C++ DistMatrix::LockedBuffer方法代码示例

const DistMatrix<T,MD,STAR>&
DistMatrix<T,MD,STAR>::operator=( const DistMatrix<T,STAR,STAR>& A )
{
#ifndef RELEASE
    CallStackEntry entry("[MD,* ] = [* ,* ]");
    this->AssertNotLocked();
    this->AssertSameGrid( A.Grid() );
#endif
    this->ResizeTo( A.Height(), A.Width() );
    if( !this->Participating() )
        return *this;

    const Int lcm = this->grid_->LCM();
    const Int colShift = this->ColShift();

    const Int width = this->Width();
    const Int localHeight = this->LocalHeight();

    const T* ABuf = A.LockedBuffer();
    const Int ALDim = A.LDim();
    T* thisBuffer = this->Buffer();
    const Int thisLDim = this->LDim();
    PARALLEL_FOR
    for( Int j=0; j<width; ++j )
    {
        T* destCol = &thisBuffer[j*thisLDim];
        const T* sourceCol = &ABuf[colShift+j*ALDim];
        for( Int iLoc=0; iLoc<localHeight; ++iLoc )
            destCol[iLoc] = sourceCol[iLoc*lcm];
    }
    return *this;
}

void Scatter
( const DistMatrix<T,CIRC,CIRC>& A,
        DistMatrix<T,STAR,STAR>& B )
{
    DEBUG_CSE
    AssertSameGrids( A, B );

    const Int height = A.Height();
    const Int width = A.Width();
    B.Resize( height, width );

    if( B.Participating() )
    {
        const Int pkgSize = mpi::Pad( height*width );
        vector<T> buffer;
        FastResize( buffer, pkgSize );

        // Pack            
        if( A.Participating() )
            util::InterleaveMatrix
            ( height, width,
              A.LockedBuffer(), 1, A.LDim(),
              buffer.data(),    1, height );

        // Broadcast from the process that packed
        mpi::Broadcast( buffer.data(), pkgSize, A.Root(), A.CrossComm() );

        // Unpack
        util::InterleaveMatrix
        ( height, width,
          buffer.data(), 1, height,
          B.Buffer(),    1, B.LDim() );
    }
}

const DistMatrix<T,STAR,STAR>&
DistMatrix<T,STAR,STAR>::operator=( const DistMatrix<T,VR,STAR>& A )
{
#ifndef RELEASE
    CallStackEntry entry("[* ,* ] = [VR,* ]");
    this->AssertNotLocked();
    this->AssertSameGrid( A.Grid() );
#endif
    const elem::Grid& g = this->Grid();
    this->ResizeTo( A.Height(), A.Width() );
    if( !this->Participating() )
        return *this;

    const Int p = g.Size();
    const Int height = this->Height();
    const Int width = this->Width();
    const Int localHeightOfA = A.LocalHeight();
    const Int maxLocalHeight = MaxLength(height,p);

    const Int portionSize = mpi::Pad( maxLocalHeight*width );
    T* buffer = this->auxMemory_.Require( (p+1)*portionSize );
    T* sendBuf = &buffer[0];
    T* recvBuf = &buffer[portionSize];

    // Pack
    const Int ALDim = A.LDim();
    const T* ABuf = A.LockedBuffer();
    PARALLEL_FOR
    for( Int j=0; j<width; ++j )
        MemCopy
        ( &sendBuf[j*localHeightOfA], &ABuf[j*ALDim], localHeightOfA );

    // Communicate
    mpi::AllGather
    ( sendBuf, portionSize,
      recvBuf, portionSize, g.VRComm() );

