예제 #1
0
파일: LN.hpp 프로젝트: mcg1969/Elemental
inline void
SyrkLN
( T alpha, const DistMatrix<T>& A, T beta, DistMatrix<T>& C, 
  bool conjugate=false )
{
#ifndef RELEASE
    PushCallStack("internal::SyrkLN");
    if( A.Grid() != C.Grid() )
        throw std::logic_error
        ("A and C must be distributed over the same grid");
    if( A.Height() != C.Height() || A.Height() != C.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal SyrkLN:\n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T> AL(g), AR(g),
                  A0(g), A1(g), A2(g);

    // Temporary distributions
    DistMatrix<T,MC,  STAR> A1_MC_STAR(g);
    DistMatrix<T,VR,  STAR> A1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > A1Trans_STAR_MR(g);

    A1_MC_STAR.AlignWith( C );
    A1_VR_STAR.AlignWith( C );
    A1Trans_STAR_MR.AlignWith( C );

    // Start the algorithm
    ScaleTrapezoid( beta, LEFT, LOWER, 0, C );
    LockedPartitionRight( A, AL, AR, 0 );
    while( AR.Width() > 0 )
    {
        LockedRepartitionRight
        ( AL, /**/ AR,
          A0, /**/ A1, A2 );

        //--------------------------------------------------------------------//
        A1_VR_STAR = A1_MC_STAR = A1;
        A1Trans_STAR_MR.TransposeFrom( A1_VR_STAR, conjugate );
        LocalTrrk( LOWER, alpha, A1_MC_STAR, A1Trans_STAR_MR, T(1), C );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #2
0
inline void
HermitianFromEVD
( UpperOrLower uplo,
        Matrix<F>& A,
  const Matrix<BASE(F)>& w,
  const Matrix<F>& Z )
{
#ifndef RELEASE
    CallStackEntry entry("HermitianFromEVD");
#endif
    typedef BASE(F) R;

    Matrix<F> ZL, ZR,
              Z0, Z1, Z2;
    Matrix<R> wT,  w0,
              wB,  w1,
                   w2;

    Matrix<F> Z1Copy, Y1;

    A.ResizeTo( Z.Height(), Z.Height() );
    if( uplo == LOWER )
        MakeTrapezoidal( UPPER, A, 1 );
    else
        MakeTrapezoidal( LOWER, A, -1 );
    LockedPartitionRight( Z, ZL, ZR, 0 );
    LockedPartitionDown
    ( w, wT,
         wB, 0 );
    while( ZL.Width() < Z.Width() )
    {
        LockedRepartitionRight
        ( ZL, /**/ ZR,
          Z0, /**/ Z1, Z2 );
        LockedRepartitionDown
        ( wT,  w0,
         /**/ /**/
               w1,
          wB,  w2 );

        //--------------------------------------------------------------------//
        Y1 = Z1Copy = Z1;
        DiagonalScale( RIGHT, NORMAL, w1, Y1 );
        Trrk( uplo, NORMAL, ADJOINT, F(1), Z1Copy, Y1, F(1), A );
        //--------------------------------------------------------------------//

        SlideLockedPartitionDown
        ( wT,  w0,
               w1,
         /**/ /**/
          wB,  w2 );
        SlideLockedPartitionRight
        ( ZL,     /**/ ZR,
          Z0, Z1, /**/ Z2 );
    }
}
예제 #3
0
파일: RLN.hpp 프로젝트: certik/Elemental
inline void
LocalTrmmAccumulateRLN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
  const DistMatrix<T,MC,  MR  >& L,
  const DistMatrix<T,STAR,MC  >& X_STAR_MC,
        DistMatrix<T,MR,  STAR>& ZAdjOrTrans_MR_STAR )
{
#ifndef RELEASE
    PushCallStack("internal::LocalTrmmAccumulateRLN");
    if( L.Grid() != X_STAR_MC.Grid() || 
        X_STAR_MC.Grid() != ZAdjOrTrans_MR_STAR.Grid() )
        throw std::logic_error
        ("{L,X,Z} must be distributed over the same grid");
    if( L.Height() != L.Width() ||
        L.Height() != X_STAR_MC.Width() ||
        L.Height() != ZAdjOrTrans_MR_STAR.Height() )
    {
        std::ostringstream msg;
        msg << "Nonconformal LocalTrmmAccumulateRLN: \n"
            << "  L ~ " << L.Height() << " x " << L.Width() << "\n"
            << "  X[* ,MC] ~ " << X_STAR_MC.Height() << " x "
                               << X_STAR_MC.Width() << "\n"
            << "  Z^H/T[MR,* ] ~ " << ZAdjOrTrans_MR_STAR.Height() << " x "
                                   << ZAdjOrTrans_MR_STAR.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
    if( X_STAR_MC.RowAlignment() != L.ColAlignment() ||
        ZAdjOrTrans_MR_STAR.ColAlignment() != L.RowAlignment() )
        throw std::logic_error("Partial matrix distributions are misaligned");
#endif
    const Grid& g = L.Grid();

    // Matrix views
    DistMatrix<T>
        LTL(g), LTR(g),  L00(g), L01(g), L02(g),
        LBL(g), LBR(g),  L10(g), L11(g), L12(g),
                         L20(g), L21(g), L22(g);

    DistMatrix<T> D11(g);

    DistMatrix<T,STAR,MC>
        XL_STAR_MC(g), XR_STAR_MC(g),
        X0_STAR_MC(g), X1_STAR_MC(g), X2_STAR_MC(g);

    DistMatrix<T,MR,STAR>
        ZTAdjOrTrans_MR_STAR(g),  Z0AdjOrTrans_MR_STAR(g),
        ZBAdjOrTrans_MR_STAR(g),  Z1AdjOrTrans_MR_STAR(g),
                                   Z2AdjOrTrans_MR_STAR(g);

    const int ratio = std::max( g.Height(), g.Width() );
    PushBlocksizeStack( ratio*Blocksize() );

    LockedPartitionDownDiagonal
    ( L, LTL, LTR,
         LBL, LBR, 0 );
    LockedPartitionRight( X_STAR_MC,  XL_STAR_MC, XR_STAR_MC, 0 );
    PartitionDown
    ( ZAdjOrTrans_MR_STAR, ZTAdjOrTrans_MR_STAR,
                            ZBAdjOrTrans_MR_STAR, 0 );
    while( LTL.Height() < L.Height() )
    {
        LockedRepartitionDownDiagonal
        ( LTL, /**/ LTR,  L00, /**/ L01, L02,
         /*************/ /******************/
               /**/       L10, /**/ L11, L12,
          LBL, /**/ LBR,  L20, /**/ L21, L22 );

        LockedRepartitionRight
        ( XL_STAR_MC, /**/ XR_STAR_MC,
          X0_STAR_MC, /**/ X1_STAR_MC, X2_STAR_MC );

        RepartitionDown
        ( ZTAdjOrTrans_MR_STAR,  Z0AdjOrTrans_MR_STAR,
         /*********************/ /*********************/
                                  Z1AdjOrTrans_MR_STAR,
          ZBAdjOrTrans_MR_STAR,  Z2AdjOrTrans_MR_STAR );

        D11.AlignWith( L11 );
        //--------------------------------------------------------------------//
        D11 = L11;
        MakeTrapezoidal( LEFT, LOWER, 0, D11 );
        if( diag == UNIT )
            SetDiagonalToOne( D11 );
        LocalGemm
        ( orientation, orientation,
          alpha, D11, X1_STAR_MC, T(1), Z1AdjOrTrans_MR_STAR );

        LocalGemm
        ( orientation, orientation,
          alpha, L21, X2_STAR_MC, T(1), Z1AdjOrTrans_MR_STAR );
        //--------------------------------------------------------------------//
        D11.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( LTL, /**/ LTR,  L00, L01, /**/ L02,
               /**/       L10, L11, /**/ L12,
         /*************/ /******************/
          LBL, /**/ LBR,  L20, L21, /**/ L22 );

        SlideLockedPartitionRight
        ( XL_STAR_MC,             /**/ XR_STAR_MC,
          X0_STAR_MC, X1_STAR_MC, /**/ X2_STAR_MC );

        SlidePartitionDown
        ( ZTAdjOrTrans_MR_STAR,  Z0AdjOrTrans_MR_STAR,
                                 Z1AdjOrTrans_MR_STAR,
         /********************/ /********************/
          ZBAdjOrTrans_MR_STAR,  Z2AdjOrTrans_MR_STAR );
    }
    PopBlocksizeStack();
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #4
0
파일: LN.hpp 프로젝트: certik/Elemental
inline void
Her2kLN
( T alpha, const DistMatrix<T,MC,MR>& A,
           const DistMatrix<T,MC,MR>& B,
  T beta,        DistMatrix<T,MC,MR>& C )
{
#ifndef RELEASE
    PushCallStack("internal::Her2kLN");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( A.Height() != C.Height() || A.Height() != C.Width() ||
        B.Height() != C.Height() || B.Height() != C.Width() ||
        A.Width() != B.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal Her2kLN:\n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views 
    DistMatrix<T,MC,MR> AL(g), AR(g),
                        A0(g), A1(g), A2(g);
    DistMatrix<T,MC,MR> BL(g), BR(g),
                        B0(g), B1(g), B2(g);

    // Temporary distributions
    DistMatrix<T,MC,  STAR> A1_MC_STAR(g);
    DistMatrix<T,MC,  STAR> B1_MC_STAR(g);
    DistMatrix<T,VR,  STAR> A1_VR_STAR(g);
    DistMatrix<T,VR,  STAR> B1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > A1Adj_STAR_MR(g);
    DistMatrix<T,STAR,MR  > B1Adj_STAR_MR(g);

    A1_MC_STAR.AlignWith( C );
    B1_MC_STAR.AlignWith( C );
    A1_VR_STAR.AlignWith( C );
    B1_VR_STAR.AlignWith( C );
    A1Adj_STAR_MR.AlignWith( C );
    B1Adj_STAR_MR.AlignWith( C );

    // Start the algorithm
    ScaleTrapezoid( beta, LEFT, LOWER, 0, C );
    LockedPartitionRight( A, AL, AR, 0 );
    LockedPartitionRight( B, BL, BR, 0 );
    while( AR.Width() > 0 )
    {
        LockedRepartitionRight
        ( AL, /**/ AR,
          A0, /**/ A1, A2 );

        LockedRepartitionRight
        ( BL, /**/ BR,
          B0, /**/ B1, B2 );

        //--------------------------------------------------------------------//
        A1_VR_STAR = A1_MC_STAR = A1;
        A1Adj_STAR_MR.AdjointFrom( A1_VR_STAR );

        B1_VR_STAR = B1_MC_STAR = B1;
        B1Adj_STAR_MR.AdjointFrom( B1_VR_STAR );

        LocalTrr2k
        ( LOWER, 
          alpha, A1_MC_STAR, B1Adj_STAR_MR, 
                 B1_MC_STAR, A1Adj_STAR_MR,
          T(1),  C );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );

        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #5
0
파일: LLN.hpp 프로젝트: ahmadia/Elemental-1
inline void
LocalTrmmAccumulateLLN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
  const DistMatrix<T,MC,  MR  >& L,
  const DistMatrix<T,STAR,MR  >& XTrans_STAR_MR,
        DistMatrix<T,MC,  STAR>& Z_MC_STAR )
{
#ifndef RELEASE
    CallStackEntry entry("internal::LocalTrmmAccumulateLLN");
    if( L.Grid() != XTrans_STAR_MR.Grid() ||
        XTrans_STAR_MR.Grid() != Z_MC_STAR.Grid() )
        throw std::logic_error
        ("{L,X,Z} must be distributed over the same grid");
    if( L.Height() != L.Width() ||
        L.Height() != XTrans_STAR_MR.Width() ||
        L.Height() != Z_MC_STAR.Height() ||
        XTrans_STAR_MR.Height() != Z_MC_STAR.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal LocalTrmmAccumulateLLN: \n" 
            << "  L ~ " << L.Height() << " x " << L.Width() << "\n"
            << "  X^H/T[* ,MR] ~ " << XTrans_STAR_MR.Height() << " x "
                                   << XTrans_STAR_MR.Width() << "\n"
            << "  Z[MC,* ] ~ " << Z_MC_STAR.Height() << " x "
                               << Z_MC_STAR.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
    if( XTrans_STAR_MR.RowAlignment() != L.RowAlignment() ||
        Z_MC_STAR.ColAlignment() != L.ColAlignment() )
        throw std::logic_error("Partial matrix distributions are misaligned");
#endif
    const Grid& g = L.Grid();

