inline void
CholeskyUVar3( DistMatrix<F>& A )
{
#ifndef RELEASE
PushCallStack("internal::CholeskyUVar3");
if( A.Height() != A.Width() )
throw std::logic_error
("Can only compute Cholesky factor of square matrices");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
// Temporary matrix distributions
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,STAR,MC > A12_STAR_MC(g);
DistMatrix<F,STAR,MR > A12_STAR_MR(g);
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ABR.Height() > 0 )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
A12_STAR_MC.AlignWith( A22 );
A12_STAR_MR.AlignWith( A22 );
A12_STAR_VR.AlignWith( A22 );
//--------------------------------------------------------------------//
A11_STAR_STAR = A11;
LocalCholesky( UPPER, A11_STAR_STAR );
A11 = A11_STAR_STAR;
A12_STAR_VR = A12;
LocalTrsm
( LEFT, UPPER, ADJOINT, NON_UNIT, F(1), A11_STAR_STAR, A12_STAR_VR );
A12_STAR_MC = A12_STAR_VR;
A12_STAR_MR = A12_STAR_VR;
LocalTrrk
( UPPER, ADJOINT, F(-1), A12_STAR_MC, A12_STAR_MR, F(1), A22 );
A12 = A12_STAR_MR;
//--------------------------------------------------------------------//
A12_STAR_MC.FreeAlignments();
A12_STAR_MR.FreeAlignments();
A12_STAR_VR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
CholeskyUVar2( DistMatrix<F>& A )
{
#ifndef RELEASE
CallStackEntry entry("hpd_inverse::CholeskyUVar2");
if( A.Height() != A.Width() )
throw std::logic_error("Nonsquare matrices cannot be triangular");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,VR, STAR> A01_VR_STAR(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,STAR,MC > A01Trans_STAR_MC(g);
DistMatrix<F,MR, STAR> A01_MR_STAR(g);
DistMatrix<F,STAR,MR > A01Adj_STAR_MR(g);
DistMatrix<F,STAR,MR > A12_STAR_MR(g);
DistMatrix<F,STAR,MC > A12_STAR_MC(g);
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
A01_VC_STAR.AlignWith( A00 );
A12_STAR_VR.AlignWith( A02 );
A01Trans_STAR_MC.AlignWith( A00 );
A01_VR_STAR.AlignWith( A00 );
A01Adj_STAR_MR.AlignWith( A00 );
A12_STAR_MR.AlignWith( A02 );
A12_STAR_MC.AlignWith( A22 );
//--------------------------------------------------------------------//
A11_STAR_STAR = A11;
LocalCholesky( UPPER, A11_STAR_STAR );
A01_VC_STAR = A01;
LocalTrsm
( RIGHT, UPPER, NORMAL, NON_UNIT, F(1), A11_STAR_STAR, A01_VC_STAR );
A12_STAR_VR = A12;
LocalTrsm
( LEFT, UPPER, ADJOINT, NON_UNIT, F(1), A11_STAR_STAR, A12_STAR_VR );
A01Trans_STAR_MC.TransposeFrom( A01_VC_STAR );
A01_VR_STAR = A01_VC_STAR;
A01Adj_STAR_MR.AdjointFrom( A01_VR_STAR );
LocalTrrk
( UPPER, TRANSPOSE,
F(1), A01Trans_STAR_MC, A01Adj_STAR_MR, F(1), A00 );
A12_STAR_MR = A12_STAR_VR;
LocalGemm
( TRANSPOSE, NORMAL, F(-1), A01Trans_STAR_MC, A12_STAR_MR, F(1), A02 );
A12_STAR_MC = A12_STAR_VR;
LocalTrrk
( UPPER, ADJOINT,
F(-1), A12_STAR_MC, A12_STAR_MR, F(1), A22 );
LocalTrsm
( RIGHT, UPPER, ADJOINT, NON_UNIT, F(1), A11_STAR_STAR, A01_VC_STAR );
LocalTrsm
( LEFT, UPPER, NORMAL, NON_UNIT, F(-1), A11_STAR_STAR, A12_STAR_VR );
LocalTriangularInverse( UPPER, NON_UNIT, A11_STAR_STAR );
LocalTrtrmm( ADJOINT, UPPER, A11_STAR_STAR );
A11 = A11_STAR_STAR;
A01 = A01_VC_STAR;
A12 = A12_STAR_VR;
//--------------------------------------------------------------------//
A01_VC_STAR.FreeAlignments();
A12_STAR_VR.FreeAlignments();
A01Trans_STAR_MC.FreeAlignments();
A01_VR_STAR.FreeAlignments();
A01Adj_STAR_MR.FreeAlignments();
A12_STAR_MR.FreeAlignments();
A12_STAR_MC.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
}
inline void
TwoSidedTrsmUVar4
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar4");
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
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
// Temporary distributions
DistMatrix<F,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,MC > A01Trans_STAR_MC(g);
