C++ DistMatrix::AlignWith方法代碼示例

/*
 * Distributes A in such a way that
 *   Layer 0 <- A(:, 0:(n/h - 1))
 *   Layer 1 <- A(:, (n/h):(2n/h - 1))
 *     .
 *     .
 *     .
 *   Layer h-1 <- A(:, ((h-1)n/h):n)
 */
void DistributeCols
( const mpi::Comm& depthComm,
  const DistMatrix<double,MC,MR>& A, 
        DistMatrix<double,MC,MR>& B )
{
    const Grid& meshGrid = A.Grid();
    const int meshSize = meshGrid.Size();
    const int depthSize = mpi::CommSize( depthComm );
    const int depthRank = mpi::CommRank( depthComm );

    const int sendCount = A.LocalHeight()*A.LocalWidth();
    const int recvCount = sendCount / depthSize;

    // For now, we will make B as large as A...
    // TODO: NOT DO THIS
    if( A.LocalHeight() != A.LocalLDim() )
        throw std::logic_error("Local height did not match local ldim");
    B.Empty();
    B.AlignWith( A );
    Zeros( A.Height(), A.Width(), B );

    // Scatter
    const int localColOffset = (A.LocalWidth()/depthSize)*depthRank;
    mpi::Scatter
    ( A.LockedLocalBuffer(), recvCount, 
      B.LocalBuffer(0,localColOffset), recvCount, 0, depthComm );
}

void RowMaxNorms
( const DistMatrix<F,U,V>& A, DistMatrix<Base<F>,U,STAR>& norms )
{
    DEBUG_CSE
    norms.AlignWith( A );
    norms.Resize( A.Height(), 1 );
    RowMaxNorms( A.LockedMatrix(), norms.Matrix() );
    AllReduce( norms, A.RowComm(), mpi::MAX );
}

void RowTwoNorms
( const DistMatrix<F,U,V>& A, DistMatrix<Base<F>,U,STAR>& norms )
{
    DEBUG_CSE
    norms.AlignWith( A );
    norms.Resize( A.Height(), 1 );
    if( A.Width() == 0 )
    {
        Zero( norms );
        return;
    }
    RowTwoNormsHelper( A.LockedMatrix(), norms.Matrix(), A.RowComm() );
}

inline void HermitianSVD
( UpperOrLower uplo, DistMatrix<F>& A, 
  DistMatrix<BASE(F),VR,STAR>& s, DistMatrix<F>& U, DistMatrix<F>& V )
{
#ifndef RELEASE
    CallStackEntry entry("HermitianSVD");
#endif
#ifdef HAVE_PMRRR
    typedef BASE(F) R;

    // Grab an eigenvalue decomposition of A
    HermitianEig( uplo, A, s, V );

    // Redistribute the singular values into an [MR,* ] distribution
    const Grid& grid = A.Grid();
    DistMatrix<R,MR,STAR> s_MR_STAR( grid );
    s_MR_STAR.AlignWith( V.DistData() );
    s_MR_STAR = s;

    // Set the singular values to the absolute value of the eigenvalues
    const Int numLocalVals = s.LocalHeight();
    for( Int iLoc=0; iLoc<numLocalVals; ++iLoc )
    {
        const R sigma = s.GetLocal(iLoc,0);
        s.SetLocal(iLoc,0,Abs(sigma));
    }

    // Copy V into U (flipping the sign as necessary)
    U.AlignWith( V );
    U.ResizeTo( V.Height(), V.Width() );
    const Int localHeight = V.LocalHeight();
    const Int localWidth = V.LocalWidth();
    for( Int jLoc=0; jLoc<localWidth; ++jLoc )
    {
        const R sigma = s_MR_STAR.GetLocal( jLoc, 0 );
        F* UCol = U.Buffer( 0, jLoc );
        const F* VCol = V.LockedBuffer( 0, jLoc );
        if( sigma >= 0 )
            for( Int iLoc=0; iLoc<localHeight; ++iLoc )
                UCol[iLoc] = VCol[iLoc];
        else
            for( Int iLoc=0; iLoc<localHeight; ++iLoc )
                UCol[iLoc] = -VCol[iLoc];
    }
#else
    U = A;
    MakeHermitian( uplo, U );
    SVD( U, s, V );
#endif // ifdef HAVE_PMRRR
}

inline void HermitianSVD
( UpperOrLower uplo, 
  DistMatrix<F>& A, DistMatrix<typename Base<F>::type,VR,STAR>& s, 
  DistMatrix<F>& U, DistMatrix<F>& V )
{
#ifndef RELEASE
    PushCallStack("HermitianSVD");
#endif
    typedef typename Base<F>::type R;