    // Matrix views
    DistMatrix<T>
        LTL(g), LTR(g),  L00(g), L01(g), L02(g),
        LBL(g), LBR(g),  L10(g), L11(g), L12(g),
                         L20(g), L21(g), L22(g);

    DistMatrix<T> D11(g);

    DistMatrix<T,STAR,MR>
        XLTrans_STAR_MR(g), XRTrans_STAR_MR(g),
        X0Trans_STAR_MR(g), X1Trans_STAR_MR(g), 
        X2Trans_STAR_MR(g);

    DistMatrix<T,MC,STAR>
        ZT_MC_STAR(g),  Z0_MC_STAR(g),
        ZB_MC_STAR(g),  Z1_MC_STAR(g),
                        Z2_MC_STAR(g);

    const int ratio = std::max( g.Height(), g.Width() );
    PushBlocksizeStack( ratio*Blocksize() );

    LockedPartitionDownDiagonal
    ( L, LTL, LTR,
         LBL, LBR, 0 );
    LockedPartitionRight
    ( XTrans_STAR_MR, XLTrans_STAR_MR, XRTrans_STAR_MR, 0 );
    PartitionDown
    ( Z_MC_STAR, ZT_MC_STAR,
                 ZB_MC_STAR, 0 );
    while( LTL.Height() < L.Height() )
    {
        LockedRepartitionDownDiagonal
        ( LTL, /**/ LTR,  L00, /**/ L01, L02,
         /*************/ /******************/
               /**/       L10, /**/ L11, L12,
          LBL, /**/ LBR,  L20, /**/ L21, L22 );

        LockedRepartitionRight
        ( XLTrans_STAR_MR, /**/ XRTrans_STAR_MR,
          X0Trans_STAR_MR, /**/ X1Trans_STAR_MR, 
                                X2Trans_STAR_MR );

        RepartitionDown
        ( ZT_MC_STAR,  Z0_MC_STAR,
         /**********/ /**********/
                       Z1_MC_STAR,
          ZB_MC_STAR,  Z2_MC_STAR );

        D11.AlignWith( L11 );
        //--------------------------------------------------------------------//
        D11 = L11;
        MakeTriangular( LOWER, D11 );
        if( diag == UNIT )
            SetDiagonal( D11, T(1) );
        LocalGemm
        ( NORMAL, orientation, alpha, D11, X1Trans_STAR_MR, T(1), Z1_MC_STAR );
        LocalGemm
        ( NORMAL, orientation, alpha, L21, X1Trans_STAR_MR, T(1), Z2_MC_STAR );
        //--------------------------------------------------------------------//
        D11.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( LTL, /**/ LTR,  L00, L01, /**/ L02,
               /**/       L10, L11, /**/ L12,
         /*************/ /******************/
          LBL, /**/ LBR,  L20, L21, /**/ L22 );

        SlideLockedPartitionRight
        ( XLTrans_STAR_MR,                  /**/ XRTrans_STAR_MR,
          X0Trans_STAR_MR, X1Trans_STAR_MR, /**/ X2Trans_STAR_MR );

        SlidePartitionDown
        ( ZT_MC_STAR,  Z0_MC_STAR,
                       Z1_MC_STAR,
         /**********/ /**********/
          ZB_MC_STAR,  Z2_MC_STAR );
    }
    PopBlocksizeStack();
}
예제 #6
0
파일: TT.hpp 프로젝트: certik/Elemental
inline void
GemmTTC
( Orientation orientationOfA, 
  Orientation orientationOfB,
  T alpha, const DistMatrix<T>& A,
           const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::GemmTTC");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( orientationOfA == NORMAL || orientationOfB == NORMAL )
        throw std::logic_error
        ("GemmTTC expects A and B to be (Conjugate)Transposed");
    if( A.Width()  != C.Height() ||
        B.Height() != C.Width()  ||
        A.Height() != B.Width()    )
    {
        std::ostringstream msg;
        msg << "Nonconformal GemmTTC: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T> AT(g),  A0(g),
                  AB(g),  A1(g),
                          A2(g);
    DistMatrix<T> BL(g), BR(g),
                  B0(g), B1(g), B2(g);

    // Temporary distributions
    DistMatrix<T,STAR,MC  > A1_STAR_MC(g);
    DistMatrix<T,VR,  STAR> B1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > B1AdjOrTrans_STAR_MR(g);

    A1_STAR_MC.AlignWith( C );
    B1_VR_STAR.AlignWith( C );
    B1AdjOrTrans_STAR_MR.AlignWith( C );
    
    // Start the algorithm    
    Scale( beta, C );
    LockedPartitionDown
    ( A, AT,
         AB, 0 ); 
    LockedPartitionRight( B, BL, BR, 0 );
    while( AB.Height() > 0 )
    {
        LockedRepartitionDown
        ( AT,  A0,
         /**/ /**/
               A1,
          AB,  A2 );

        LockedRepartitionRight
        ( BL, /**/     BR,
          B0, /**/ B1, B2 );

        //--------------------------------------------------------------------//
        A1_STAR_MC = A1; 
        B1_VR_STAR = B1;
        if( orientationOfB == ADJOINT )
            B1AdjOrTrans_STAR_MR.AdjointFrom( B1_VR_STAR );
        else
            B1AdjOrTrans_STAR_MR.TransposeFrom( B1_VR_STAR );

        // C[MC,MR] += alpha (A1[*,MC])^[T/H] (B1[MR,*])^[T/H]
        //           = alpha (A1^[T/H])[MC,*] (B1^[T/H])[*,MR]
        LocalGemm
        ( orientationOfA, NORMAL, 
          alpha, A1_STAR_MC, B1AdjOrTrans_STAR_MR, T(1), C );
        //--------------------------------------------------------------------//

        SlideLockedPartitionDown
        ( AT,  A0,
               A1,
         /**/ /**/
          AB,  A2 );

        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #7
0
파일: NN.hpp 프로젝트: jimgoo/Elemental
inline void 
GemmNNDot
( T alpha, const DistMatrix<T>& A,
           const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::GemmNNDot");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( A.Height() != C.Height() ||
        B.Width()  != C.Width()  ||
        A.Width()  != B.Height() )
    {
        std::ostringstream msg;
        msg << "Nonconformal GemmNNDot: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
#endif
    const Grid& g = A.Grid();

    if( A.Height() > B.Width() )
    {
        // Matrix views
        DistMatrix<T> AT(g), AB(g),
                      A0(g), A1(g), A2(g);         
        DistMatrix<T> BL(g),  B0(g),
                      BR(g),  B1(g),
                              B2(g);
        DistMatrix<T> CT(g), C0(g), C1L(g), C1R(g),
                      CB(g), C1(g), C10(g), C11(g), C12(g),
                             C2(g);

        // Temporary distributions
        DistMatrix<T,STAR,VC> A1_STAR_VC(g);
        DistMatrix<T,VC,STAR> B1_VC_STAR(g);
        DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);

        // Star the algorithm
        Scale( beta, C );
        LockedPartitionDown
        ( A, AT,
             AB, 0 );
        PartitionDown
        ( C, CT,
             CB, 0 );
        while( AB.Height() > 0 )
        {
            LockedRepartitionDown
            ( AT,  A0,
             /**/ /**/
                   A1,
              AB,  A2 );

            RepartitionDown
            ( CT,  C0,
             /**/ /**/
                   C1,
              CB,  C2 );

            A1_STAR_VC = A1; 
            B1_VC_STAR.AlignWith( A1_STAR_VC );

            LockedPartitionRight( B, BL, BR, 0 );
            PartitionRight( C1, C1L, C1R, 0 );
            while( BR.Width() > 0 )
            {
                LockedRepartitionRight
                ( BL, /**/ BR,
                  B0, /**/ B1, B2 );

                RepartitionRight
                ( C1L, /**/ C1R,
                  C10, /**/ C11, C12 );

                Zeros( C11.Height(), C11.Width(), C11_STAR_STAR );
                //------------------------------------------------------------//
                B1_VC_STAR = B1;
                LocalGemm
                ( NORMAL, NORMAL, 
                  alpha, A1_STAR_VC, B1_VC_STAR, T(0), C11_STAR_STAR );
                C11.SumScatterUpdate( T(1), C11_STAR_STAR );
                //------------------------------------------------------------//

                SlideLockedPartitionRight
                ( BL,     /**/ BR,
                  B0, B1, /**/ B2 );

                SlidePartitionRight
                ( C1L,      /**/ C1R,
                  C10, C11, /**/ C12 );
            }
            B1_VC_STAR.FreeAlignments();

            SlideLockedPartitionDown
            ( AT,  A0,
                   A1,
             /**/ /**/
              AB,  A2 );

            SlidePartitionDown
            ( CT,  C0,
                   C1,
             /**/ /**/
              CB,  C2 );
        }
    }
    else
    {
        // Matrix views
        DistMatrix<T> AT(g), AB(g),
                      A0(g), A1(g), A2(g);         
        DistMatrix<T> BL(g),  B0(g),
                      BR(g),  B1(g),
                              B2(g);
        DistMatrix<T> 
            CL(g), CR(g),         C1T(g),  C01(g),
            C0(g), C1(g), C2(g),  C1B(g),  C11(g),
                                           C21(g);

        // Temporary distributions
        DistMatrix<T,STAR,VR> A1_STAR_VR(g);
        DistMatrix<T,VR,STAR> B1_VR_STAR(g);
        DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);

        // Star the algorithm
        Scale( beta, C );
        LockedPartitionRight( B, BL, BR, 0 );
        PartitionRight( C, CL, CR, 0 );
        while( BR.Width() > 0 )
        {
            LockedRepartitionRight
            ( BL, /**/ BR,
              B0, /**/ B1, B2 );

            RepartitionRight
            ( CL, /**/ CR,
              C0, /**/ C1, C2 );

            B1_VR_STAR = B1;
            A1_STAR_VR.AlignWith( B1_VR_STAR );

            LockedPartitionDown
            ( A, AT,
                 AB, 0 );
            PartitionDown
            ( C1, C1T,
                  C1B, 0 );
            while( AB.Height() > 0 )
            {
                LockedRepartitionDown
                ( AT,  A0,
                 /**/ /**/
                       A1,
                  AB,  A2 );

                RepartitionDown
                ( C1T,  C01,
                 /***/ /***/
                        C11,
                  C1B,  C21 );

                Zeros( C11.Height(), C11.Width(), C11_STAR_STAR );
                //------------------------------------------------------------//
                A1_STAR_VR = A1;
                LocalGemm
                ( NORMAL, NORMAL, 
                  alpha, A1_STAR_VR, B1_VR_STAR, T(0), C11_STAR_STAR );
                C11.SumScatterUpdate( T(1), C11_STAR_STAR );
                //------------------------------------------------------------//

                SlideLockedPartitionDown
                ( AT,  A0,
                       A1,
                 /**/ /**/
                  AB,  A2 );

                SlidePartitionDown
                ( C1T,  C01,
                        C11,
                 /***/ /***/
                  C1B,  C21 );
            }
            A1_STAR_VR.FreeAlignments();

            SlideLockedPartitionRight
            ( BL,     /**/ BR,
              B0, B1, /**/ B2 ); 

            SlidePartitionRight
            ( CL,     /**/ CR,
              C0, C1, /**/ C2 );
        }
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #8
0
파일: NN.hpp 프로젝트: jimgoo/Elemental
inline void
GemmNNA
( T alpha, const DistMatrix<T>& A,
           const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::GemmNNA");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( A.Height() != C.Height() ||
        B.Width()  != C.Width()  ||
        A.Width()  != B.Height() )
    {
        std::ostringstream msg;
        msg << "Nonconformal GemmNNA: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T> BL(g), BR(g),
                  B0(g), B1(g), B2(g);
    DistMatrix<T> CL(g), CR(g),
                  C0(g), C1(g), C2(g);

    // Temporary distributions
    DistMatrix<T,VR,STAR> B1_VR_STAR(g);
    DistMatrix<T,STAR,MR> B1Trans_STAR_MR(g);
    DistMatrix<T,MC,STAR> D1_MC_STAR(g);

