHB
( Conjugation conjugation, int offset,
const Matrix<Complex<R> >& H,
const Matrix<Complex<R> >& t,
Matrix<Complex<R> >& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsLUHB");
if( offset < 0 || offset > H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Height() )
throw std::logic_error
("Width of transforms must equal height of target matrix");
if( t.Height() != H.DiagonalLength( offset ) )
throw std::logic_error("t must be the same length as H's offset diag");
#endif
typedef Complex<R> C;
Matrix<C>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<C> ABottom;
Matrix<C>
tT, t0,
tB, t1,
t2;
Matrix<C> SInv, Z;
LockedPartitionUpDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
LockedPartitionUp
( t, tT,
tB, 0 );
while( HBR.Height() < H.Height() && HBR.Width() < H.Width() )
{
LockedRepartitionUpDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
const int HPanWidth = H11.Width() + H12.Width();
const int HPanHeight =
std::min( H11.Height(), std::max(HPanWidth-offset,0) );
const int leftover = A.Height()-HPanWidth;
HPan.LockedView( H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
LockedRepartitionUp
( tT, t0,
t1,
/**/ /**/
tB, t2, HPanHeight );
ABottom.View( A, leftover, 0, HPanWidth, A.Width() );
Zeros( HPanHeight, ABottom.Width(), Z );
Zeros( HPanHeight, HPanHeight, SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, UPPER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
Herk( UPPER, NORMAL, C(1), HPanCopy, C(0), SInv );
FixDiagonal( conjugation, t1, SInv );
Gemm( NORMAL, NORMAL, C(1), HPanCopy, ABottom, C(0), Z );
Trsm( LEFT, UPPER, NORMAL, NON_UNIT, C(1), SInv, Z );
Gemm( ADJOINT, NORMAL, C(-1), HPanCopy, Z, C(1), ABottom );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
SlideLockedPartitionUp
( tT, t0,
/**/ /**/
t1,
tB, t2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
Inverse( Matrix<F>& A )
{
#ifndef RELEASE
PushCallStack("Inverse");
if( A.Height() != A.Width() )
throw std::logic_error("Cannot invert non-square matrices");
#endif
Matrix<int> p;
LU( A, p );
TriangularInverse( UPPER, NON_UNIT, A );
// Solve inv(A) L = inv(U) for inv(A)
Matrix<F> ATL, ATR,
ABL, ABR;
Matrix<F> A00, A01, A02,
A10, A11, A12,
A20, A21, A22;
Matrix<F> A1, A2;
Matrix<F> L11,
L21;
PartitionUpDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ABR.Height() < A.Height() )
{
RepartitionUpDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
View( A1, A, 0, A00.Width(), A.Height(), A01.Width() );
View( A2, A, 0, A00.Width()+A01.Width(), A.Height(), A02.Width() );
//--------------------------------------------------------------------//
// Copy out L1
L11 = A11;
L21 = A21;
// Zero the strictly lower triangular portion of A1
MakeTrapezoidal( LEFT, UPPER, 0, A11 );
Zero( A21 );
// Perform the lazy update of A1
Gemm( NORMAL, NORMAL, F(-1), A2, L21, F(1), A1 );
// Solve against this diagonal block of L11
Trsm( RIGHT, LOWER, NORMAL, UNIT, F(1), L11, A1 );
//--------------------------------------------------------------------//
SlidePartitionUpDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /*******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
}
// inv(A) := inv(A) P
ApplyInverseColumnPivots( A, p );
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
ApplyPackedReflectorsLUHB
( int offset, const Matrix<R>& H, Matrix<R>& A )
{
#ifndef RELEASE
PushCallStack("internal::ApplyPackedReflectorsLUHB");
if( offset < 0 || offset > H.Width() )
throw std::logic_error("Transforms out of bounds");
if( H.Width() != A.Height() )
throw std::logic_error
("Width of transforms must equal height of target matrix");
#endif
Matrix<R>
HTL, HTR, H00, H01, H02, HPan, HPanCopy,
HBL, HBR, H10, H11, H12,
H20, H21, H22;
Matrix<R> ABottom;
Matrix<R> SInv, Z;
LockedPartitionUpDiagonal
( H, HTL, HTR,
HBL, HBR, 0 );
while( HBR.Height() < H.Height() && HBR.Width() < H.Width() )
{
LockedRepartitionUpDiagonal