    // Unpack
    T* thisBuf = this->Buffer();
    const Int thisLDim = this->LDim();
    const Int colAlignmentOfA = A.ColAlignment();
    OUTER_PARALLEL_FOR
    for( Int k=0; k<p; ++k )
    {
        const T* data = &recvBuf[k*portionSize];
        const Int colShift = Shift_( k, colAlignmentOfA, p );
        const Int localHeight = Length_( height, colShift, p );
        INNER_PARALLEL_FOR
        for( Int j=0; j<width; ++j )
        {
            T* destCol = &thisBuf[colShift+j*thisLDim];
            const T* sourceCol = &data[j*localHeight];
            for( Int iLoc=0; iLoc<localHeight; ++iLoc )
                destCol[iLoc*p] = sourceCol[iLoc];
        }
    }
    this->auxMemory_.Release();
    return *this;
}

void AllGather
( const DistMatrix<T,        U,           V   >& A,
        DistMatrix<T,Collect<U>(),Collect<V>()>& B )
{
    EL_DEBUG_CSE
    AssertSameGrids( A, B );

    const Int height = A.Height();
    const Int width = A.Width();
    B.SetGrid( A.Grid() );
    B.Resize( height, width );

    if( A.Participating() )
    {
        if( A.DistSize() == 1 )
        {
            Copy( A.LockedMatrix(), B.Matrix() );
        }
        else
        {
            const Int colStride = A.ColStride();
            const Int rowStride = A.RowStride();
            const Int distStride = colStride*rowStride;
            const Int maxLocalHeight = MaxLength(height,colStride);
            const Int maxLocalWidth = MaxLength(width,rowStride);
            const Int portionSize = mpi::Pad( maxLocalHeight*maxLocalWidth );
            vector<T> buf;
            FastResize( buf, (distStride+1)*portionSize );
            T* sendBuf = &buf[0];
            T* recvBuf = &buf[portionSize];

            // Pack
            util::InterleaveMatrix
            ( A.LocalHeight(), A.LocalWidth(),
              A.LockedBuffer(), 1, A.LDim(),
              sendBuf,          1, A.LocalHeight() );

            // Communicate
            mpi::AllGather
            ( sendBuf, portionSize, recvBuf, portionSize, A.DistComm() );

            // Unpack
            util::StridedUnpack
            ( height, width,
              A.ColAlign(), colStride,
              A.RowAlign(), rowStride,
              recvBuf, portionSize,
              B.Buffer(), B.LDim() );
        }
    }
    if( A.Grid().InGrid() && A.CrossComm() != mpi::COMM_SELF )
        El::Broadcast( B, A.CrossComm(), A.Root() );
}

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
}

void Filter
( const DistMatrix<T,Collect<U>(),Collect<V>()>& A,
        DistMatrix<T,        U,           V   >& B )
{
    DEBUG_CSE
    AssertSameGrids( A, B );

    B.Resize( A.Height(), A.Width() );
    if( !B.Participating() )
        return;

    const Int colShift = B.ColShift();
    const Int rowShift = B.RowShift();
    util::InterleaveMatrix
    ( B.LocalHeight(), B.LocalWidth(),
      A.LockedBuffer(colShift,rowShift), B.ColStride(), B.RowStride()*A.LDim(),
      B.Buffer(),                        1,             B.LDim() );
}

void GetMappedDiagonal
( const DistMatrix<T,U,V>& A,
        AbstractDistMatrix<S>& dPre,
        function<S(const T&)> func,
        Int offset )
{
    EL_DEBUG_CSE
    EL_DEBUG_ONLY(AssertSameGrids( A, dPre ))
    ElementalProxyCtrl ctrl;
    ctrl.colConstrain = true;
    ctrl.colAlign = A.DiagonalAlign(offset);
    ctrl.rootConstrain = true;
    ctrl.root = A.DiagonalRoot(offset);

    DistMatrixWriteProxy<S,S,DiagCol<U,V>(),DiagRow<U,V>()> dProx( dPre, ctrl );
    auto& d = dProx.Get();

    d.Resize( A.DiagonalLength(offset), 1 );
    if( d.Participating() )
    {
        const Int diagShift = d.ColShift();
        const Int iStart = diagShift + Max(-offset,0);
        const Int jStart = diagShift + Max( offset,0);

        const Int colStride = A.ColStride();
        const Int rowStride = A.RowStride();
        const Int iLocStart = (iStart-A.ColShift()) / colStride;
        const Int jLocStart = (jStart-A.RowShift()) / rowStride;
        const Int iLocStride = d.ColStride() / colStride;
        const Int jLocStride = d.ColStride() / rowStride;