    B1_VR_STAR.AlignWith( A );
    B1Trans_STAR_MR.AlignWith( A );
    D1_MC_STAR.AlignWith( A );

    // Start the algorithm
    Scale( beta, C );
    LockedPartitionRight( B, BL, BR, 0 );
    PartitionRight( C, CL, CR, 0 );
    while( BR.Width() > 0 )
    {
        LockedRepartitionRight
        ( BL, /**/     BR,
          B0, /**/ B1, B2 );

        RepartitionRight
        ( CL, /**/     CR,
          C0, /**/ C1, C2 );

        Zeros( C1.Height(), C1.Width(), D1_MC_STAR );
        //--------------------------------------------------------------------//
        B1_VR_STAR = B1;
        B1Trans_STAR_MR.TransposeFrom( B1_VR_STAR );

        // D1[MC,*] := alpha A[MC,MR] B1[MR,*]
        LocalGemm
        ( NORMAL, TRANSPOSE, alpha, A, B1Trans_STAR_MR, T(0), D1_MC_STAR );

        // C1[MC,MR] += scattered result of D1[MC,*] summed over grid rows
        C1.SumScatterUpdate( T(1), D1_MC_STAR );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );

        SlidePartitionRight
        ( CL,     /**/ CR,
          C0, C1, /**/ C2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #9
0
inline void
internal::LocalTrmmAccumulateLUN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
  const DistMatrix<T,MC,  MR  >& U,
  const DistMatrix<T,STAR,MR  >& XAdjOrTrans_STAR_MR,
        DistMatrix<T,MC,  STAR>& Z_MC_STAR )
{
#ifndef RELEASE
    PushCallStack("internal::LocalTrmmAccumulateLUN");
    if( U.Grid() != XAdjOrTrans_STAR_MR.Grid() ||
        XAdjOrTrans_STAR_MR.Grid() != Z_MC_STAR.Grid() )
        throw std::logic_error
        ("{U,X,Z} must be distributed over the same grid");
    if( U.Height() != U.Width() ||
        U.Height() != XAdjOrTrans_STAR_MR.Width() ||
        U.Height() != Z_MC_STAR.Height() ||
        XAdjOrTrans_STAR_MR.Height() != Z_MC_STAR.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal LocalTrmmAccumulateLUN: \n"
            << "  U ~ " << U.Height() << " x " << U.Width() << "\n"
            << "  X^H/T[* ,MR] ~ " << XAdjOrTrans_STAR_MR.Height() << " x "
                                   << XAdjOrTrans_STAR_MR.Width() << "\n"
            << "  Z[MC,* ] ~ " << Z_MC_STAR.Height() << " x "
                               << Z_MC_STAR.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
    if( XAdjOrTrans_STAR_MR.RowAlignment() != U.RowAlignment() ||
        Z_MC_STAR.ColAlignment() != U.ColAlignment() )
        throw std::logic_error("Partial matrix distributions are misaligned");
#endif
    const Grid& g = U.Grid();

    // Matrix views
    DistMatrix<T,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);

    DistMatrix<T,MC,MR> D11(g);

    DistMatrix<T,STAR,MR>
        XLAdjOrTrans_STAR_MR(g), XRAdjOrTrans_STAR_MR(g),
        X0AdjOrTrans_STAR_MR(g), X1AdjOrTrans_STAR_MR(g), 
        X2AdjOrTrans_STAR_MR(g);

    DistMatrix<T,MC,STAR>
        ZT_MC_STAR(g),  Z0_MC_STAR(g),
        ZB_MC_STAR(g),  Z1_MC_STAR(g),
                        Z2_MC_STAR(g);

    const int ratio = std::max( g.Height(), g.Width() );
    PushBlocksizeStack( ratio*Blocksize() );

    LockedPartitionDownDiagonal
    ( U, UTL, UTR,
         UBL, UBR, 0 );
    LockedPartitionRight
    ( XAdjOrTrans_STAR_MR, XLAdjOrTrans_STAR_MR, XRAdjOrTrans_STAR_MR, 0 );
    PartitionDown
    ( Z_MC_STAR, ZT_MC_STAR,
                 ZB_MC_STAR, 0 );
    while( UTL.Height() < U.Height() )
    {
        LockedRepartitionDownDiagonal
        ( UTL, /**/ UTR,  U00, /**/ U01, U02,
         /*************/ /******************/
               /**/       U10, /**/ U11, U12,
          UBL, /**/ UBR,  U20, /**/ U21, U22 );

        LockedRepartitionRight
        ( XLAdjOrTrans_STAR_MR, /**/ XRAdjOrTrans_STAR_MR,
          X0AdjOrTrans_STAR_MR, /**/ X1AdjOrTrans_STAR_MR, X2AdjOrTrans_STAR_MR 
        );

        RepartitionDown
        ( ZT_MC_STAR,  Z0_MC_STAR,
         /**********/ /**********/
                       Z1_MC_STAR,
          ZB_MC_STAR,  Z2_MC_STAR );

        D11.AlignWith( U11 );
        //--------------------------------------------------------------------//
        D11 = U11;
        MakeTrapezoidal( LEFT, UPPER, 0, D11 );
        if( diag == UNIT )
            SetDiagonalToOne( D11 );
        internal::LocalGemm
        ( NORMAL, orientation, alpha, D11, X1AdjOrTrans_STAR_MR,
          (T)1, Z1_MC_STAR );

        internal::LocalGemm
        ( NORMAL, orientation, alpha, U01, X1AdjOrTrans_STAR_MR,
          (T)1, Z0_MC_STAR );
        //--------------------------------------------------------------------//
        D11.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( UTL, /**/ UTR,  U00, U01, /**/ U02,
               /**/       U10, U11, /**/ U12,
         /*************/ /******************/
          UBL, /**/ UBR,  U20, U21, /**/ U22 );

        SlideLockedPartitionRight
        ( XLAdjOrTrans_STAR_MR,                       /**/ XRAdjOrTrans_STAR_MR,
          X0AdjOrTrans_STAR_MR, X1AdjOrTrans_STAR_MR, /**/ X2AdjOrTrans_STAR_MR 
        );

        SlidePartitionDown
        ( ZT_MC_STAR,  Z0_MC_STAR,
                       Z1_MC_STAR,
         /**********/ /**********/
          ZB_MC_STAR,  Z2_MC_STAR );
    }
    PopBlocksizeStack();
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #10
0
inline void
internal::GemmTNA
( Orientation orientationOfA,
  T alpha, const DistMatrix<T,MC,MR>& A,
           const DistMatrix<T,MC,MR>& B,
  T beta,        DistMatrix<T,MC,MR>& C )
{
#ifndef RELEASE
    PushCallStack("internal::GemmTNA");    
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( orientationOfA == NORMAL )
        throw std::logic_error("GemmTNA assumes A is (Conjugate)Transposed");
    if( A.Width()  != C.Height() ||
        B.Width()  != C.Width()  ||
        A.Height() != B.Height()   )
    {
        std::ostringstream msg;
        msg << "Nonconformal GemmTNA: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T,MC,MR> BL(g), BR(g),
                        B0(g), B1(g), B2(g);

    DistMatrix<T,MC,MR> CL(g), CR(g),
                        C0(g), C1(g), C2(g);

    // Temporary distributions
    DistMatrix<T,MC,STAR> B1_MC_STAR(g);
    DistMatrix<T,MR,STAR> D1_MR_STAR(g);
    DistMatrix<T,MR,MC  > D1_MR_MC(g);
    DistMatrix<T,MC,MR  > D1(g);

    // Start the algorithm
    Scal( beta, C );
    LockedPartitionRight( B, BL, BR, 0 );
    PartitionRight( C, CL, CR, 0 );
    while( BR.Width() > 0 )
    {
        LockedRepartitionRight
        ( BL, /**/     BR,
          B0, /**/ B1, B2 );
 
        RepartitionRight
        ( CL, /**/     CR,
          C0, /**/ C1, C2 );

        B1_MC_STAR.AlignWith( A );
        D1_MR_STAR.AlignWith( A );
        D1_MR_STAR.ResizeTo( C1.Height(), C1.Width() );
        D1.AlignWith( C1 );
        //--------------------------------------------------------------------//
        B1_MC_STAR = B1; // B1[MC,*] <- B1[MC,MR]

        // D1[MR,*] := alpha (A1[MC,MR])^T B1[MC,*]
        //           = alpha (A1^T)[MR,MC] B1[MC,*]
        internal::LocalGemm
        ( orientationOfA, NORMAL, alpha, A, B1_MC_STAR, (T)0, D1_MR_STAR );

        // C1[MC,MR] += scattered & transposed D1[MR,*] summed over grid cols
        D1_MR_MC.SumScatterFrom( D1_MR_STAR );
        D1 = D1_MR_MC; 
        Axpy( (T)1, D1, C1 );
        //--------------------------------------------------------------------//
        B1_MC_STAR.FreeAlignments();
        D1_MR_STAR.FreeAlignments();
        D1.FreeAlignments();

        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );

        SlidePartitionRight
        ( CL,     /**/ CR,
          C0, C1, /**/ C2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #11
0
inline void
internal::GemmTNB
( Orientation orientationOfA,
  T alpha, const DistMatrix<T,MC,MR>& A,
           const DistMatrix<T,MC,MR>& B,
  T beta,        DistMatrix<T,MC,MR>& C )
{
#ifndef RELEASE
    PushCallStack("internal::GemmTNB");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( orientationOfA == NORMAL )
        throw std::logic_error("GemmTNB assumes A is (Conjugate)Transposed");
    if( A.Width()  != C.Height() ||
        B.Width()  != C.Width()  ||
        A.Height() != B.Height()   )
    {
        std::ostringstream msg;
        msg << "Nonconformal GemmTNB: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T,MC,MR> AL(g), AR(g),
                        A0(g), A1(g), A2(g);

    DistMatrix<T,MC,MR> CT(g),  C0(g),
                        CB(g),  C1(g),
                                C2(g);

    // Temporary distributions
    DistMatrix<T,MC,STAR> A1_MC_STAR(g);
    DistMatrix<T,STAR,MR> D1_STAR_MR(g);

    // Start the algorithm
    Scal( beta, C );
    LockedPartitionRight( A, AL, AR, 0 );
    PartitionDown
    ( C, CT,
         CB, 0 );
    while( AR.Width() > 0 )
    {
        LockedRepartitionRight
        ( AL, /**/     AR,
          A0, /**/ A1, A2 );

        RepartitionDown
        ( CT,  C0,
         /**/ /**/
               C1,
          CB,  C2 );

        A1_MC_STAR.AlignWith( B );
        D1_STAR_MR.AlignWith( B );
        D1_STAR_MR.ResizeTo( C1.Height(), C1.Width() );
        //--------------------------------------------------------------------//
        A1_MC_STAR = A1; // A1[MC,*] <- A1[MC,MR]

        // D1[*,MR] := alpha (A1[MC,*])^T B[MC,MR]
        //           = alpha (A1^T)[*,MC] B[MC,MR]
        internal::LocalGemm
        ( orientationOfA, NORMAL, alpha, A1_MC_STAR, B, (T)0, D1_STAR_MR );

        // C1[MC,MR] += scattered result of D1[*,MR] summed over grid cols
        C1.SumScatterUpdate( (T)1, D1_STAR_MR );
        //--------------------------------------------------------------------//
        A1_MC_STAR.FreeAlignments();
        D1_STAR_MR.FreeAlignments();

        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );

        SlidePartitionDown
        ( CT,  C0,
               C1,
         /**/ /**/
          CB,  C2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #12
0
inline void
ReformHermitianMatrix
( UpperOrLower uplo,
        DistMatrix<R,MC,MR>& A,
  const DistMatrix<R,VR,STAR>& w,
  const DistMatrix<R,MC,MR>& Z,
  const RealFunctor& f )
{
#ifndef RELEASE
    PushCallStack("hermitian_function::ReformHermitianMatrix");
#endif
    const Grid& g = A.Grid();

    DistMatrix<R,MC,MR> ZL(g), ZR(g),
                        Z0(g), Z1(g), Z2(g);
    DistMatrix<R,VR,STAR> wT(g),  w0(g),
                          wB(g),  w1(g),
                                  w2(g);