( HTL, /**/ HTR, H00, H01, /**/ H02,
/**/ H10, H11, /**/ H12,
/*************/ /******************/
HBL, /**/ HBR, H20, H21, /**/ H22 );
const int HPanWidth = H11.Width() + H12.Width();
const int HPanHeight =
std::min( H11.Height(), std::max(HPanWidth-offset,0) );
const int leftover = A.Height()-HPanWidth;
HPan.LockedView( H, H00.Height(), H00.Width(), HPanHeight, HPanWidth );
ABottom.View( A, leftover, 0, HPanWidth, A.Width() );
Zeros( HPanHeight, ABottom.Width(), Z );
Zeros( HPanHeight, HPanHeight, SInv );
//--------------------------------------------------------------------//
HPanCopy = HPan;
MakeTrapezoidal( LEFT, UPPER, offset, HPanCopy );
SetDiagonalToOne( LEFT, offset, HPanCopy );
Syrk( UPPER, NORMAL, R(1), HPanCopy, R(0), SInv );
HalveMainDiagonal( SInv );
Gemm( NORMAL, NORMAL, R(1), HPanCopy, ABottom, R(0), Z );
Trsm( LEFT, UPPER, NORMAL, NON_UNIT, R(1), SInv, Z );
Gemm( TRANSPOSE, NORMAL, R(-1), HPanCopy, Z, R(1), ABottom );
//--------------------------------------------------------------------//
SlideLockedPartitionUpDiagonal
( HTL, /**/ HTR, H00, /**/ H01, H02,
/*************/ /******************/
/**/ H10, /**/ H11, H12,
HBL, /**/ HBR, H20, /**/ H21, H22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
bool NnlsBlockpivot(const DenseMatrix<T>& LHS,
const DenseMatrix<T>& RHS,
DenseMatrix<T>& X, // input as xinit
DenseMatrix<T>& Y) // gradX
{
// Solve (LHS)*X = RHS for X by block principal pivoting. Matrix LHS
// is assumed to be symmetric positive definite.
const int PBAR = 3;
const unsigned int WIDTH = RHS.Width();
const unsigned int HEIGHT = RHS.Height();
const unsigned int MAX_ITER = HEIGHT*5;
BitMatrix passive_set = (X > T(0));
std::vector<unsigned int> tmp_indices(WIDTH);
for (unsigned int i=0; i<WIDTH; ++i)
tmp_indices[i] = i;
MakeZeros(X);
if (!BppSolveNormalEqNoGroup(tmp_indices, passive_set, LHS, RHS, X))
return false;
// Y = LHS * X - RHS
Gemm(NORMAL, NORMAL, T(1), LHS, X, T(0), Y);
Axpy( T(-1), RHS, Y);
std::vector<int> P(WIDTH, PBAR), Ninf(WIDTH, HEIGHT+1);
BitMatrix nonopt_set = (Y < T(0)) & ~passive_set;
BitMatrix infeas_set = (X < T(0)) & passive_set;
std::vector<int> col_sums(WIDTH);
std::vector<int> not_good(WIDTH);
nonopt_set.SumColumns(not_good);
infeas_set.SumColumns(col_sums);
not_good += col_sums;
BitMatrix not_opt_cols = (not_good > 0);
BitMatrix not_opt_mask;
std::vector<unsigned int> non_opt_col_indices(WIDTH);
not_opt_cols.Find(non_opt_col_indices);
DenseMatrix<double> RHSsub(HEIGHT, WIDTH);
DenseMatrix<double> Xsub(HEIGHT, WIDTH);
DenseMatrix<double> Ysub(HEIGHT, WIDTH);
unsigned int iter = 0;
while (!non_opt_col_indices.empty())
{
// exit if not getting anywhere
if (iter >= MAX_ITER)
return false;
UpdatePassiveSet(passive_set, PBAR, HEIGHT,
not_opt_cols, nonopt_set, infeas_set,
not_good, P, Ninf);
// equivalent of repmat(NotOptCols, HEIGHT, 1)
not_opt_mask = MatrixFromColumnMask(not_opt_cols, HEIGHT);
// Setup for the normal equation solver by extracting submatrices
// from RHS and X. The normal equation solver will extract further
// subproblems from RHSsub and Xsub and write all updated values
// back into RHSsub and Xsub.
RHS.SubmatrixFromCols(RHSsub, non_opt_col_indices);
X.SubmatrixFromCols(Xsub, non_opt_col_indices);
if (!BppSolveNormalEqNoGroup(non_opt_col_indices, passive_set, LHS, RHSsub, Xsub))
return false;
ZeroizeSmallValues(Xsub, 1.0e-12);
// compute Ysub = LHS * Xsub - RHSsub
Ysub.Resize(RHSsub.Height(), RHSsub.Width());
Gemm(NORMAL, NORMAL, T(1), LHS, Xsub, T(0), Ysub);
Axpy( T(-1), RHSsub, Ysub);
// update Y and X using the new values in Ysub and Xsub
OverwriteCols(Y, Ysub, non_opt_col_indices, non_opt_col_indices.size());
OverwriteCols(X, Xsub, non_opt_col_indices, non_opt_col_indices.size());
ZeroizeSmallValues(X, 1.0e-12);
ZeroizeSmallValues(Y, 1.0e-12);
// Check optimality - BppUpdateSets does the equivalent of the next two lines.
