Example #1
0
void SUMMA_NTC
( Orientation orientB,
  T alpha,
  const AbstractDistMatrix<T>& APre,
  const AbstractDistMatrix<T>& BPre,
        AbstractDistMatrix<T>& CPre )
{
    EL_DEBUG_CSE
    const Int sumDim = APre.Width();
    const Int bsize = Blocksize();
    const Grid& g = APre.Grid();
    const bool conjugate = ( orientB == ADJOINT );

    DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
    DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
    DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
    auto& A = AProx.GetLocked();
    auto& B = BProx.GetLocked();
    auto& C = CProx.Get();

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

    A1_MC_STAR.AlignWith( C );
    B1_VR_STAR.AlignWith( C );
    B1Trans_STAR_MR.AlignWith( C );

    for( Int k=0; k<sumDim; k+=bsize )
    {
        const Int nb = Min(bsize,sumDim-k);
        auto A1 = A( ALL, IR(k,k+nb) );
        auto B1 = B( ALL, IR(k,k+nb) );

        A1_MC_STAR = A1;
        B1_VR_STAR = B1;
        Transpose( B1_VR_STAR, B1Trans_STAR_MR, conjugate );

        // C[MC,MR] += alpha A1[MC,*] (B1[MR,*])^T
        LocalGemm
        ( NORMAL, NORMAL, alpha, A1_MC_STAR, B1Trans_STAR_MR, T(1), C );
    }
}
Example #2
0
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
}
Example #3
0
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
}
Example #4
0
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
}
Example #5
0
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
}
Example #6
0
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
}
Example #7
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 );
    }
}
Example #8
0
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
}
Example #9
0
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 );
    }
}