inline void
TwoSidedTrsmUVar1
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("internal::TwoSidedTrsmUVar1");
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,STAR,STAR> A11_STAR_STAR(g);
DistMatrix<F,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,VC, STAR> U01_VC_STAR(g);
DistMatrix<F,VR, STAR> U01_VR_STAR(g);
DistMatrix<F,STAR,MR > U01Adj_STAR_MR(g);
DistMatrix<F,STAR,STAR> X11_STAR_STAR(g);
DistMatrix<F,MR, MC > Z01_MR_MC(g);
DistMatrix<F,MC, STAR> Z01_MC_STAR(g);
DistMatrix<F,MR, STAR> Z01_MR_STAR(g);
DistMatrix<F> Y01(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( A01 );
U01_MC_STAR.AlignWith( A00 );
U01_VR_STAR.AlignWith( A00 );
U01_VC_STAR.AlignWith( A00 );
U01Adj_STAR_MR.AlignWith( A00 );
Y01.AlignWith( A01 );
Z01_MR_MC.AlignWith( A01 );
Z01_MC_STAR.AlignWith( A00 );
Z01_MR_STAR.AlignWith( A00 );
//--------------------------------------------------------------------//
// Y01 := A00 U01
U01_MC_STAR = U01;
U01_VR_STAR = U01_MC_STAR;
U01Adj_STAR_MR.AdjointFrom( U01_VR_STAR );
Zeros( Z01_MC_STAR, A01.Height(), A01.Width() );
Zeros( Z01_MR_STAR, A01.Height(), A01.Width() );
LocalSymmetricAccumulateLU
( ADJOINT,
F(1), A00, U01_MC_STAR, U01Adj_STAR_MR, Z01_MC_STAR, Z01_MR_STAR );
Z01_MR_MC.SumScatterFrom( Z01_MR_STAR );
Y01 = Z01_MR_MC;
Y01.SumScatterUpdate( F(1), Z01_MC_STAR );
// A01 := inv(U00)' A01
//
// This is the bottleneck because A01 only has blocksize columns
Trsm( LEFT, UPPER, ADJOINT, diag, F(1), U00, A01 );
// A01 := A01 - 1/2 Y01
Axpy( F(-1)/F(2), Y01, A01 );
// A11 := A11 - (U01' A01 + A01' U01)
A01_VC_STAR = A01;
U01_VC_STAR = U01_MC_STAR;
Zeros( X11_STAR_STAR, A11.Height(), A11.Width() );
Her2k
( UPPER, ADJOINT,
F(-1), A01_VC_STAR.Matrix(), U01_VC_STAR.Matrix(),
F(0), X11_STAR_STAR.Matrix() );
A11.SumScatterUpdate( F(1), X11_STAR_STAR );
// A11 := inv(U11)' A11 inv(U11)
A11_STAR_STAR = A11;
U11_STAR_STAR = U11;
LocalTwoSidedTrsm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
// A01 := A01 - 1/2 Y01
Axpy( F(-1)/F(2), Y01, A01 );
// A01 := A01 inv(U11)
A01_VC_STAR = A01;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, A01_VC_STAR );
A01 = A01_VC_STAR;
//--------------------------------------------------------------------//
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
TwoSidedTrmmUVar5
( UnitOrNonUnit diag, DistMatrix<F>& A, const DistMatrix<F>& U )
{
#ifndef RELEASE
PushCallStack("internal::TwoSidedTrmmUVar5");
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,MC, STAR> A01_MC_STAR(g);
DistMatrix<F,MR, STAR> A01_MR_STAR(g);
DistMatrix<F,VC, STAR> A01_VC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,MC, STAR> U01_MC_STAR(g);
DistMatrix<F,MR, STAR> U01_MR_STAR(g);
DistMatrix<F,VC, STAR> U01_VC_STAR(g);
DistMatrix<F,VC, STAR> Y01_VC_STAR(g);
DistMatrix<F> Y01(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_MC_STAR.AlignWith( A00 );
A01_MR_STAR.AlignWith( A00 );
A01_VC_STAR.AlignWith( A00 );
U01_MC_STAR.AlignWith( A00 );
U01_MR_STAR.AlignWith( A00 );
U01_VC_STAR.AlignWith( A00 );
Y01.AlignWith( A01 );
Y01_VC_STAR.AlignWith( A01 );
//--------------------------------------------------------------------//
// Y01 := U01 A11
A11_STAR_STAR = A11;
U01_VC_STAR = U01;
Y01_VC_STAR.ResizeTo( A01.Height(), A01.Width() );
Hemm
( RIGHT, UPPER,
F(1), A11_STAR_STAR.LocalMatrix(), U01_VC_STAR.LocalMatrix(),
F(0), Y01_VC_STAR.LocalMatrix() );
Y01 = Y01_VC_STAR;
// A01 := U00 A01
Trmm( LEFT, UPPER, NORMAL, diag, F(1), U00, A01 );
// A01 := A01 + 1/2 Y01
Axpy( F(1)/F(2), Y01, A01 );
// A00 := A00 + (U01 A01' + A01 U01')
A01_MC_STAR = A01;
U01_MC_STAR = U01;
A01_VC_STAR = A01_MC_STAR;
A01_MR_STAR = A01_VC_STAR;
U01_MR_STAR = U01_MC_STAR;
LocalTrr2k
( UPPER, ADJOINT, ADJOINT,