// nonopt_set = not_opt_mask & (Y < T(0)) & ~passive_set;
// infeas_set = not_opt_mask & (X < T(0)) & passive_set;
BppUpdateSets(nonopt_set, infeas_set, not_opt_mask, X, Y, passive_set);
nonopt_set.SumColumns(not_good);
infeas_set.SumColumns(col_sums);
not_good += col_sums;
not_opt_cols = (not_good > 0);
not_opt_cols.Find(non_opt_col_indices);
++iter;
}
return true;
}
void TransformRows
( const Matrix<F>& V,
DistMatrix<F,MC,MR,BLOCK>& A )
{
DEBUG_CSE
const Int height = A.Height();
const Grid& grid = A.Grid();
const Int blockHeight = A.BlockHeight();
const Int firstBlockHeight = blockHeight - A.ColCut();
if( height <= firstBlockHeight || grid.Height() == 1 )
{
if( grid.Row() == A.RowOwner(0) )
{
// This process row can locally update its portion of A
TransformRows( V, A.Matrix() );
}
}
else if( height <= firstBlockHeight + blockHeight )
{
const int firstRow = A.RowOwner( 0 );
const int secondRow = A.RowOwner( firstBlockHeight );
if( grid.Row() == firstRow )
{
//
// Replace A with
//
// | VLeft, VRight |' | ATop |,
// | ABottom |
//
// where ATop is owned by this process row and ABottom by the next.
//
auto VLeft = V( ALL, IR(0,firstBlockHeight) );
// Partition space for the combined matrix
Matrix<F> ACombine( height, A.LocalWidth() );
auto ATop = ACombine( IR(0,firstBlockHeight), ALL );
auto ABottom = ACombine( IR(firstBlockHeight,END), ALL );
// Copy our portion into the combined matrix
ATop = A.LockedMatrix();
// Exchange the data
El::SendRecv( ATop, ABottom, A.ColComm(), secondRow, secondRow );
// Form our portion of the result
Gemm( ADJOINT, NORMAL, F(1), VLeft, ACombine, A.Matrix() );
}
else if( grid.Row() == secondRow )
{
//
// Replace A with
//
// | VLeft, VRight |' | ATop |,
// | ABottom |
//
// where ATop is owned by the previous process row and ABottom by
// this one.
//
auto VRight = V( ALL, IR(firstBlockHeight,END) );
// Partition space for the combined matrix
Matrix<F> ACombine( height, A.LocalWidth() );
auto ATop = ACombine( IR(0,firstBlockHeight), ALL );
auto ABottom = ACombine( IR(firstBlockHeight,END), ALL );
// Copy our portion into the combined matrix
ABottom = A.LockedMatrix();
// Exchange the data
El::SendRecv( ABottom, ATop, A.ColComm(), firstRow, firstRow );
// Form our portion of the result
Gemm( ADJOINT, NORMAL, F(1), VRight, ACombine, A.Matrix() );
}
}
else
{
// Fall back to the entire process column interacting.
// TODO(poulson): Only form the subset of the result that we need.
DistMatrix<F,STAR,MR,BLOCK> A_STAR_MR( A );
Matrix<F> ALocCopy( A_STAR_MR.Matrix() );
Gemm( ADJOINT, NORMAL, F(1), V, ALocCopy, A_STAR_MR.Matrix() );
A = A_STAR_MR;
}
}
bool NnlsHals(const MatrixType<T>& A,
DenseMatrix<T>& W,
DenseMatrix<T>& H,
const T tol,
const bool verbose,
const unsigned int max_iter)
{
unsigned int n = A.Width();
unsigned int k = W.Width();
if (static_cast<unsigned int>(W.Height()) != static_cast<unsigned int>(A.Height()))
throw std::logic_error("NnlsHals: W and A must have identical height");
if (static_cast<unsigned int>(H.Width()) != static_cast<unsigned int>(A.Width()))
throw std::logic_error("NnlsHals: H and A must have identical width");
if (H.Height() != W.Width())
throw std::logic_error("NnlsHals: non-conformant W and H");
DenseMatrix<T> WtW(k, k), WtA(k, n), WtWH_r(1, n), gradH(k, n);
if (verbose)
std::cout << "\nRunning NNLS solver..." << std::endl;
// compute W'W and W'A for the normal equations
Gemm(TRANSPOSE, NORMAL, T(1.0), W, W, T(0.0), WtW);
Gemm(TRANSPOSE, NORMAL, T(1.0), W, A, T(0.0), WtA);
bool success = false;
T pg0 = T(0), pg;
for (unsigned int i=0; i<max_iter; ++i)
{
// compute the new matrix H
UpdateH_Hals(H, WtWH_r, WtW, WtA);
// compute gradH = WtW*H - WtA
Gemm(NORMAL, NORMAL, T(1.0), WtW, H, T(0.0), gradH);
Axpy( T(-1.0), WtA, gradH);
// compute progress metric
if (0 == i)
{
pg0 = ProjectedGradientNorm(gradH, H);
if (verbose)
ReportProgress(i+1, T(1.0));
continue;
}
else
{
pg = ProjectedGradientNorm(gradH, H);
}
if (verbose)
ReportProgress(i+1, pg/pg0);