        const Int localDiagLength = d.LocalHeight();
        S* dBuf = d.Buffer();
        const T* ABuf = A.LockedBuffer();
        const Int ldim = A.LDim();
        EL_PARALLEL_FOR
        for( Int k=0; k<localDiagLength; ++k )
        {
            const Int iLoc = iLocStart + k*iLocStride;
            const Int jLoc = jLocStart + k*jLocStride;
            dBuf[k] = func(ABuf[iLoc+jLoc*ldim]);
        }
    }
}

inline void AddInLocalData
( const DistMatrix<F,VC,STAR>& X1, DistMatrix<F,STAR,STAR>& Z )
{
#ifndef RELEASE
    PushCallStack("internal::AddInLocalData");
#endif
    const int width = X1.Width();
    const int localHeight = X1.LocalHeight();
    const int stride = X1.Grid().Size();
    const int offset = X1.ColShift();
    for( int j=0; j<width; ++j )
    {
        F* ZColBuffer = Z.Buffer(0,j);
        const F* X1ColBuffer = X1.LockedBuffer(0,j);
        for( int iLocal=0; iLocal<localHeight; ++iLocal )
            ZColBuffer[offset+stride*iLocal] += X1ColBuffer[iLocal];
    }
#ifndef RELEASE
    PopCallStack();
#endif
}

void AccumulateRHS( const DistMatrix<F,VC,STAR>& X, DistMatrix<F,STAR,STAR>& Z )
{
    const Int height = X.Height();
    const Int width = X.Width();
    Z.Empty();
    Zeros( Z, height, width );

    const Int localHeight = X.LocalHeight();
    const Int colShift = X.ColShift();
    const int commSize = X.Grid().Size();
    const F* XBuffer = X.LockedBuffer();
    F* ZBuffer = Z.Buffer();
    const Int XLDim = X.LDim();
    const Int ZLDim = Z.LDim();
    for( Int iLoc=0; iLoc<localHeight; ++iLoc )
    {
        const Int i = colShift + iLoc*commSize;
        for( Int j=0; j<width; ++j )
            ZBuffer[i+j*ZLDim] = XBuffer[iLoc+j*XLDim];
    }
    mpi::AllReduce( ZBuffer, ZLDim*width, mpi::SUM, X.Grid().VCComm() );
}

void FormDiagonalBlocks
( const DistMatrix<F,VC,STAR>& L, DistMatrix<F,STAR,STAR>& D, bool conjugate )
{
    const Grid& g = L.Grid();

    const Int height = L.Width();
    const Int blocksize = Blocksize();

    const int commRank = g.VCRank();
    const int commSize = g.Size();

    const Int localHeight = Length(height,commRank,commSize);
    const Int maxLocalHeight = MaxLength(height,commSize);
    const Int portionSize = maxLocalHeight*blocksize;

    std::vector<F> sendBuffer( portionSize );
    const Int colShift = L.ColShift();
    const Int LLDim = L.LDim();
    const F* LBuffer = L.LockedBuffer();
    if( conjugate )
    {
        for( Int iLoc=0; iLoc<localHeight; ++iLoc )
        {
            const Int i = colShift + iLoc*commSize;
            const Int block = i / blocksize;
            const Int jStart = block*blocksize;
            const Int b = std::min(height-jStart,blocksize);
            for( Int jOff=0; jOff<b; ++jOff )
                sendBuffer[iLoc*blocksize+jOff] = 
                    Conj(LBuffer[iLoc+(jStart+jOff)*LLDim]);
        }
    }
    else
    {
        for( Int iLoc=0; iLoc<localHeight; ++iLoc )
        {
            const Int i = colShift + iLoc*commSize;
            const Int block = i / blocksize;
            const Int jStart = block*blocksize;
            const Int b = std::min(height-jStart,blocksize);
            for( Int jOff=0; jOff<b; ++jOff )
                sendBuffer[iLoc*blocksize+jOff] = 
                    LBuffer[iLoc+(jStart+jOff)*LLDim];
        }
    }

    std::vector<F> recvBuffer( portionSize*commSize );
    mpi::AllGather
    ( &sendBuffer[0], portionSize, &recvBuffer[0], portionSize, g.VCComm() );
    SwapClear( sendBuffer );
    