    DistMatrix<R,MC,  STAR> Z1_MC_STAR(g);
    DistMatrix<R,VR,  STAR> Z1_VR_STAR(g);
    DistMatrix<R,STAR,MR  > Z1Trans_STAR_MR(g);
    DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);

    if( uplo == LOWER )
        MakeTrapezoidal( LEFT, UPPER, 1, A );
    else
        MakeTrapezoidal( LEFT, LOWER, -1, A );
    LockedPartitionRight( Z, ZL, ZR, 0 );
    LockedPartitionDown
    ( w, wT,
         wB, 0 );
    while( ZL.Width() < Z.Width() )
    {
        LockedRepartitionRight
        ( ZL, /**/ ZR,
          Z0, /**/ Z1, Z2 );
        LockedRepartitionDown
        ( wT,  w0,
         /**/ /**/
               w1,
          wB,  w2 );

        Z1_MC_STAR.AlignWith( A );
        Z1_VR_STAR.AlignWith( A );
        Z1Trans_STAR_MR.AlignWith( A );
        //--------------------------------------------------------------------//
        Z1_MC_STAR = Z1;
        Z1_VR_STAR = Z1_MC_STAR;
        w1_STAR_STAR = w1;

        // Scale Z1[VR,* ] with the modified eigenvalues
        const int width = Z1_VR_STAR.Width();
        const int localHeight = Z1_VR_STAR.LocalHeight();
        for( int j=0; j<width; ++j )
        {
            const R omega = f(w1_STAR_STAR.GetLocalEntry(j,0));
            R* buffer = Z1_VR_STAR.LocalBuffer(0,j);
            for( int iLocal=0; iLocal<localHeight; ++iLocal )
                buffer[iLocal] *= omega;
        }

        Z1Trans_STAR_MR.TransposeFrom( Z1_VR_STAR );
        internal::LocalTrrk( uplo, (R)1, Z1_MC_STAR, Z1Trans_STAR_MR, (R)1, A );
        //--------------------------------------------------------------------//
        Z1Trans_STAR_MR.FreeAlignments();
        Z1_VR_STAR.FreeAlignments();
        Z1_MC_STAR.FreeAlignments();

        SlideLockedPartitionDown
        ( wT,  w0,
               w1,
         /**/ /**/
          wB,  w2 );
        SlideLockedPartitionRight
        ( ZL,     /**/ ZR,
          Z0, Z1, /**/ Z2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #13
0
inline void
ReformNormalMatrix
(       DistMatrix<Complex<R>,MC,MR  >& A,
  const DistMatrix<R,         VR,STAR>& w,
  const DistMatrix<Complex<R>,MC,MR  >& Z,
  const ComplexFunctor& f )
{
#ifndef RELEASE
    PushCallStack("hermitian_function::ReformNormalMatrix");
#endif
    const Grid& g = A.Grid();
    typedef Complex<R> C;

    DistMatrix<C,MC,MR> ZL(g), ZR(g),
                        Z0(g), Z1(g), Z2(g);
    DistMatrix<R,VR,STAR> wT(g),  w0(g),
                          wB(g),  w1(g),
                                  w2(g);

    DistMatrix<C,MC,  STAR> Z1_MC_STAR(g);
    DistMatrix<C,VR,  STAR> Z1_VR_STAR(g);
    DistMatrix<C,STAR,MR  > Z1Adj_STAR_MR(g);
    DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);

    Zero( A );
    LockedPartitionRight( Z, ZL, ZR, 0 );
    LockedPartitionDown
    ( w, wT,
         wB, 0 );
    while( ZL.Width() < Z.Width() )
    {
        LockedRepartitionRight
        ( ZL, /**/ ZR,
          Z0, /**/ Z1, Z2 );
        LockedRepartitionDown
        ( wT,  w0,
         /**/ /**/
               w1,
          wB,  w2 );

        Z1_MC_STAR.AlignWith( A );
        Z1_VR_STAR.AlignWith( A );
        Z1Adj_STAR_MR.AlignWith( A );
        //--------------------------------------------------------------------//
        Z1_MC_STAR = Z1;
        Z1_VR_STAR = Z1_MC_STAR;
        w1_STAR_STAR = w1;

        // Scale Z1[VR,* ] with the modified eigenvalues
        const int width = Z1_VR_STAR.Width();
        const int localHeight = Z1_VR_STAR.LocalHeight();
        for( int j=0; j<width; ++j )
        {
            const C conjOmega = Conj(f(w1_STAR_STAR.GetLocalEntry(j,0)));
            C* buffer = Z1_VR_STAR.LocalBuffer(0,j);
            for( int iLocal=0; iLocal<localHeight; ++iLocal )
                buffer[iLocal] *= conjOmega;
        }

        Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
        internal::LocalGemm
        ( NORMAL, NORMAL, (C)1, Z1_MC_STAR, Z1Adj_STAR_MR, (C)1, A );
        //--------------------------------------------------------------------//
        Z1Adj_STAR_MR.FreeAlignments();
        Z1_VR_STAR.FreeAlignments();
        Z1_MC_STAR.FreeAlignments();

        SlideLockedPartitionDown
        ( wT,  w0,
               w1,
         /**/ /**/
          wB,  w2 );
        SlideLockedPartitionRight
        ( ZL,     /**/ ZR,
          Z0, Z1, /**/ Z2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #14
0
inline void
HermitianFromEVD
( UpperOrLower uplo,
        DistMatrix<F>& A,
  const DistMatrix<BASE(F),VR,STAR>& w,
  const DistMatrix<F>& Z )
{
#ifndef RELEASE
    CallStackEntry entry("HermitianFromEVD");
#endif
    const Grid& g = A.Grid();
    typedef BASE(F) R;

    DistMatrix<F> ZL(g), ZR(g),
                  Z0(g), Z1(g), Z2(g);
    DistMatrix<R,VR,STAR> wT(g),  w0(g),
                          wB(g),  w1(g),
                                  w2(g);

    DistMatrix<F,MC,  STAR> Z1_MC_STAR(g);
    DistMatrix<F,VR,  STAR> Z1_VR_STAR(g);
    DistMatrix<F,STAR,MR  > Z1Adj_STAR_MR(g);
    DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);

    A.ResizeTo( Z.Height(), Z.Height() );
    if( uplo == LOWER )
        MakeTrapezoidal( UPPER, A, 1 );
    else
        MakeTrapezoidal( LOWER, A, -1 );
    LockedPartitionRight( Z, ZL, ZR, 0 );
    LockedPartitionDown
    ( w, wT,
         wB, 0 );
    while( ZL.Width() < Z.Width() )
    {
        LockedRepartitionRight
        ( ZL, /**/ ZR,
          Z0, /**/ Z1, Z2 );
        LockedRepartitionDown
        ( wT,  w0,
         /**/ /**/
               w1,
          wB,  w2 );

        Z1_MC_STAR.AlignWith( A );
        Z1_VR_STAR.AlignWith( A );
        Z1Adj_STAR_MR.AlignWith( A );
        //--------------------------------------------------------------------//
        Z1_MC_STAR = Z1;
        Z1_VR_STAR = Z1_MC_STAR;
        w1_STAR_STAR = w1;

        DiagonalScale( RIGHT, NORMAL, w1_STAR_STAR, Z1_VR_STAR );

        Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
        LocalTrrk( uplo, F(1), Z1_MC_STAR, Z1Adj_STAR_MR, F(1), A );
        //--------------------------------------------------------------------//

        SlideLockedPartitionDown
        ( wT,  w0,
               w1,
         /**/ /**/
          wB,  w2 );
        SlideLockedPartitionRight
        ( ZL,     /**/ ZR,
          Z0, Z1, /**/ Z2 );
    }
}
예제 #15
0
파일: TTTT.hpp 프로젝트: jimgoo/Elemental
inline void
Trr2kTTTT
( UpperOrLower uplo,
  Orientation orientationOfA, Orientation orientationOfB,
  Orientation orientationOfC, Orientation orientationOfD, 
  T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
           const DistMatrix<T>& C, const DistMatrix<T>& D,
  T beta,        DistMatrix<T>& E )
{
#ifndef RELEASE
    PushCallStack("internal::Trr2kTTTT");
    if( E.Height() != E.Width()  || A.Height() != C.Height() ||
        A.Width()  != E.Height() || C.Width()  != E.Height() ||
        B.Height() != E.Width()  || D.Height() != E.Width()  ||
        A.Height() != B.Width()  || C.Height() != D.Width() )
        throw std::logic_error("Nonconformal Trr2kTTTT");
#endif
    const Grid& g = E.Grid();

    DistMatrix<T> AT(g),  A0(g),
                  AB(g),  A1(g),
                          A2(g);
    DistMatrix<T> BL(g), BR(g),
                  B0(g), B1(g), B2(g);

    DistMatrix<T> CT(g),  C0(g),
                  CB(g),  C1(g),
                          C2(g);
    DistMatrix<T> DL(g), DR(g),
                  D0(g), D1(g), D2(g);

    DistMatrix<T,STAR,MC  > A1_STAR_MC(g);
    DistMatrix<T,VR,  STAR> B1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > B1AdjOrTrans_STAR_MR(g);
    DistMatrix<T,STAR,MC  > C1_STAR_MC(g);
    DistMatrix<T,VR,  STAR> D1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > D1AdjOrTrans_STAR_MR(g);

    A1_STAR_MC.AlignWith( E );
    B1_VR_STAR.AlignWith( E );
    B1AdjOrTrans_STAR_MR.AlignWith( E );
    C1_STAR_MC.AlignWith( E );
    D1_VR_STAR.AlignWith( E );
    D1AdjOrTrans_STAR_MR.AlignWith( E );

    LockedPartitionDown
    ( A, AT,
         AB, 0 );
    LockedPartitionRight( B, BL, BR, 0 );
    LockedPartitionDown
    ( C, CT,
         CB, 0 );
    LockedPartitionRight( D, DL, DR, 0 );
    while( AT.Height() < A.Height() )
    {
        LockedRepartitionDown
        ( AT,  A0,
         /**/ /**/
               A1,
          AB,  A2 );
        LockedRepartitionRight
        ( BL, /**/ BR,
          B0, /**/ B1, B2 );
        LockedRepartitionDown
        ( CT,  C0,
         /**/ /**/
               C1,
          CB,  C2 );
        LockedRepartitionRight
        ( DL, /**/ DR,
          D0, /**/ D1, D2 );

        //--------------------------------------------------------------------//
        A1_STAR_MC = A1;
        C1_STAR_MC = C1;
        B1_VR_STAR = B1;
        D1_VR_STAR = D1;
        if( orientationOfB == ADJOINT )
            B1AdjOrTrans_STAR_MR.AdjointFrom( B1_VR_STAR );
        else
            B1AdjOrTrans_STAR_MR.TransposeFrom( B1_VR_STAR );
        if( orientationOfD == ADJOINT )
            D1AdjOrTrans_STAR_MR.AdjointFrom( D1_VR_STAR );
        else
            D1AdjOrTrans_STAR_MR.TransposeFrom( D1_VR_STAR );
        LocalTrr2k
        ( uplo, orientationOfA, orientationOfC,
          alpha, A1_STAR_MC, B1AdjOrTrans_STAR_MR,
                 C1_STAR_MC, D1AdjOrTrans_STAR_MR,
          beta,  E );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight
        ( DL,     /**/ DR,
          D0, D1, /**/ D2 );
        SlideLockedPartitionDown
        ( CT,  C0,
               C1,
         /**/ /**/
          CB,  C2 );
        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );
        SlideLockedPartitionDown
        ( AT,  A0,
               A1,
         /**/ /**/
          AB,  A2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #16
0
void Trr2kNTTN
( UpperOrLower uplo,
  Orientation orientationOfB, Orientation orientationOfC,
  T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
           const DistMatrix<T>& C, const DistMatrix<T>& D,
  T beta,        DistMatrix<T>& E )
{
#ifndef RELEASE
    CallStackEntry entry("internal::Trr2kNTTN");
    if( E.Height() != E.Width()  || A.Width()  != C.Height() ||
        A.Height() != E.Height() || C.Width()  != E.Height() ||
        B.Height() != E.Width()  || D.Width()  != E.Width()  ||
        A.Width()  != B.Width()  || C.Height() != D.Height() )
        LogicError("Nonconformal Trr2kNTTN");
#endif
    const Grid& g = E.Grid();

    DistMatrix<T> AL(g), AR(g),
                  A0(g), A1(g), A2(g);
    DistMatrix<T> BL(g), BR(g),
                  B0(g), B1(g), B2(g);