// check progress vs. desired tolerance
if (pg < tol * pg0)
{
success = true;
NormalizeAndScale<T>(W, H);
break;
}
}
if (!success)
std::cerr << "NNLS solver reached iteration limit." << std::endl;
return success;
}
inline void
TwoSidedTrsmLVar2( UnitOrNonUnit diag, Matrix<F>& A, const Matrix<F>& L )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrsmLVar2");
if( A.Height() != A.Width() )
throw std::logic_error("A must be square");
if( L.Height() != L.Width() )
throw std::logic_error("Triangular matrices must be square");
if( A.Height() != L.Height() )
throw std::logic_error("A and L must be the same size");
#endif
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<F>
LTL, LTR, L00, L01, L02,
LBL, LBR, L10, L11, L12,
L20, L21, L22;
// Temporary products
Matrix<F> X11;
Matrix<F> Y10;
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
//--------------------------------------------------------------------//
// Y10 := L10 A00
Zeros( L10.Height(), A00.Width(), Y10 );
Hemm( RIGHT, LOWER, F(1), A00, L10, F(0), Y10 );
// A10 := A10 - 1/2 Y10
Axpy( F(-1)/F(2), Y10, A10 );
// A11 := A11 - (A10 L10' + L10 A10')
Her2k( LOWER, NORMAL, F(-1), A10, L10, F(1), A11 );
// A11 := inv(L11) A11 inv(L11)'
TwoSidedTrsmLUnb( diag, A11, L11 );
// A21 := A21 - A20 L10'
Gemm( NORMAL, ADJOINT, F(-1), A20, L10, F(1), A21 );
// A21 := A21 inv(L11)'
Trsm( RIGHT, LOWER, ADJOINT, diag, F(1), L11, A21 );
// A10 := A10 - 1/2 Y10
Axpy( F(-1)/F(2), Y10, A10 );
// A10 := inv(L11) A10
Trsm( LEFT, LOWER, NORMAL, diag, F(1), L11, A10 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/**********************************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LU( Matrix<F>& A, Matrix<int>& p )
{
#ifndef RELEASE
PushCallStack("LU");
if( p.Viewing() &&
(std::min(A.Height(),A.Width()) != p.Height() || p.Width() != 1) )
throw std::logic_error
("p must be a vector of the same height as the min dimension of A.");
#endif
if( !p.Viewing() )
p.ResizeTo( std::min(A.Height(),A.Width()), 1 );
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02, ABRL, ABRR,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<int>
pT, p0,
pB, p1,
p2;
// Pivot composition
std::vector<int> image, preimage;
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( p, pT,
pB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( pT, p0,
/**/ /**/
p1,
pB, p2 );
PartitionRight( ABR, ABRL, ABRR, A11.Width() );
const int pivotOffset = A01.Height();
//--------------------------------------------------------------------//
internal::PanelLU( ABRL, p1, pivotOffset );
internal::ComposePanelPivots( p1, pivotOffset, image, preimage );
ApplyRowPivots( ABL, image, preimage );
ApplyRowPivots( ABRR, image, preimage );
Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 );
Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlidePartitionDown
( pT, p0,
p1,
/**/ /**/
pB, p2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LU( Matrix<F>& A, Matrix<Int>& p )
{
#ifndef RELEASE
CallStackEntry entry("LU");
#endif
p.ResizeTo( Min(A.Height(),A.Width()), 1 );
// Matrix views
Matrix<F>
ATL, ATR, A00, A01, A02, ABRL, ABRR,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
Matrix<Int>
pT, p0,
pB, p1,
p2;
// Pivot composition
std::vector<Int> image, preimage;
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( p, pT,
pB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
RepartitionDown
( pT, p0,
/**/ /**/
p1,
pB, p2 );
PartitionRight( ABR, ABRL, ABRR, A11.Width() );
const Int pivotOffset = A01.Height();
//--------------------------------------------------------------------//
lu::Panel( ABRL, p1, pivotOffset );
ComposePivots( p1, pivotOffset, image, preimage );
ApplyRowPivots( ABL, image, preimage );
ApplyRowPivots( ABRR, image, preimage );
Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 );
Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlidePartitionDown
( pT, p0,
p1,
/**/ /**/
pB, p2 );
}
}
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;
}
