Пример #1
0
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 );
    }
}
Пример #2
0
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 );
    }
}
Пример #3
0
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
}
Пример #4
0
inline void
internal::HegstLUVar2( DistMatrix<F,MC,MR>& A, const DistMatrix<F,MC,MR>& U )
{
#ifndef RELEASE
    PushCallStack("internal::HegstLUVar2");
    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,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);

    DistMatrix<F,MC,MR>
        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,STAR> A11_STAR_STAR(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,VR  > U12_STAR_VR(g);
    DistMatrix<F,MR,  STAR> U12Adj_MR_STAR(g);
    DistMatrix<F,VC,  STAR> U12Adj_VC_STAR(g);
    DistMatrix<F,MC,  STAR> X01_MC_STAR(g);
    DistMatrix<F,STAR,STAR> X11_STAR_STAR(g);
    DistMatrix<F,MC,  MR  > Y12(g);
    DistMatrix<F,MC,  MR  > Z12Adj(g);
    DistMatrix<F,MR,  MC  > Z12Adj_MR_MC(g);
    DistMatrix<F,MC,  STAR> Z12Adj_MC_STAR(g);
    DistMatrix<F,MR,  STAR> Z12Adj_MR_STAR(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_VR.AlignWith( A12 );
        U12_STAR_MC.AlignWith( A22 );
        U12_STAR_VR.AlignWith( A12 );
        U12Adj_MR_STAR.AlignWith( A22 );
        U12Adj_VC_STAR.AlignWith( A22 );
        X01_MC_STAR.AlignWith( A01 );
        Y12.AlignWith( A12 );
        Z12Adj.AlignWith( A12 );
        Z12Adj_MR_MC.AlignWith( A12 );
        Z12Adj_MC_STAR.AlignWith( A22 );
        Z12Adj_MR_STAR.AlignWith( A22 );
        //--------------------------------------------------------------------//
        // A01 := A01 U11'
        U11_STAR_STAR = U11;
        A01_VC_STAR = A01;
        internal::LocalTrmm
        ( RIGHT, UPPER, ADJOINT, NON_UNIT, (F)1, U11_STAR_STAR, A01_VC_STAR );
        A01 = A01_VC_STAR;

        // A01 := A01 + A02 U12'
        U12Adj_MR_STAR.AdjointFrom( U12 );
        X01_MC_STAR.ResizeTo( A01.Height(), A01.Width() );
        internal::LocalGemm
        ( NORMAL, NORMAL,
          (F)1, A02, U12Adj_MR_STAR, (F)0, X01_MC_STAR );
        A01.SumScatterUpdate( (F)1, X01_MC_STAR );

        // Y12 := U12 A22
        U12Adj_VC_STAR = U12Adj_MR_STAR;
        U12_STAR_MC.AdjointFrom( U12Adj_VC_STAR );
        Z12Adj_MC_STAR.ResizeTo( A12.Width(), A12.Height() );
        Z12Adj_MR_STAR.ResizeTo( A12.Width(), A12.Height() );
        Zero( Z12Adj_MC_STAR );
        Zero( Z12Adj_MR_STAR );
        internal::LocalSymmetricAccumulateRU
        ( ADJOINT, 
          (F)1, A22, U12_STAR_MC, U12Adj_MR_STAR, 
          Z12Adj_MC_STAR, Z12Adj_MR_STAR );
        Z12Adj.SumScatterFrom( Z12Adj_MC_STAR );
        Z12Adj_MR_MC = Z12Adj;
        Z12Adj_MR_MC.SumScatterUpdate( (F)1, Z12Adj_MR_STAR );
        Y12.ResizeTo( A12.Height(), A12.Width() );
        Adjoint( Z12Adj_MR_MC.LockedLocalMatrix(), Y12.LocalMatrix() );

        // A12 := U11 A12
        A12_STAR_VR = A12;
        U11_STAR_STAR = U11;
        internal::LocalTrmm
        ( LEFT, UPPER, NORMAL, NON_UNIT, (F)1, U11_STAR_STAR, A12_STAR_VR );
        A12 = A12_STAR_VR;

        // A12 := A12 + 1/2 Y12
        Axpy( (F)0.5, Y12, A12 );

        // A11 := U11 A11 U11'
        A11_STAR_STAR = A11;
        internal::LocalHegst( LEFT, UPPER, A11_STAR_STAR, U11_STAR_STAR );
        A11 = A11_STAR_STAR;

        // A11 := A11 + (A12 U12' + U12 A12')
        A12_STAR_VR = A12;
        U12_STAR_VR = U12;
        X11_STAR_STAR.ResizeTo( A11.Height(), A11.Width() );
        Her2k
        ( UPPER, NORMAL,
          (F)1, A12_STAR_VR.LocalMatrix(), U12_STAR_VR.LocalMatrix(),
          (F)0, X11_STAR_STAR.LocalMatrix() );
        A11.SumScatterUpdate( (F)1, X11_STAR_STAR );

        // A12 := A12 + 1/2 Y12
        Axpy( (F)0.5, Y12, A12 );
        //--------------------------------------------------------------------//
        A12_STAR_VR.FreeAlignments();
        U12_STAR_MC.FreeAlignments();
        U12_STAR_VR.FreeAlignments();
        U12Adj_MR_STAR.FreeAlignments();
        U12Adj_VC_STAR.FreeAlignments();
        X01_MC_STAR.FreeAlignments();
        Y12.FreeAlignments();
        Z12Adj.FreeAlignments();
        Z12Adj_MR_MC.FreeAlignments(); 
        Z12Adj_MC_STAR.FreeAlignments();
        Z12Adj_MR_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
}