    DistMatrix<T> CT(g),  C0(g),
                  CB(g),  C1(g),
                          C2(g);
    DistMatrix<T> DT(g),  D0(g),
                  DB(g),  D1(g),
                          D2(g);

    DistMatrix<T,MC,  STAR> A1_MC_STAR(g);
    DistMatrix<T,VR,  STAR> B1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > B1AdjOrTrans_STAR_MR(g);
    DistMatrix<T,STAR,MC  > C1_STAR_MC(g);
    DistMatrix<T,MR,  STAR> D1Trans_MR_STAR(g);

    A1_MC_STAR.AlignWith( E );
    B1_VR_STAR.AlignWith( E );
    B1AdjOrTrans_STAR_MR.AlignWith( E );
    C1_STAR_MC.AlignWith( E );
    D1Trans_MR_STAR.AlignWith( E );

    LockedPartitionRight( A, AL, AR, 0 );
    LockedPartitionRight( B, BL, BR, 0 );
    LockedPartitionDown
    ( C, CT,
         CB, 0 );
    LockedPartitionDown
    ( D, DT,
         DB, 0 );
    while( AL.Width() < A.Width() )
    {
        LockedRepartitionRight
        ( AL, /**/ AR,
          A0, /**/ A1, A2 );
        LockedRepartitionRight
        ( BL, /**/ BR,
          B0, /**/ B1, B2 );
        LockedRepartitionDown
        ( CT,  C0,
         /**/ /**/
               C1,
          CB,  C2 );
        LockedRepartitionDown
        ( DT,  D0,
         /**/ /**/
               D1,
          DB,  D2 );

        //--------------------------------------------------------------------//
        A1_MC_STAR = A1;
        C1_STAR_MC = C1;
        B1_VR_STAR = B1;
        if( orientationOfB == ADJOINT )
            B1AdjOrTrans_STAR_MR.AdjointFrom( B1_VR_STAR );
        else
            B1AdjOrTrans_STAR_MR.TransposeFrom( B1_VR_STAR );
        D1Trans_MR_STAR.TransposeFrom( D1 );
        LocalTrr2k 
        ( uplo, orientationOfC, TRANSPOSE,
          alpha, A1_MC_STAR, B1AdjOrTrans_STAR_MR,
                 C1_STAR_MC, D1Trans_MR_STAR,
          beta,  E );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );
        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );
        SlideLockedPartitionDown
        ( CT,  C0,
               C1,
         /**/ /**/
          CB,  C2 );
        SlideLockedPartitionDown
        ( DT,  D0,
               D1,
         /**/ /**/
          DB,  D2 );
    }
}
예제 #17
0
파일: U.hpp 프로젝트: jimgoo/Elemental
inline void
LocalSymvRowAccumulateU
( T alpha, 
  const DistMatrix<T>& A,
  const DistMatrix<T,STAR,MC>& x_STAR_MC,
  const DistMatrix<T,STAR,MR>& x_STAR_MR,
        DistMatrix<T,STAR,MC>& z_STAR_MC,
        DistMatrix<T,STAR,MR>& z_STAR_MR )
{
#ifndef RELEASE
    PushCallStack("internal::LocalSymvRowAccumulateU");
    if( A.Grid() != x_STAR_MC.Grid() ||
        x_STAR_MC.Grid() != x_STAR_MR.Grid() ||
        x_STAR_MR.Grid() != z_STAR_MC.Grid() ||
        z_STAR_MC.Grid() != z_STAR_MR.Grid() )
        throw std::logic_error
        ("{A,x,z} must be distributed over the same grid");
    if( x_STAR_MC.Height() != 1 || x_STAR_MR.Height() != 1 ||
        z_STAR_MC.Height() != 1 || z_STAR_MR.Height() != 1 )
        throw std::logic_error("Expected x and z to be row vectors");
    if( A.Height() != A.Width() || 
        A.Height() != x_STAR_MC.Width() ||
        A.Height() != x_STAR_MR.Width() ||
        A.Height() != z_STAR_MC.Width() ||
        A.Height() != z_STAR_MR.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal LocalSymvRowAccumulateU: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  x[* ,MC] ~ " << x_STAR_MC.Height() << " x " 
                               << x_STAR_MC.Width() << "\n"
            << "  x[* ,MR] ~ " << x_STAR_MR.Height() << " x " 
                               << x_STAR_MR.Width() << "\n"
            << "  z[* ,MC] ~ " << z_STAR_MC.Height() << " x " 
                               << z_STAR_MC.Width() << "\n"
            << "  z[* ,MR] ~ " << z_STAR_MR.Height() << " x " 
                               << z_STAR_MR.Width() << "\n";
        throw std::logic_error( msg.str() );
    }
    if( x_STAR_MC.RowAlignment() != A.ColAlignment() ||
        x_STAR_MR.RowAlignment() != A.RowAlignment() ||
        z_STAR_MC.RowAlignment() != A.ColAlignment() ||
        z_STAR_MR.RowAlignment() != A.RowAlignment() )
        throw std::logic_error("Partial matrix distributions are misaligned");
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T> A11(g), A12(g);
    DistMatrix<T> D11(g);

    DistMatrix<T,STAR,MC> x1_STAR_MC(g);
    DistMatrix<T,STAR,MR> 
        xL_STAR_MR(g), xR_STAR_MR(g),
        x0_STAR_MR(g), x1_STAR_MR(g), x2_STAR_MR(g);
    DistMatrix<T,STAR,MC> z1_STAR_MC(g);
    DistMatrix<T,STAR,MR> z1_STAR_MR(g), z2_STAR_MR(g);

    // We want our local gemvs to be of width blocksize, so we will 
    // temporarily change to max(r,c) times the current blocksize
    const int ratio = std::max( g.Height(), g.Width() );
    PushBlocksizeStack( ratio*LocalSymvBlocksize<T>() );
                 
    LockedPartitionRight( x_STAR_MR,  xL_STAR_MR, xR_STAR_MR, 0 );
    while( xL_STAR_MR.Width() < x_STAR_MR.Width() )
    {
        LockedRepartitionRight
        ( xL_STAR_MR, /**/ xR_STAR_MR, 
          x0_STAR_MR, /**/ x1_STAR_MR, x2_STAR_MR );

        const int n0 = x0_STAR_MR.Width();
        const int n1 = x1_STAR_MR.Width();
        const int n2 = x2_STAR_MR.Width();
        LockedView( A11, A, n0, n0,    n1, n1 );
        LockedView( A12, A, n0, n0+n1, n1, n2 );
        LockedView( x1_STAR_MC, x_STAR_MC, 0, n0, 1, n1 );
        View( z1_STAR_MC, z_STAR_MC, 0, n0,    1, n1 );
        View( z1_STAR_MR, z_STAR_MR, 0, n0,    1, n1 );
        View( z2_STAR_MR, z_STAR_MR, 0, n0+n1, 1, n2 );

        D11.AlignWith( A11 );
        //--------------------------------------------------------------------//
        // TODO: These diagonal block updates can be greatly improved
        D11 = A11;
        MakeTrapezoidal( LEFT, UPPER, 0, D11 );
        Gemv
        ( NORMAL, 
          alpha, D11.LockedLocalMatrix(), 
                 x1_STAR_MR.LockedLocalMatrix(),
          T(1),  z1_STAR_MC.LocalMatrix() );
        MakeTrapezoidal( LEFT, UPPER, 1, D11 );
        Gemv
        ( TRANSPOSE,
          alpha, D11.LockedLocalMatrix(),
                 x1_STAR_MC.LockedLocalMatrix(),
          T(1),  z1_STAR_MR.LocalMatrix() );

        Gemv
        ( NORMAL,
          alpha, A12.LockedLocalMatrix(),
                 x2_STAR_MR.LockedLocalMatrix(),
          T(1),  z1_STAR_MC.LocalMatrix() );
        Gemv
        ( TRANSPOSE,
          alpha, A12.LockedLocalMatrix(),
                 x1_STAR_MC.LockedLocalMatrix(),
          T(1),  z2_STAR_MR.LocalMatrix() );
        //--------------------------------------------------------------------//
        D11.FreeAlignments();

        SlideLockedPartitionRight
        ( xL_STAR_MR,             /**/ xR_STAR_MR,
          x0_STAR_MR, x1_STAR_MR, /**/ x2_STAR_MR );
    }
    PopBlocksizeStack();
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #18
0
파일: NN.hpp 프로젝트: jimgoo/Elemental
inline void 
GemmNNC
( T alpha, const DistMatrix<T>& A,
           const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::GemmNNC");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( A.Height() != C.Height() ||
        B.Width()  != C.Width()  ||
        A.Width()  != B.Height() )
    {
        std::ostringstream msg;
        msg << "Nonconformal GemmNNC: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T> AL(g), AR(g),
                  A0(g), A1(g), A2(g);         
    DistMatrix<T> BT(g),  B0(g),
                  BB(g),  B1(g),
                          B2(g);

    // Temporary distributions
    DistMatrix<T,MC,STAR> A1_MC_STAR(g);
    DistMatrix<T,MR,STAR> B1Trans_MR_STAR(g); 

    A1_MC_STAR.AlignWith( C );
    B1Trans_MR_STAR.AlignWith( C );

    // Start the algorithm
    Scale( beta, C );
    LockedPartitionRight( A, AL, AR, 0 ); 
    LockedPartitionDown
    ( B, BT, 
         BB, 0 ); 
    while( AR.Width() > 0 )
    {
        LockedRepartitionRight( AL, /**/ AR,
                                A0, /**/ A1, A2 );

        LockedRepartitionDown( BT,  B0,
                              /**/ /**/
                                    B1, 
                               BB,  B2 );

        //--------------------------------------------------------------------//
        A1_MC_STAR = A1; 
        B1Trans_MR_STAR.TransposeFrom( B1 );

        // C[MC,MR] += alpha A1[MC,*] (B1^T[MR,*])^T
        //           = alpha A1[MC,*] B1[*,MR]
        LocalGemm
        ( NORMAL, TRANSPOSE, alpha, A1_MC_STAR, B1Trans_MR_STAR, T(1), C );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight( AL,     /**/ AR,
                                   A0, A1, /**/ A2 );

        SlideLockedPartitionDown( BT,  B0,
                                       B1,
                                 /**/ /**/
                                  BB,  B2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #19
0
파일: UN.hpp 프로젝트: ahmadia/Elemental-1
inline void
Syr2kUN
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C,
  bool conjugate=false )
{
#ifndef RELEASE
    CallStackEntry entry("internal::Syr2kUN");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( A.Height() != C.Height() || A.Height() != C.Width() ||
        B.Height() != C.Height() || B.Height() != C.Width() ||
        A.Width() != B.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal Syr2kUN:\n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
#endif
    const Grid& g = C.Grid();

    // Matrix views 
    DistMatrix<T> AL(g), AR(g),
                  A0(g), A1(g), A2(g);
    DistMatrix<T> BL(g), BR(g),
                  B0(g), B1(g), B2(g);

    // Temporary distributions
    DistMatrix<T,MC,  STAR> A1_MC_STAR(g);
    DistMatrix<T,MC,  STAR> B1_MC_STAR(g);
    DistMatrix<T,VR,  STAR> A1_VR_STAR(g);
    DistMatrix<T,VR,  STAR> B1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > A1Trans_STAR_MR(g);
    DistMatrix<T,STAR,MR  > B1Trans_STAR_MR(g);

    A1_MC_STAR.AlignWith( C );
    B1_MC_STAR.AlignWith( C );
    A1_VR_STAR.AlignWith( C );
    B1_VR_STAR.AlignWith( C );
    A1Trans_STAR_MR.AlignWith( C );
    B1Trans_STAR_MR.AlignWith( C );

    // Start the algorithm
    ScaleTrapezoid( beta, LEFT, UPPER, 0, C );
    LockedPartitionRight( A, AL, AR, 0 );
    LockedPartitionRight( B, BL, BR, 0 );
    while( AR.Width() > 0 )
    {
        LockedRepartitionRight
        ( AL, /**/ AR,
          A0, /**/ A1, A2 );

        LockedRepartitionRight
        ( BL, /**/ BR,
          B0, /**/ B1, B2 );

        //--------------------------------------------------------------------//
        A1_VR_STAR = A1_MC_STAR = A1;
        A1Trans_STAR_MR.TransposeFrom( A1_VR_STAR, conjugate );