    // Grab an eigenvalue decomposition of A
    HermitianEig( uplo, A, s, V ); 
    
    // Redistribute the singular values into an [MR,* ] distribution
    const Grid& grid = A.Grid();
    DistMatrix<R,MR,STAR> s_MR_STAR( grid );
    s_MR_STAR.AlignWith( V );
    s_MR_STAR = s;

    // Set the singular values to the absolute value of the eigenvalues
    const int numLocalVals = s.LocalHeight();
    for( int iLocal=0; iLocal<numLocalVals; ++iLocal )
    {
        const R sigma = s.GetLocal(iLocal,0);
        s.SetLocal(iLocal,0,Abs(sigma));
    }

    // Copy V into U (flipping the sign as necessary)
    U.AlignWith( V );
    U.ResizeTo( V.Height(), V.Width() );
    const int localHeight = V.LocalHeight();
    const int localWidth = V.LocalWidth();
    for( int jLocal=0; jLocal<localWidth; ++jLocal )
    {
        const R sigma = s_MR_STAR.GetLocal( jLocal, 0 );
        F* UCol = U.LocalBuffer( 0, jLocal );
        const F* VCol = V.LockedLocalBuffer( 0, jLocal );
        if( sigma >= 0 )
            for( int iLocal=0; iLocal<localHeight; ++iLocal )
                UCol[iLocal] = VCol[iLocal];
        else
            for( int iLocal=0; iLocal<localHeight; ++iLocal )
                UCol[iLocal] = -VCol[iLocal];
    }
#ifndef RELEASE
    PopCallStack();
#endif
}

// Create a new set of distributed matrices, so that, 
//    if depthRank == 0, B = A,
//    otherwise,         B = 0.
void CopyOrReset
( const DistMatrix<double,MC,MR>& A, 
        DistMatrix<double,MC,MR>& B )
{
    const int rank = mpi::CommRank( mpi::COMM_WORLD );
    const Grid& meshGrid = A.Grid();
    const int meshSize = meshGrid.Size();
    const int depthRank = rank / meshSize;

    //Layer 0
    if( depthRank == 0 )
        B = A;
    else
    {
        B.AlignWith( A );
        Zeros( A.Height(), A.Width(), B );
    }
}

void ScaLAPACKHelper
( DistMatrix<F,MC,MR,BLOCK>& A,
  DistMatrix<F,MR,STAR,BLOCK>& householderScalars )
{
    EL_DEBUG_CSE
    AssertScaLAPACKSupport();
#ifdef EL_HAVE_SCALAPACK
    const Int m = A.Height();
    const Int n = A.Width();
    const Int minDim = Min(m,n);
    householderScalars.AlignWith( A );
    householderScalars.Resize( minDim, 1 );

    auto descA = FillDesc( A );
    scalapack::QR
    ( m, n, A.Buffer(), descA.data(), householderScalars.Buffer() );
#endif
}

void QR
( DistMatrix<F,MC,MR,BLOCK>& A,
  DistMatrix<F,MR,STAR,BLOCK>& phase )
{
    DEBUG_CSE
    AssertScaLAPACKSupport();
#ifdef EL_HAVE_SCALAPACK
    const Int m = A.Height();
    const Int n = A.Width();
    const Int minDim = Min(m,n);
    phase.AlignWith( A );
    phase.Resize( minDim, 1 ); 

    const int bHandle = blacs::Handle( A );
    const int context = blacs::GridInit( bHandle, A );
    auto descA = FillDesc( A, context );

    scalapack::QR( m, n, A.Buffer(), descA.data(), phase.Buffer() );