F(1), U01_MC_STAR, A01_MR_STAR,
A01_MC_STAR, U01_MR_STAR,
F(1), A00 );
// A01 := A01 + 1/2 Y01
Axpy( F(1)/F(2), Y01_VC_STAR, A01_VC_STAR );
// A01 := A01 U11'
U11_STAR_STAR = U11;
LocalTrmm
( RIGHT, UPPER, ADJOINT, diag, F(1), U11_STAR_STAR, A01_VC_STAR );
A01 = A01_VC_STAR;
// A11 := U11 A11 U11'
LocalTwoSidedTrmm( UPPER, diag, A11_STAR_STAR, U11_STAR_STAR );
A11 = A11_STAR_STAR;
//--------------------------------------------------------------------//
A01_MC_STAR.FreeAlignments();
A01_MR_STAR.FreeAlignments();
A01_VC_STAR.FreeAlignments();
U01_MC_STAR.FreeAlignments();
U01_MR_STAR.FreeAlignments();
U01_VC_STAR.FreeAlignments();
Y01.FreeAlignments();
Y01_VC_STAR.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
UVar3( UnitOrNonUnit diag, DistMatrix<F>& U )
{
#ifndef RELEASE
CallStackEntry entry("triangular_inverse::UVar3");
if( U.Height() != U.Width() )
LogicError("Nonsquare matrices cannot be triangular");
#endif
const Grid& g = U.Grid();
// Matrix views
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> U01_VC_STAR(g);
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,VR > U12_STAR_VR(g);
DistMatrix<F,STAR,MC > U01Trans_STAR_MC(g);
DistMatrix<F,MR, STAR> U12Trans_MR_STAR(g);
// Start the algorithm
PartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( UBR.Height() < U.Height() )
{
RepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
U01Trans_STAR_MC.AlignWith( U02 );
U12Trans_MR_STAR.AlignWith( U02 );
//--------------------------------------------------------------------//
U01_VC_STAR = U01;
U11_STAR_STAR = U11;
LocalTrsm
( RIGHT, UPPER, NORMAL, diag, F(-1), U11_STAR_STAR, U01_VC_STAR );
// We transpose before the communication to avoid cache-thrashing
// in the unpacking stage.
U12Trans_MR_STAR.TransposeFrom( U12 );
U01Trans_STAR_MC.TransposeFrom( U01_VC_STAR );
LocalGemm
( TRANSPOSE, TRANSPOSE,
F(1), U01Trans_STAR_MC, U12Trans_MR_STAR, F(1), U02 );
U01.TransposeFrom( U01Trans_STAR_MC );
U12_STAR_VR.TransposeFrom( U12Trans_MR_STAR );
LocalTrsm
( LEFT, UPPER, NORMAL, diag, F(1), U11_STAR_STAR, U12_STAR_VR );
LocalTriangularInverse( UPPER, diag, U11_STAR_STAR );
U11 = U11_STAR_STAR;
U12 = U12_STAR_VR;
//--------------------------------------------------------------------//
SlidePartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
}
}
inline void
TrtrmmUVar1( Orientation orientation, DistMatrix<T>& U )
{
#ifndef RELEASE
PushCallStack("internal::TrtrmmUVar1");
if( U.Height() != U.Width() )
throw std::logic_error("U must be square");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T>
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<T,MC, STAR> U01_MC_STAR(g);
DistMatrix<T,VC, STAR> U01_VC_STAR(g);
DistMatrix<T,VR, STAR> U01_VR_STAR(g);
DistMatrix<T,STAR,MR > U01AdjOrTrans_STAR_MR(g);
DistMatrix<T,STAR,STAR> U11_STAR_STAR(g);
U01_MC_STAR.AlignWith( U );
U01_VC_STAR.AlignWith( U );
U01_VR_STAR.AlignWith( U );
U01AdjOrTrans_STAR_MR.AlignWith( U );
PartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
while( UTL.Height() < U.Height() && UTL.Width() < U.Height() )
{
RepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
//--------------------------------------------------------------------//
U01_MC_STAR = U01;
U01_VC_STAR = U01_MC_STAR;
U01_VR_STAR = U01_VC_STAR;
if( orientation == ADJOINT )
U01AdjOrTrans_STAR_MR.AdjointFrom( U01_VR_STAR );
else
U01AdjOrTrans_STAR_MR.TransposeFrom( U01_VR_STAR );
LocalTrrk( UPPER, T(1), U01_MC_STAR, U01AdjOrTrans_STAR_MR, T(1), U00 );
U11_STAR_STAR = U11;
LocalTrmm
( RIGHT, UPPER, orientation, NON_UNIT,
T(1), U11_STAR_STAR, U01_VC_STAR );
U01 = U01_VC_STAR;
LocalTrtrmm( orientation, UPPER, U11_STAR_STAR );
U11 = U11_STAR_STAR;
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