        B1_VR_STAR = B1_MC_STAR = B1;
        B1Trans_STAR_MR.TransposeFrom( B1_VR_STAR, conjugate );

        LocalTrr2k
        ( UPPER, 
          alpha, A1_MC_STAR, B1Trans_STAR_MR,
                 B1_MC_STAR, A1Trans_STAR_MR,
          T(1),  C );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );

        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );
    }
}
예제 #20
0
파일: TT.hpp 프로젝트: certik/Elemental
inline void
GemmTTB
( Orientation orientationOfA, 
  Orientation orientationOfB,
  T alpha, const DistMatrix<T>& A,
           const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::GemmTTB");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( orientationOfA == NORMAL || orientationOfB == NORMAL )
        throw std::logic_error
        ("GemmTTB expects A and B to be (Conjugate)Transposed");
    if( A.Width()  != C.Height() ||
        B.Height() != C.Width()  ||
        A.Height() != B.Width()    )
    {
        std::ostringstream msg;
        msg << "Nonconformal GemmTTB: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T> AL(g), AR(g),
                  A0(g), A1(g), A2(g);
    DistMatrix<T> CT(g),  C0(g),
                  CB(g),  C1(g),
                          C2(g);

    // Temporary distributions
    DistMatrix<T,VR,  STAR> A1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > A1AdjOrTrans_STAR_MR(g);
    DistMatrix<T,STAR,MC  > D1_STAR_MC(g);
    DistMatrix<T,MR,  MC  > D1_MR_MC(g);
    DistMatrix<T> D1(g);

    A1_VR_STAR.AlignWith( B );
    A1AdjOrTrans_STAR_MR.AlignWith( B );
    D1_STAR_MC.AlignWith( B );

    // Start the algorithm 
    Scale( beta, C );
    LockedPartitionRight( A, AL, AR, 0 );
    PartitionDown
    ( C, CT,
         CB, 0 );
    while( AR.Width() > 0 )
    {
        LockedRepartitionRight
        ( AL, /**/     AR,
          A0, /**/ A1, A2 );
 
        RepartitionDown
        ( CT,  C0,
         /**/ /**/
               C1,
          CB,  C2 );

        D1.AlignWith( C1 );
        Zeros( C1.Height(), C1.Width(), D1_STAR_MC );
        //--------------------------------------------------------------------//
        A1_VR_STAR = A1;
        if( orientationOfA == ADJOINT )
            A1AdjOrTrans_STAR_MR.AdjointFrom( A1_VR_STAR );
        else
            A1AdjOrTrans_STAR_MR.TransposeFrom( A1_VR_STAR );
 
        // D1[*,MC] := alpha (A1[MR,*])^[T/H] (B[MC,MR])^[T/H]
        //           = alpha (A1^[T/H])[*,MR] (B^[T/H])[MR,MC]
        LocalGemm
        ( NORMAL, orientationOfB, 
          alpha, A1AdjOrTrans_STAR_MR, B, T(0), D1_STAR_MC );

        // C1[MC,MR] += scattered & transposed D1[*,MC] summed over grid rows
        D1_MR_MC.SumScatterFrom( D1_STAR_MC );
        D1 = D1_MR_MC; 
        Axpy( T(1), D1, C1 );
        //--------------------------------------------------------------------//
        D1.FreeAlignments();

        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );

        SlidePartitionDown
        ( CT,  C0,
               C1,
         /**/ /**/
          CB,  C2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #21
0
파일: RU.hpp 프로젝트: certik/Elemental
inline void
HemmRUC
( T alpha, const DistMatrix<T>& A,
           const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::HemmRUC");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error("{A,B,C} must be distributed on the same grid");
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T> 
        ATL(g), ATR(g),  A00(g), A01(g), A02(g),  AColPan(g),
        ABL(g), ABR(g),  A10(g), A11(g), A12(g),  ARowPan(g),
                         A20(g), A21(g), A22(g);
    DistMatrix<T> BL(g), BR(g),
                  B0(g), B1(g), B2(g);
    DistMatrix<T> CL(g), CR(g),
                  C0(g), C1(g), C2(g),
                  CLeft(g), CRight(g);

    // Temporary distributions
    DistMatrix<T,MC,STAR> B1_MC_STAR(g);
    DistMatrix<T,VR,  STAR> AColPan_VR_STAR(g);
    DistMatrix<T,STAR,MR  > AColPanAdj_STAR_MR(g);
    DistMatrix<T,MR,  STAR> ARowPanAdj_MR_STAR(g);

    B1_MC_STAR.AlignWith( C );

    // Start the algorithm
    Scale( beta, C );
    LockedPartitionDownDiagonal
    ( A, ATL, ATR,
         ABL, ABR, 0 );
    LockedPartitionRight( B, BL, BR, 0 );
    PartitionRight( C, CL, CR, 0 );
    while( CR.Width() > 0 )
    {
        LockedRepartitionDownDiagonal
        ( ATL, /**/ ATR,  A00, /**/ A01, A02,
         /*************/ /******************/
               /**/       A10, /**/ A11, A12,
          ABL, /**/ ABR,  A20, /**/ A21, A22 );

        LockedRepartitionRight
        ( BL, /**/ BR,
          B0, /**/ B1, B2 );

        RepartitionRight
        ( CL, /**/ CR,
          C0, /**/ C1, C2 );

        ARowPan.LockedView1x2( A11, A12 );
        AColPan.LockedView2x1
        ( A01,
          A11 );

        CLeft.View1x2( C0, C1 );
        CRight.View1x2( C1, C2 );

        AColPan_VR_STAR.AlignWith( CLeft );
        AColPanAdj_STAR_MR.AlignWith( CLeft );
        ARowPanAdj_MR_STAR.AlignWith( CRight );
        //--------------------------------------------------------------------//
        B1_MC_STAR = B1;

        AColPan_VR_STAR = AColPan;
        AColPanAdj_STAR_MR.AdjointFrom( AColPan_VR_STAR );
        ARowPanAdj_MR_STAR.AdjointFrom( ARowPan );
        MakeTrapezoidal( LEFT,  LOWER,  0, ARowPanAdj_MR_STAR );
        MakeTrapezoidal( RIGHT, LOWER, -1, AColPanAdj_STAR_MR );

        LocalGemm
        ( NORMAL, ADJOINT, 
          alpha, B1_MC_STAR, ARowPanAdj_MR_STAR, T(1), CRight );

        LocalGemm
        ( NORMAL, NORMAL,
          alpha, B1_MC_STAR, AColPanAdj_STAR_MR, T(1), CLeft );
        //--------------------------------------------------------------------//
        AColPan_VR_STAR.FreeAlignments();
        AColPanAdj_STAR_MR.FreeAlignments();
        ARowPanAdj_MR_STAR.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( ATL, /**/ ATR,  A00, A01, /**/ A02,
               /**/       A10, A11, /**/ A12,
         /*************/ /******************/
          ABL, /**/ ABR,  A20, A21, /**/ A22 );

        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );

        SlidePartitionRight
        ( CL,     /**/ CR,
          C0, C1, /**/ C2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #22
0
파일: LU.hpp 프로젝트: jimgoo/Elemental
inline void
LocalSymmetricAccumulateLU
( Orientation orientation, T alpha,
  const DistMatrix<T>& A,
  const DistMatrix<T,MC,  STAR>& B_MC_STAR,
  const DistMatrix<T,STAR,MR  >& BAdjOrTrans_STAR_MR,
        DistMatrix<T,MC,  STAR>& Z_MC_STAR,
        DistMatrix<T,MR,  STAR>& Z_MR_STAR )
{
#ifndef RELEASE
    PushCallStack("internal::LocalSymmetricAccumulateLU");
    if( A.Grid() != B_MC_STAR.Grid() ||
        B_MC_STAR.Grid() != BAdjOrTrans_STAR_MR.Grid() ||
        BAdjOrTrans_STAR_MR.Grid() != Z_MC_STAR.Grid() ||
        Z_MC_STAR.Grid() != Z_MR_STAR.Grid() )
        throw std::logic_error
        ("{A,B,Z} must be distributed over the same grid");
    if( A.Height() != A.Width() ||
        A.Height() != B_MC_STAR.Height() ||
        A.Height() != BAdjOrTrans_STAR_MR.Width() ||
        A.Height() != Z_MC_STAR.Height() ||
        A.Height() != Z_MR_STAR.Height() ||
        B_MC_STAR.Width() != BAdjOrTrans_STAR_MR.Height() ||
        BAdjOrTrans_STAR_MR.Height() != Z_MC_STAR.Width() ||
        Z_MC_STAR.Width() != Z_MR_STAR.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal LocalSymmetricAccumulateLU: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B[MC,* ] ~ " << B_MC_STAR.Height() << " x "
                               << B_MC_STAR.Width() << "\n"
            << "  B^H/T[* ,MR] ~ " << BAdjOrTrans_STAR_MR.Height() << " x "
                                   << BAdjOrTrans_STAR_MR.Width() << "\n"
            << "  Z[MC,* ] ~ " << Z_MC_STAR.Height() << " x "
                               << Z_MC_STAR.Width() << "\n"
            << "  Z[MR,* ] ` " << Z_MR_STAR.Height() << " x "
                               << Z_MR_STAR.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
    if( B_MC_STAR.ColAlignment() != A.ColAlignment() ||
        BAdjOrTrans_STAR_MR.RowAlignment() != A.RowAlignment() ||
        Z_MC_STAR.ColAlignment() != A.ColAlignment() ||
        Z_MR_STAR.ColAlignment() != A.RowAlignment() )
        throw std::logic_error("Partial matrix distributions are misaligned");
#endif
    const Grid& g = A.Grid();

    DistMatrix<T>
        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<T> D11(g);

    DistMatrix<T,MC,STAR>
        BT_MC_STAR(g),  B0_MC_STAR(g),
        BB_MC_STAR(g),  B1_MC_STAR(g),
                        B2_MC_STAR(g);

    DistMatrix<T,STAR,MR>
        BLAdjOrTrans_STAR_MR(g), BRAdjOrTrans_STAR_MR(g),
        B0AdjOrTrans_STAR_MR(g), B1AdjOrTrans_STAR_MR(g), 
        B2AdjOrTrans_STAR_MR(g);

    DistMatrix<T,MC,STAR>
        ZT_MC_STAR(g),  Z0_MC_STAR(g),
        ZB_MC_STAR(g),  Z1_MC_STAR(g),
                        Z2_MC_STAR(g);

    DistMatrix<T,MR,STAR>
        ZT_MR_STAR(g),  Z0_MR_STAR(g),
        ZB_MR_STAR(g),  Z1_MR_STAR(g),
                        Z2_MR_STAR(g);

    const int ratio = std::max( g.Height(), g.Width() );
    PushBlocksizeStack( ratio*Blocksize() );

    LockedPartitionDownDiagonal
    ( A, ATL, ATR,
         ABL, ABR, 0 );
    LockedPartitionDown
    ( B_MC_STAR, BT_MC_STAR,
                 BB_MC_STAR, 0 );
    LockedPartitionRight
    ( BAdjOrTrans_STAR_MR, BLAdjOrTrans_STAR_MR, BRAdjOrTrans_STAR_MR, 0 );
    PartitionDown
    ( Z_MC_STAR, ZT_MC_STAR,
                 ZB_MC_STAR, 0 );
    PartitionDown
    ( Z_MR_STAR, ZT_MR_STAR,
                 ZB_MR_STAR, 0 );
    while( ATL.Height() < A.Height() )
    {
        LockedRepartitionDownDiagonal
        ( ATL, /**/ ATR,  A00, /**/ A01, A02,
          /************/ /******************/
               /**/       A10, /**/ A11, A12,
          ABL, /**/ ABR,  A20, /**/ A21, A22 );

        LockedRepartitionDown
        ( BT_MC_STAR,  B0_MC_STAR,
         /**********/ /**********/
                       B1_MC_STAR,
          BB_MC_STAR,  B2_MC_STAR );

        LockedRepartitionRight
        ( BLAdjOrTrans_STAR_MR, /**/ BRAdjOrTrans_STAR_MR,
          B0AdjOrTrans_STAR_MR, /**/ B1AdjOrTrans_STAR_MR, 
                                     B2AdjOrTrans_STAR_MR );

        RepartitionDown
        ( ZT_MC_STAR,  Z0_MC_STAR,
         /**********/ /**********/
                       Z1_MC_STAR,
          ZB_MC_STAR,  Z2_MC_STAR );