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,STAR,VC > A12_STAR_VC(g);
DistMatrix<F,STAR,MC > A12_STAR_MC(g);
DistMatrix<F,STAR,MR > A12_STAR_MR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MR, STAR> U12Trans_MR_STAR(g);
DistMatrix<F,VR, STAR> U12Trans_VR_STAR(g);
DistMatrix<F,STAR,VR > U12_STAR_VR(g);
DistMatrix<F,STAR,VC > U12_STAR_VC(g);
DistMatrix<F,STAR,MC > U12_STAR_MC(g);
DistMatrix<F,STAR,VR > Y12_STAR_VR(g);
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_VC_STAR.AlignWith( A02 );
A01Trans_STAR_MC.AlignWith( A02 );
A12_STAR_VR.AlignWith( A22 );
A12_STAR_VC.AlignWith( A22 );
A12_STAR_MC.AlignWith( A22 );
A12_STAR_MR.AlignWith( A22 );
U12Trans_MR_STAR.AlignWith( A02 );
U12Trans_VR_STAR.AlignWith( A02 );
U12_STAR_VR.AlignWith( A02 );
U12_STAR_VC.AlignWith( A22 );
U12_STAR_MC.AlignWith( A22 );
Y12_STAR_VR.AlignWith( A12 );
//--------------------------------------------------------------------//
// A01 := A01 inv(U11)
A01_VC_STAR = A01;
U11_STAR_STAR = U11;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, A01_VC_STAR );
A01 = A01_VC_STAR;
// A11 := inv(U11)' A11 inv(U11)
A11_STAR_STAR = A11;
LocalTwoSidedTrsm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
// A02 := A02 - A01 U12
A01Trans_STAR_MC.TransposeFrom( A01_VC_STAR );
U12Trans_MR_STAR.TransposeFrom( U12 );
LocalGemm
( TRANSPOSE, TRANSPOSE,
F(-1), A01Trans_STAR_MC, U12Trans_MR_STAR, F(1), A02 );
// Y12 := A11 U12
U12Trans_VR_STAR = U12Trans_MR_STAR;
Zeros( U12_STAR_VR, A12.Height(), A12.Width() );
Transpose( U12Trans_VR_STAR.Matrix(), U12_STAR_VR.Matrix() );
Zeros( Y12_STAR_VR, A12.Height(), A12.Width() );
Hemm
( LEFT, UPPER,
F(1), A11_STAR_STAR.Matrix(), U12_STAR_VR.Matrix(),
F(0), Y12_STAR_VR.Matrix() );
// A12 := inv(U11)' A12
A12_STAR_VR = A12;
LocalTrsm
( LEFT, UPPER, ADJOINT, diag, F(1), U11_STAR_STAR, A12_STAR_VR );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12_STAR_VR, A12_STAR_VR );
// A22 := A22 - (A12' U12 + U12' A12)
A12_STAR_MR = A12_STAR_VR;
A12_STAR_VC = A12_STAR_VR;
U12_STAR_VC = U12_STAR_VR;
A12_STAR_MC = A12_STAR_VC;
U12_STAR_MC = U12_STAR_VC;
LocalTrr2k
( UPPER, ADJOINT, TRANSPOSE, ADJOINT,
F(-1), A12_STAR_MC, U12Trans_MR_STAR,
U12_STAR_MC, A12_STAR_MR,
F(1), A22 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12_STAR_VR, A12_STAR_VR );
A12 = A12_STAR_VR;
//--------------------------------------------------------------------//
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
TwoSidedTrsmUVar5
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrsmUVar5");
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
const Grid& g = A.Grid();
// Matrix views
DistMatrix<F>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<F>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,MC > A12_STAR_MC(g);
DistMatrix<F,STAR,MR > A12_STAR_MR(g);
DistMatrix<F,STAR,VC > A12_STAR_VC(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,MC > U12_STAR_MC(g);
DistMatrix<F,STAR,MR > U12_STAR_MR(g);
DistMatrix<F,STAR,VC > U12_STAR_VC(g);
DistMatrix<F,STAR,VR > U12_STAR_VR(g);
DistMatrix<F,STAR,VR > Y12_STAR_VR(g);
DistMatrix<F> Y12(g);
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 );
A12_STAR_MC.AlignWith( A22 );
A12_STAR_MR.AlignWith( A22 );
A12_STAR_VC.AlignWith( A22 );
A12_STAR_VR.AlignWith( A22 );
U12_STAR_MC.AlignWith( A22 );
U12_STAR_MR.AlignWith( A22 );
U12_STAR_VC.AlignWith( A22 );
U12_STAR_VR.AlignWith( A22 );
Y12.AlignWith( A12 );
Y12_STAR_VR.AlignWith( A12 );
//--------------------------------------------------------------------//
// A11 := inv(U11)' A11 inv(U11)
U11_STAR_STAR = U11;
A11_STAR_STAR = A11;
LocalTwoSidedTrsm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
// Y12 := A11 U12
U12_STAR_VR = U12;
Y12_STAR_VR.ResizeTo( A12.Height(), A12.Width() );
Hemm
( LEFT, UPPER,
F(1), A11_STAR_STAR.LocalMatrix(), U12_STAR_VR.LocalMatrix(),
