Ejemplo n.º 1
0
void LT_C
( T alpha,
  const AbstractDistMatrix<T>& APre,
  AbstractDistMatrix<T>& CPre,
  bool conjugate=false )
{
    EL_DEBUG_CSE
    const Int r = APre.Height();
    const Int bsize = Blocksize();
    const Grid& g = APre.Grid();
    const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );

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

    // Temporary distributions
    DistMatrix<T,MR,  STAR> A1Trans_MR_STAR(g);
    DistMatrix<T,STAR,VR  > A1_STAR_VR(g);
    DistMatrix<T,STAR,MC  > A1_STAR_MC(g);

    A1Trans_MR_STAR.AlignWith( C );
    A1_STAR_MC.AlignWith( C );

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

        Transpose( A1, A1Trans_MR_STAR );
        Transpose( A1Trans_MR_STAR, A1_STAR_VR );
        A1_STAR_MC = A1_STAR_VR;

        LocalTrrk
        ( LOWER, orientation, TRANSPOSE,
          alpha, A1_STAR_MC, A1Trans_MR_STAR, T(1), C );
    }
}
Ejemplo n.º 2
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
}
Ejemplo n.º 3
0
inline void
Syr2kUT
( T alpha, const DistMatrix<T>& A, const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C,
  bool conjugate=false )
{
#ifndef RELEASE
    CallStackEntry entry("internal::Syr2kUT");
    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.Width() != C.Height() || 
        A.Width() != C.Width()  ||
        B.Width() != C.Height() ||
        B.Width() != C.Width()  ||
        A.Height() != B.Height()  )
    {
        std::ostringstream msg;
        msg << "Nonconformal Syr2kUT:\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();
    const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );

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

    // Temporary distributions
    DistMatrix<T,MR,  STAR> A1Trans_MR_STAR(g);
    DistMatrix<T,MR,  STAR> B1Trans_MR_STAR(g);
    DistMatrix<T,STAR,VR  > A1_STAR_VR(g);
    DistMatrix<T,STAR,VR  > B1_STAR_VR(g);
    DistMatrix<T,STAR,MC  > A1_STAR_MC(g);
    DistMatrix<T,STAR,MC  > B1_STAR_MC(g);

    A1Trans_MR_STAR.AlignWith( C );
    B1Trans_MR_STAR.AlignWith( C );
    A1_STAR_MC.AlignWith( C );
    B1_STAR_MC.AlignWith( C );

    // Start the algorithm
    ScaleTrapezoid( beta, LEFT, UPPER, 0, C );
    LockedPartitionDown
    ( A, AT, 
         AB, 0 );
    LockedPartitionDown
    ( B, BT,
         BB, 0 );
    while( AB.Height() > 0 )
    {
        LockedRepartitionDown
        ( AT,  A0,
         /**/ /**/
               A1,
          AB,  A2 );

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

        //--------------------------------------------------------------------//
        A1Trans_MR_STAR.TransposeFrom( A1 );
        A1_STAR_VR.TransposeFrom( A1Trans_MR_STAR );
        A1_STAR_MC = A1_STAR_VR;

        B1Trans_MR_STAR.TransposeFrom( B1 );
        B1_STAR_VR.TransposeFrom( B1Trans_MR_STAR );
        B1_STAR_MC = B1_STAR_VR;

        LocalTrr2k
        ( UPPER, orientation, TRANSPOSE, orientation, TRANSPOSE, 
          alpha, A1_STAR_MC, B1Trans_MR_STAR,
                 B1_STAR_MC, A1Trans_MR_STAR,
          T(1),  C );
        //--------------------------------------------------------------------//

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

        SlideLockedPartitionDown
        ( BT,  B0,
               B1,
         /**/ /**/
          BB,  B2 );
    }
}
Ejemplo n.º 4
0
inline void
HerkLC( T alpha, const DistMatrix<T>& A, T beta, DistMatrix<T>& C )
{
#ifndef RELEASE
    PushCallStack("internal::HerkLC");
    if( A.Grid() != C.Grid() )
        throw std::logic_error
        ("A and C must be distributed over the same grid");
    if( A.Width() != C.Height() || A.Width() != C.Width() )
    {
        std::ostringstream msg;
        msg << "Nonconformal HerkLC:\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> AT(g),  A0(g),
                  AB(g),  A1(g),
                          A2(g);

    // Temporary distributions
    DistMatrix<T,MR,  STAR> A1Trans_MR_STAR(g);
    DistMatrix<T,STAR,VR  > A1_STAR_VR(g);
    DistMatrix<T,STAR,MC  > A1_STAR_MC(g);

    A1Trans_MR_STAR.AlignWith( C );
    A1_STAR_MC.AlignWith( C );

    // Start the algorithm
    ScaleTrapezoid( beta, LEFT, LOWER, 0, C );
    LockedPartitionDown
    ( A, AT, 
         AB, 0 );
    while( AB.Height() > 0 )
    {
        LockedRepartitionDown
        ( AT,  A0,
         /**/ /**/
               A1,
          AB,  A2 );

        //--------------------------------------------------------------------//
        A1Trans_MR_STAR.TransposeFrom( A1 );
        A1_STAR_VR.TransposeFrom( A1Trans_MR_STAR );
        A1_STAR_MC = A1_STAR_VR;

        LocalTrrk
        ( LOWER, ADJOINT, TRANSPOSE,
          alpha, A1_STAR_MC, A1Trans_MR_STAR, T(1), C );
        //--------------------------------------------------------------------//

        SlideLockedPartitionDown
        ( AT,  A0,
               A1,
         /**/ /**/
          AB,  A2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}