        RepartitionDown
        ( ZT_MR_STAR,  Z0_MR_STAR,
         /**********/ /**********/
                       Z1_MR_STAR,
          ZB_MR_STAR,  Z2_MR_STAR );

        D11.AlignWith( A11 );
        //--------------------------------------------------------------------//
        D11 = A11;
        MakeTrapezoidal( LEFT, UPPER, 0, D11 );
        LocalGemm
        ( NORMAL, orientation, 
          alpha, D11, B1AdjOrTrans_STAR_MR, T(1), Z1_MC_STAR );
        MakeTrapezoidal( LEFT, UPPER, 1, D11 );

        LocalGemm
        ( orientation, NORMAL, alpha, D11, B1_MC_STAR, T(1), Z1_MR_STAR );

        LocalGemm
        ( NORMAL, orientation, 
          alpha, A12, B2AdjOrTrans_STAR_MR, T(1), Z1_MC_STAR );

        LocalGemm
        ( orientation, NORMAL, alpha, A12, B1_MC_STAR, T(1), Z2_MR_STAR );
        //--------------------------------------------------------------------//
        D11.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( ATL, /**/ ATR,  A00, A01, /**/ A02,
               /**/       A10, A11, /**/ A12,
         /*************/ /******************/
          ABL, /**/ ABR,  A20, A21, /**/ A22 );

        SlideLockedPartitionDown
        ( BT_MC_STAR,  B0_MC_STAR,
                       B1_MC_STAR,
         /**********/ /**********/
          BB_MC_STAR,  B2_MC_STAR );

        SlideLockedPartitionRight
        ( BLAdjOrTrans_STAR_MR,                       /**/ BRAdjOrTrans_STAR_MR,
          B0AdjOrTrans_STAR_MR, B1AdjOrTrans_STAR_MR, /**/ B2AdjOrTrans_STAR_MR         );

        SlidePartitionDown
        ( ZT_MC_STAR,  Z0_MC_STAR,
                       Z1_MC_STAR,
         /**********/ /**********/
          ZB_MC_STAR,  Z2_MC_STAR );

        SlidePartitionDown
        ( ZT_MR_STAR,  Z0_MR_STAR,
                       Z1_MR_STAR,
         /**********/ /**********/
          ZB_MR_STAR,  Z2_MR_STAR );
    }
    PopBlocksizeStack();
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #23
0
파일: NNTN.hpp 프로젝트: jimgoo/Elemental
inline void
Trr2kNNTN
( UpperOrLower uplo,
  Orientation orientationOfC,
  T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
           const DistMatrix<T>& C, const DistMatrix<T>& D,
  T beta,        DistMatrix<T>& E )
{
#ifndef RELEASE
    PushCallStack("internal::Trr2kNNTN");
    if( E.Height() != E.Width()  || A.Width()  != C.Height() ||
        A.Height() != E.Height() || C.Width()  != E.Height() ||
        B.Width()  != E.Width()  || D.Width()  != E.Width()  ||
        A.Width()  != B.Height() || C.Height() != D.Height() )
        throw std::logic_error("Nonconformal Trr2kNNTN");
#endif
    const Grid& g = E.Grid();

    DistMatrix<T> AL(g), AR(g),
                  A0(g), A1(g), A2(g);
    DistMatrix<T> BT(g),  B0(g),
                  BB(g),  B1(g),
                          B2(g);

    DistMatrix<T> CT(g),  C0(g),
                  CB(g),  C1(g),
                          C2(g);
    DistMatrix<T> DT(g),  D0(g),
                  DB(g),  D1(g),
                          D2(g);

    DistMatrix<T,MC,  STAR> A1_MC_STAR(g);
    DistMatrix<T,MR,  STAR> B1Trans_MR_STAR(g);
    DistMatrix<T,STAR,MC  > C1_STAR_MC(g);
    DistMatrix<T,MR,  STAR> D1Trans_MR_STAR(g);

    A1_MC_STAR.AlignWith( E );
    B1Trans_MR_STAR.AlignWith( E );
    C1_STAR_MC.AlignWith( E );
    D1Trans_MR_STAR.AlignWith( E );

    LockedPartitionRight( A, AL, AR, 0 );
    LockedPartitionDown
    ( B, BT,
         BB, 0 );
    LockedPartitionDown
    ( C, CT,
         CB, 0 );
    LockedPartitionDown
    ( D, DT,
         DB, 0 );
    while( AL.Width() < A.Width() )
    {
        LockedRepartitionRight
        ( AL, /**/ AR,
          A0, /**/ A1, A2 );
        LockedRepartitionDown
        ( BT,  B0,
         /**/ /**/
               B1,
          BB,  B2 );
        LockedRepartitionDown
        ( CT,  C0,
         /**/ /**/
               C1,
          CB,  C2 );
        LockedRepartitionDown
        ( DT,  D0,
         /**/ /**/
               D1,
          DB,  D2 );

        //--------------------------------------------------------------------//
        A1_MC_STAR = A1;
        C1_STAR_MC = C1;
        B1Trans_MR_STAR.TransposeFrom( B1 );
        D1Trans_MR_STAR.TransposeFrom( D1 );
        LocalTrr2k
        ( uplo, TRANSPOSE, orientationOfC, TRANSPOSE,
          alpha, A1_MC_STAR, B1Trans_MR_STAR, 
                 C1_STAR_MC, D1Trans_MR_STAR,
          beta,  E );
        //--------------------------------------------------------------------//

        SlideLockedPartitionDown
        ( DT,  D0,
               D1,
         /**/ /**/
          DB,  D2 );
        SlideLockedPartitionDown
        ( CT,  C0,
               C1,
         /**/ /**/
          CB,  C2 );
        SlideLockedPartitionDown
        ( BT,  B0,
               B1,
         /**/ /**/
          BB,  B2 );
        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #24
0
파일: NNNT.hpp 프로젝트: mcg1969/Elemental
void Trr2kNNNT
( UpperOrLower uplo,
  Orientation orientationOfD,
  T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
           const DistMatrix<T>& C, const DistMatrix<T>& D,
  T beta,        DistMatrix<T>& E )
{
#ifndef RELEASE
    PushCallStack("internal::Trr2kNNNT");
    if( E.Height() != E.Width()  || A.Width()  != C.Width()  ||
        A.Height() != E.Height() || C.Height() != E.Height() ||
        B.Width()  != E.Width()  || D.Height() != E.Width()  ||
        A.Width()  != B.Height() || C.Width()  != D.Width() )
        throw std::logic_error("Nonconformal Trr2kNNNT");
#endif
    const Grid& g = E.Grid();

    DistMatrix<T> AL(g), AR(g),
                  A0(g), A1(g), A2(g);
    DistMatrix<T> BT(g),  B0(g),
                  BB(g),  B1(g),
                          B2(g);

    DistMatrix<T> CL(g), CR(g),
                  C0(g), C1(g), C2(g);
    DistMatrix<T> DL(g), DR(g),
                  D0(g), D1(g), D2(g);

    DistMatrix<T,MC,  STAR> A1_MC_STAR(g);
    DistMatrix<T,MR,  STAR> B1Trans_MR_STAR(g);
    DistMatrix<T,MC,  STAR> C1_MC_STAR(g);
    DistMatrix<T,VR,  STAR> D1_VR_STAR(g);
    DistMatrix<T,STAR,MR  > D1AdjOrTrans_STAR_MR(g);

    A1_MC_STAR.AlignWith( E );
    B1Trans_MR_STAR.AlignWith( E );
    C1_MC_STAR.AlignWith( E );
    D1_VR_STAR.AlignWith( E );
    D1AdjOrTrans_STAR_MR.AlignWith( E );

    LockedPartitionRight( A, AL, AR, 0 );
    LockedPartitionDown
    ( B, BT,
         BB, 0 );
    LockedPartitionRight( C, CL, CR, 0 );
    LockedPartitionRight( D, DL, DR, 0 );
    while( AL.Width() < A.Width() )
    {
        LockedRepartitionRight
        ( AL, /**/ AR,
          A0, /**/ A1, A2 );
        LockedRepartitionDown
        ( BT,  B0,
         /**/ /**/
               B1,
          BB,  B2 );
        LockedRepartitionRight
        ( CL, /**/ CR,
          C0, /**/ C1, C2 );
        LockedRepartitionRight
        ( CL, /**/ CR,
          C0, /**/ C1, C2 );

        //--------------------------------------------------------------------//
        A1_MC_STAR = A1;
        C1_MC_STAR = C1;
        B1Trans_MR_STAR.TransposeFrom( B1 );
        D1_VR_STAR = D1;
        if( orientationOfD == ADJOINT )
            D1AdjOrTrans_STAR_MR.AdjointFrom( D1_VR_STAR );
        else
            D1AdjOrTrans_STAR_MR.TransposeFrom( D1_VR_STAR );
        LocalTrr2k
        ( uplo, TRANSPOSE, 
          alpha, A1_MC_STAR, B1Trans_MR_STAR, 
                 C1_MC_STAR, D1AdjOrTrans_STAR_MR,
          beta,  E );
        //--------------------------------------------------------------------//

        SlideLockedPartitionRight
        ( DL,     /**/ DR,
          D0, D1, /**/ D2 );
        SlideLockedPartitionRight
        ( CL,     /**/ CR,
          C0, C1, /**/ C2 );
        SlideLockedPartitionDown
        ( BT,  B0,
               B1,
         /**/ /**/
          BB,  B2 );
        SlideLockedPartitionRight
        ( AL,     /**/ AR,
          A0, A1, /**/ A2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #25
0
파일: RUN.hpp 프로젝트: ahmadia/Elemental-1
inline void
LocalTrmmAccumulateRUN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
  const DistMatrix<T,MC,  MR  >& U,
  const DistMatrix<T,STAR,MC  >& X_STAR_MC,
        DistMatrix<T,MR,  STAR>& ZTrans_MR_STAR )
{
#ifndef RELEASE
    CallStackEntry entry("internal::LocalTrmmAccumulateRUN");
    if( U.Grid() != X_STAR_MC.Grid() ||
        X_STAR_MC.Grid() != ZTrans_MR_STAR.Grid() )
        throw std::logic_error
        ("{U,X,Z} must be distributed over the same grid");
    if( U.Height() != U.Width() ||
        U.Height() != X_STAR_MC.Width() ||
        U.Height() != ZTrans_MR_STAR.Height() )
    {
        std::ostringstream msg;
        msg << "Nonconformal LocalTrmmAccumulateRUN: \n"
            << "  U ~ " << U.Height() << " x " << U.Width() << "\n"
            << "  X[* ,MC] ~ " << X_STAR_MC.Height() << " x "
                               << X_STAR_MC.Width() << "\n"
            << "  Z^H/T[MR,* ] ~ " << ZTrans_MR_STAR.Height() << " x "
                                   << ZTrans_MR_STAR.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
    if( X_STAR_MC.RowAlignment() != U.ColAlignment() ||
        ZTrans_MR_STAR.ColAlignment() != U.RowAlignment() )
        throw std::logic_error("Partial matrix distributions are misaligned");
#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);

    DistMatrix<T> D11(g);

    DistMatrix<T,STAR,MC>
        XL_STAR_MC(g), XR_STAR_MC(g),
        X0_STAR_MC(g), X1_STAR_MC(g), X2_STAR_MC(g);

    DistMatrix<T,MR,STAR>
        ZTTrans_MR_STAR(g),  Z0Trans_MR_STAR(g),
        ZBTrans_MR_STAR(g),  Z1Trans_MR_STAR(g),
                             Z2Trans_MR_STAR(g);

    const int ratio = std::max( g.Height(), g.Width() );
    PushBlocksizeStack( ratio*Blocksize() );

    LockedPartitionDownDiagonal
    ( U, UTL, UTR,
         UBL, UBR, 0 );
    LockedPartitionRight( X_STAR_MC,  XL_STAR_MC, XR_STAR_MC, 0 );
    PartitionDown
    ( ZTrans_MR_STAR, ZTTrans_MR_STAR,
                      ZBTrans_MR_STAR, 0 );
    while( UTL.Height() < U.Height() )
    {
        LockedRepartitionDownDiagonal
        ( UTL, /**/ UTR,  U00, /**/ U01, U02,
         /*************/ /******************/
               /**/       U10, /**/ U11, U12,
          UBL, /**/ UBR,  U20, /**/ U21, U22 );