    blacs::FreeGrid( context );
    blacs::FreeHandle( bHandle );
#endif
}

inline void
Var3( Orientation orientation, DistMatrix<F>& A, DistMatrix<F,MC,STAR>& d )
{
#ifndef RELEASE
    PushCallStack("ldl::Var3");
    if( orientation == NORMAL )
        throw std::logic_error("Can only perform LDL^T and LDL^H");
    if( A.Height() != A.Width() )
        throw std::logic_error("A must be square");
    if( A.Grid() != d.Grid() )
        throw std::logic_error("A and d must use the same grid");
    if( d.Viewing() && (d.Height() != A.Height() || d.Width() != 1) )
        throw std::logic_error
        ("d must be a column vector of the same height as A");
    if( d.Viewing() && d.ColAlignment() != A.ColAlignment() )
        throw std::logic_error("d must be aligned with A");
#endif
    const Grid& g = A.Grid();
    if( !d.Viewing() )
    {
        d.AlignWith( A );
        d.ResizeTo( A.Height(), 1 );
    }

    // Matrix views
    DistMatrix<F>
        ATL(g), ATR(g),  A00(g), A01(g), A02(g),
        ABL(g), ABR(g),  A10(g), A11(g), A12(g),
                         A20(g), A21(g), A22(g);
    DistMatrix<F,MC,STAR>
        dT(g),  d0(g),
        dB(g),  d1(g),
                d2(g);

    // Temporary matrices
    DistMatrix<F,STAR,STAR> A11_STAR_STAR(g);
    DistMatrix<F,STAR,STAR> d1_STAR_STAR(g);
    DistMatrix<F,VC,  STAR> A21_VC_STAR(g);
    DistMatrix<F,VR,  STAR> A21_VR_STAR(g);
    DistMatrix<F,STAR,MC  > S21Trans_STAR_MC(g);
    DistMatrix<F,STAR,MR  > A21AdjOrTrans_STAR_MR(g);

    const bool conjugate = ( orientation == ADJOINT );

    // Start the algorithm
    PartitionDownDiagonal
    ( A, ATL, ATR,
         ABL, ABR, 0 );
    PartitionDown
    ( d, dT,
         dB, 0 );
    while( ABR.Height() > 0 )
    {
        RepartitionDownDiagonal
        ( ATL, /**/ ATR,  A00, /**/ A01, A02,
         /*************/ /******************/
               /**/       A10, /**/ A11, A12,
          ABL, /**/ ABR,  A20, /**/ A21, A22 );

        RepartitionDown
        ( dT,  d0,
         /**/ /**/
               d1,
          dB,  d2 );

        A21_VC_STAR.AlignWith( A22 );
        A21_VR_STAR.AlignWith( A22 );
        S21Trans_STAR_MC.AlignWith( A22 );
        A21AdjOrTrans_STAR_MR.AlignWith( A22 );
        //--------------------------------------------------------------------//
        A11_STAR_STAR = A11;
        LocalLDL( orientation, A11_STAR_STAR, d1_STAR_STAR );
        A11 = A11_STAR_STAR;
        d1 = d1_STAR_STAR;

        A21_VC_STAR = A21;
        LocalTrsm
        ( RIGHT, LOWER, orientation, UNIT,
          F(1), A11_STAR_STAR, A21_VC_STAR );

        S21Trans_STAR_MC.TransposeFrom( A21_VC_STAR );
        DiagonalSolve( RIGHT, NORMAL, d1_STAR_STAR, A21_VC_STAR );
        A21_VR_STAR = A21_VC_STAR;
        A21AdjOrTrans_STAR_MR.TransposeFrom( A21_VR_STAR, conjugate );
        LocalTrrk
        ( LOWER, TRANSPOSE,
          F(-1), S21Trans_STAR_MC, A21AdjOrTrans_STAR_MR, F(1), A22 );

        A21 = A21_VC_STAR;
        //--------------------------------------------------------------------//
        A21_VC_STAR.FreeAlignments();
        A21_VR_STAR.FreeAlignments();
        S21Trans_STAR_MC.FreeAlignments();
        A21AdjOrTrans_STAR_MR.FreeAlignments();

        SlidePartitionDown
        ( dT,  d0,
               d1,
         /**/ /**/
          dB,  d2 );
//.........這裏部分代碼省略.........