F(0), Y12_STAR_VR.LocalMatrix() );
Y12 = Y12_STAR_VR;
// A12 := inv(U11)' A12
A12_STAR_VR = A12;
LocalTrsm
( LEFT, UPPER, ADJOINT, diag, F(1), U11_STAR_STAR, A12_STAR_VR );
A12 = A12_STAR_VR;
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
// A22 := A22 - (A12' U12 + U12' A12)
A12_STAR_VR = A12;
A12_STAR_VC = A12_STAR_VR;
U12_STAR_VC = U12_STAR_VR;
A12_STAR_MC = A12_STAR_VC;
U12_STAR_MC = U12_STAR_VC;
A12_STAR_MR = A12_STAR_VR;
U12_STAR_MR = U12_STAR_VR;
LocalTrr2k
( UPPER, ADJOINT, ADJOINT,
F(-1), U12_STAR_MC, A12_STAR_MR,
A12_STAR_MC, U12_STAR_MR,
F(1), A22 );
// A12 := A12 - 1/2 Y12
Axpy( F(-1)/F(2), Y12, A12 );
// A12 := A12 inv(U22)
//
// This is the bottleneck because A12 only has blocksize rows
Trsm( RIGHT, UPPER, NORMAL, diag, F(1), U22, A12 );
//--------------------------------------------------------------------//
A12_STAR_MC.FreeAlignments();
A12_STAR_MR.FreeAlignments();
A12_STAR_VC.FreeAlignments();
A12_STAR_VR.FreeAlignments();
U12_STAR_MC.FreeAlignments();
U12_STAR_MR.FreeAlignments();
U12_STAR_VC.FreeAlignments();
U12_STAR_VR.FreeAlignments();
Y12.FreeAlignments();
Y12_STAR_VR.FreeAlignments();
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
internal::CholeskyUVar3Square( DistMatrix<F,MC,MR>& A )
{
#ifndef RELEASE
PushCallStack("internal::CholeskyUVar3Square");
if( A.Height() != A.Width() )
throw std::logic_error
("Can only compute Cholesky factor of square matrices.");
if( A.Grid().Height() != A.Grid().Width() )
throw std::logic_error
("CholeskyUVar3Square assumes a square process grid.");
#endif
const Grid& g = A.Grid();
// Find the process holding our transposed data
const int r = g.Height();
int transposeRank;
{
const int colAlignment = A.ColAlignment();
const int rowAlignment = A.RowAlignment();
const int colShift = A.ColShift();
const int rowShift = A.RowShift();
const int transposeRow = (colAlignment+rowShift) % r;
const int transposeCol = (rowAlignment+colShift) % r;
transposeRank = transposeRow + r*transposeCol;
}
const bool onDiagonal = ( transposeRank == g.VCRank() );
// Matrix views
DistMatrix<F,MC,MR>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
// Temporary matrix distributions
DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,STAR,VR > A12_STAR_VR(g);
DistMatrix<F,STAR,MC > A12_STAR_MC(g);
DistMatrix<F,STAR,MR > A12_STAR_MR(g);
// Start the algorithm
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ABR.Height() > 0 )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
A12_STAR_MC.AlignWith( A22 );
A12_STAR_MR.AlignWith( A22 );
A12_STAR_VR.AlignWith( A22 );
//--------------------------------------------------------------------//
A11_STAR_STAR = A11;
internal::LocalCholesky( UPPER, A11_STAR_STAR );
A11 = A11_STAR_STAR;
A12_STAR_VR = A12;
internal::LocalTrsm
( LEFT, UPPER, ADJOINT, NON_UNIT, (F)1, A11_STAR_STAR, A12_STAR_VR );
A12_STAR_MR = A12_STAR_VR;
// SendRecv to form A12[* ,MC] from A12[* ,MR]
A12_STAR_MC.ResizeTo( A12.Height(), A12.Width() );
{
if( onDiagonal )
{
const int size = A11.Height()*A22.LocalWidth();
MemCopy
( A12_STAR_MC.LocalBuffer(),
A12_STAR_MR.LocalBuffer(), size );
}
else
{
const int sendSize = A11.Height()*A22.LocalWidth();
const int recvSize = A11.Width()*A22.LocalHeight();
// We know that the ldim is the height since we have manually
// created both temporary matrices.
mpi::SendRecv
( A12_STAR_MR.LocalBuffer(), sendSize, transposeRank, 0,
A12_STAR_MC.LocalBuffer(), recvSize, transposeRank, 0,
g.VCComm() );
}
}
internal::LocalTrrk
( UPPER, ADJOINT, (F)-1, A12_STAR_MC, A12_STAR_MR, (F)1, A22 );
A12 = A12_STAR_MR;
//--------------------------------------------------------------------//
A12_STAR_MC.FreeAlignments();
A12_STAR_MR.FreeAlignments();
A12_STAR_VR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}