        LockedRepartitionRight
        ( XL_STAR_MC, /**/ XR_STAR_MC,
          X0_STAR_MC, /**/ X1_STAR_MC, X2_STAR_MC );

        RepartitionDown
        ( ZTTrans_MR_STAR,  Z0Trans_MR_STAR,
         /***************/ /***************/
                            Z1Trans_MR_STAR,
          ZBTrans_MR_STAR,  Z2Trans_MR_STAR );

        D11.AlignWith( U11 );
        //--------------------------------------------------------------------//
        D11 = U11;
        MakeTriangular( UPPER, D11 );
        if( diag == UNIT )
            SetDiagonal( D11, T(1) );
        LocalGemm
        ( orientation, orientation,
          alpha, D11, X1_STAR_MC, T(1), Z1Trans_MR_STAR );
        LocalGemm
        ( orientation, orientation,
          alpha, U01, X0_STAR_MC, T(1), Z1Trans_MR_STAR );
        //--------------------------------------------------------------------//
        D11.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( UTL, /**/ UTR,  U00, U01, /**/ U02,
               /**/       U10, U11, /**/ U12,
         /*************/ /******************/
          UBL, /**/ UBR,  U20, U21, /**/ U22 );

        SlideLockedPartitionRight
        ( XL_STAR_MC,             /**/ XR_STAR_MC,
          X0_STAR_MC, X1_STAR_MC, /**/ X2_STAR_MC );

        SlidePartitionDown
        ( ZTTrans_MR_STAR,  Z0Trans_MR_STAR,
                            Z1Trans_MR_STAR,
         /***************/ /***************/
          ZBTrans_MR_STAR,  Z2Trans_MR_STAR );
    }
    PopBlocksizeStack();
}
예제 #26
0
파일: LL.hpp 프로젝트: jimgoo/Elemental
inline void
SymmLLA
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::SymmLLA");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
#endif
    const Grid& g = A.Grid();

    DistMatrix<T> 
        BL(g), BR(g),
        B0(g), B1(g), B2(g);

    DistMatrix<T>
        CL(g), CR(g),
        C0(g), C1(g), C2(g);

    DistMatrix<T,MC,STAR> B1_MC_STAR(g);
    DistMatrix<T,VR,STAR> B1_VR_STAR(g);
    DistMatrix<T,STAR,MR> B1Trans_STAR_MR(g);
    DistMatrix<T> Z1(g);
    DistMatrix<T,MC,STAR> Z1_MC_STAR(g);
    DistMatrix<T,MR,STAR> Z1_MR_STAR(g);
    DistMatrix<T,MR,MC  > Z1_MR_MC(g);

    B1_MC_STAR.AlignWith( A );
    B1_VR_STAR.AlignWith( A );
    B1Trans_STAR_MR.AlignWith( A );
    Z1_MC_STAR.AlignWith( A );
    Z1_MR_STAR.AlignWith( A );

    Scale( beta, C );
    LockedPartitionRight
    ( B, BL, BR, 0 );
    PartitionRight
    ( C, CL, CR, 0 );
    while( CL.Width() < C.Width() )
    {
        LockedRepartitionRight 
        ( BL, /**/ BR,
          B0, /**/ B1, B2 );

        RepartitionRight
        ( CL, /**/ CR,
          C0, /**/ C1, C2 );

        Z1.AlignWith( C1 );
        Zeros( C1.Height(), C1.Width(), Z1_MC_STAR );
        Zeros( C1.Height(), C1.Width(), Z1_MR_STAR );
        //--------------------------------------------------------------------//
        B1_MC_STAR = B1;
        B1_VR_STAR = B1_MC_STAR;
        B1Trans_STAR_MR.TransposeFrom( B1_VR_STAR );
        LocalSymmetricAccumulateLL
        ( TRANSPOSE, 
          alpha, A, B1_MC_STAR, B1Trans_STAR_MR, Z1_MC_STAR, Z1_MR_STAR );

        Z1_MR_MC.SumScatterFrom( Z1_MR_STAR );
        Z1 = Z1_MR_MC;
        Z1.SumScatterUpdate( T(1), Z1_MC_STAR );
        Axpy( T(1), Z1, C1 );
        //--------------------------------------------------------------------//
        Z1.FreeAlignments();

        SlideLockedPartitionRight
        ( BL,     /**/ BR,
          B0, B1, /**/ B2 );

        SlidePartitionRight
        ( CL,     /**/ CR,
          C0, C1, /**/ C2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
예제 #27
0
파일: RU.hpp 프로젝트: mcg1969/Elemental
inline void
LocalSymmetricAccumulateRU
( Orientation orientation, T alpha,
  const DistMatrix<T,MC,  MR  >& A,
  const DistMatrix<T,STAR,MC  >& B_STAR_MC,
  const DistMatrix<T,MR,  STAR>& BTrans_MR_STAR,
        DistMatrix<T,MC,  STAR>& ZTrans_MC_STAR,
        DistMatrix<T,MR,  STAR>& ZTrans_MR_STAR )
{
#ifndef RELEASE
    PushCallStack("internal::LocalSymmetricAccumulateRU");
    if( A.Grid() != B_STAR_MC.Grid() ||
        B_STAR_MC.Grid() != BTrans_MR_STAR.Grid() ||
        BTrans_MR_STAR.Grid() != ZTrans_MC_STAR.Grid() ||
        ZTrans_MC_STAR.Grid() != ZTrans_MR_STAR.Grid() )
        throw std::logic_error
        ("{A,B,C} must be distributed over the same grid");
    if( A.Height() != A.Width() ||
        A.Height() != B_STAR_MC.Width() ||
        A.Height() != BTrans_MR_STAR.Height() ||
        A.Height() != ZTrans_MC_STAR.Height() ||
        A.Height() != ZTrans_MR_STAR.Height() ||
        B_STAR_MC.Height() != BTrans_MR_STAR.Width() ||
        BTrans_MR_STAR.Width() != ZTrans_MC_STAR.Width() ||
        ZTrans_MC_STAR.Width() != ZTrans_MR_STAR.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal LocalSymmetricAccumulateRU: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B[* ,MC] ~ " << B_STAR_MC.Height() << " x "
                               << B_STAR_MC.Width() << "\n"
            << "  B^H/T[MR,* ] ~ " << BTrans_MR_STAR.Height() << " x "
                                   << BTrans_MR_STAR.Width() << "\n"
            << "  Z^H/T[MC,* ] ~ " << ZTrans_MC_STAR.Height() << " x "
                                   << ZTrans_MC_STAR.Width() << "\n"
            << "  Z^H/T[MR,* ] ~ " << ZTrans_MR_STAR.Height() << " x "
                                   << ZTrans_MR_STAR.Width() << "\n";
        throw std::logic_error( msg.str().c_str() );
    }
    if( B_STAR_MC.RowAlignment() != A.ColAlignment() ||
        BTrans_MR_STAR.ColAlignment() != A.RowAlignment() ||
        ZTrans_MC_STAR.ColAlignment() != A.ColAlignment() ||
        ZTrans_MR_STAR.ColAlignment() != A.RowAlignment() )
        throw std::logic_error("Partial matrix distributions are misaligned");
#endif
    const Grid& g = A.Grid();

    // Matrix views
    DistMatrix<T>
        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<T> D11(g);

    DistMatrix<T,STAR,MC>
        BL_STAR_MC(g), BR_STAR_MC(g),
        B0_STAR_MC(g), B1_STAR_MC(g), B2_STAR_MC(g);

    DistMatrix<T,MR,STAR>
        BTTrans_MR_STAR(g),  B0Trans_MR_STAR(g),
        BBTrans_MR_STAR(g),  B1Trans_MR_STAR(g),
                             B2Trans_MR_STAR(g);

    DistMatrix<T,MC,STAR>
        ZTTrans_MC_STAR(g),  Z0Trans_MC_STAR(g),
        ZBTrans_MC_STAR(g),  Z1Trans_MC_STAR(g),
                             Z2Trans_MC_STAR(g);

    DistMatrix<T,MR,STAR>
        ZBTrans_MR_STAR(g),  Z0Trans_MR_STAR(g),
        ZTTrans_MR_STAR(g),  Z1Trans_MR_STAR(g),
                             Z2Trans_MR_STAR(g);

    const int ratio = std::max( g.Height(), g.Width() );
    PushBlocksizeStack( ratio*Blocksize() );

    LockedPartitionDownDiagonal
    ( A, ATL, ATR,
         ABL, ABR, 0 );
    LockedPartitionRight( B_STAR_MC,  BL_STAR_MC, BR_STAR_MC, 0 );
    LockedPartitionDown
    ( BTrans_MR_STAR, BTTrans_MR_STAR,
                      BBTrans_MR_STAR, 0 );
    PartitionDown
    ( ZTrans_MC_STAR, ZTTrans_MC_STAR,
                      ZBTrans_MC_STAR, 0 );
    PartitionDown
    ( ZTrans_MR_STAR, ZTTrans_MR_STAR,
                      ZBTrans_MR_STAR, 0 );
    while( ATL.Height() < A.Height() )
    {
        LockedRepartitionDownDiagonal
        ( ATL, /**/ ATR,  A00, /**/ A01, A02,
         /*************/ /******************/
               /**/       A10, /**/ A11, A12,
          ABL, /**/ ABR,  A20, /**/ A21, A22 );

        LockedRepartitionRight
        ( BL_STAR_MC, /**/ BR_STAR_MC,
          B0_STAR_MC, /**/ B1_STAR_MC, B2_STAR_MC );

        LockedRepartitionDown
        ( BTTrans_MR_STAR,  B0Trans_MR_STAR,
         /***************/ /***************/
                            B1Trans_MR_STAR,
          BBTrans_MR_STAR,  B2Trans_MR_STAR );

        RepartitionDown
        ( ZTTrans_MC_STAR,  Z0Trans_MC_STAR,
         /***************/ /***************/
                            Z1Trans_MC_STAR,
          ZBTrans_MC_STAR,  Z2Trans_MC_STAR );

        RepartitionDown
        ( ZTTrans_MR_STAR,  Z0Trans_MR_STAR,
         /***************/ /***************/
                            Z1Trans_MR_STAR,
          ZBTrans_MR_STAR,  Z2Trans_MR_STAR );

        D11.AlignWith( A11 );
        //--------------------------------------------------------------------//
        D11 = A11;
        MakeTriangular( UPPER, D11 );
        LocalGemm
        ( orientation, orientation,
          alpha, D11, B1_STAR_MC, T(1), Z1Trans_MR_STAR );
        SetDiagonal( D11, T(0) );

        LocalGemm
        ( NORMAL, NORMAL, alpha, D11, B1Trans_MR_STAR, T(1), Z1Trans_MC_STAR );

        LocalGemm
        ( orientation, orientation, 
          alpha, A12, B1_STAR_MC, T(1), Z2Trans_MR_STAR );

        LocalGemm
        ( NORMAL, NORMAL, alpha, A12, B2Trans_MR_STAR, T(1), Z1Trans_MC_STAR );
        //--------------------------------------------------------------------//
        D11.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( ATL, /**/ ATR,  A00, A01, /**/ A02,
               /**/       A10, A11, /**/ A12,
         /*************/ /******************/
          ABL, /**/ ABR,  A20, A21, /**/ A22 );

        SlideLockedPartitionRight
        ( BL_STAR_MC,             /**/ BR_STAR_MC,
          B0_STAR_MC, B1_STAR_MC, /**/ B2_STAR_MC );

        SlideLockedPartitionDown
        ( BTTrans_MR_STAR,  B0Trans_MR_STAR,
                            B1Trans_MR_STAR,
         /***************/ /***************/
          BBTrans_MR_STAR,  B2Trans_MR_STAR );

        SlidePartitionDown
        ( ZTTrans_MC_STAR,  Z0Trans_MC_STAR,
                            Z1Trans_MC_STAR,
         /***************/ /***************/
          ZBTrans_MC_STAR,  Z2Trans_MC_STAR );

        SlidePartitionDown
        ( ZTTrans_MR_STAR,  Z0Trans_MR_STAR,
                            Z1Trans_MR_STAR,
         /***************/ /***************/
          ZBTrans_MR_STAR,  Z2Trans_MR_STAR );
    }
    PopBlocksizeStack();
#ifndef RELEASE
    PopCallStack();
#endif
}