    D.Resize( blocksize, height );
    F* DBuffer = D.Buffer();
    const Int DLDim = D.LDim();
    for( Int proc=0; proc<commSize; ++proc )
    {
        const F* procRecv = &recvBuffer[proc*portionSize];
        const Int procLocalHeight = Length(height,proc,commSize);
        for( Int iLoc=0; iLoc<procLocalHeight; ++iLoc )
        {
            const Int i = proc + iLoc*commSize;
            for( Int jOff=0; jOff<blocksize; ++jOff )
                DBuffer[jOff+i*DLDim] = procRecv[jOff+iLoc*blocksize];
        }
    }
}

void ColAllToAllPromote
( const DistMatrix<T,        U,                     V   >& A,
        DistMatrix<T,Partial<U>(),PartialUnionRow<U,V>()>& B )
{
    DEBUG_CSE
    AssertSameGrids( A, B );

    const Int height = A.Height();
    const Int width = A.Width();
    B.AlignColsAndResize
    ( Mod(A.ColAlign(),B.ColStride()), height, width, false, false );
    if( !B.Participating() )
        return;

    const Int colStride = A.ColStride();
    const Int colStridePart = A.PartialColStride();
    const Int colStrideUnion = A.PartialUnionColStride();
    const Int colRankPart = A.PartialColRank();
    const Int colDiff = B.ColAlign() - Mod(A.ColAlign(),colStridePart);

    const Int maxLocalHeight = MaxLength(height,colStride);
    const Int maxLocalWidth = MaxLength(width,colStrideUnion);
    const Int portionSize = mpi::Pad( maxLocalHeight*maxLocalWidth );

    if( colDiff == 0 )
    {
        if( A.PartialUnionColStride() == 1 )
        {
            Copy( A.LockedMatrix(), B.Matrix() );
        }
        else
        {
            vector<T> buffer;
            FastResize( buffer, 2*colStrideUnion*portionSize );
            T* firstBuf  = &buffer[0];
            T* secondBuf = &buffer[colStrideUnion*portionSize];

            // Pack            
            util::RowStridedPack
            ( A.LocalHeight(), width,
              B.RowAlign(), colStrideUnion,
              A.LockedBuffer(), A.LDim(),
              firstBuf,         portionSize );

            // Simultaneously Gather in columns and Scatter in rows
            mpi::AllToAll
            ( firstBuf,  portionSize,
              secondBuf, portionSize, A.PartialUnionColComm() );

            // Unpack
            util::PartialColStridedUnpack 
            ( height, B.LocalWidth(),
              A.ColAlign(), colStride,
              colStrideUnion, colStridePart, colRankPart,
              B.ColShift(),
              secondBuf,  portionSize,
              B.Buffer(), B.LDim() );
        }
    }
    else
    {
#ifdef EL_UNALIGNED_WARNINGS
        if( A.Grid().Rank() == 0 )
            cerr << "Unaligned PartialColAllToAllPromote" << endl;
#endif
        const Int sendColRankPart = Mod( colRankPart+colDiff, colStridePart );
        const Int recvColRankPart = Mod( colRankPart-colDiff, colStridePart );

        vector<T> buffer;
        FastResize( buffer, 2*colStrideUnion*portionSize );
        T* firstBuf  = &buffer[0];
        T* secondBuf = &buffer[colStrideUnion*portionSize];

        // Pack
        util::RowStridedPack
        ( A.LocalHeight(), width,
          B.RowAlign(), colStrideUnion,
          A.LockedBuffer(), A.LDim(),
          secondBuf,        portionSize );

        // Realign the input
        mpi::SendRecv
        ( secondBuf, colStrideUnion*portionSize, sendColRankPart,
          firstBuf,  colStrideUnion*portionSize, recvColRankPart,
          A.PartialColComm() );

        // Simultaneously Scatter in columns and Gather in rows
        mpi::AllToAll
        ( firstBuf,  portionSize,
          secondBuf, portionSize, A.PartialUnionColComm() );

        // Unpack
        util::PartialColStridedUnpack 
        ( height, B.LocalWidth(),
          A.ColAlign(), colStride,
          colStrideUnion, colStridePart, recvColRankPart,
          B.ColShift(),
          secondBuf,  portionSize,
          B.Buffer(), B.LDim() );
    }
//.........这里部分代码省略.........

//.........这里部分代码省略.........
    // necessary processes at each step.
    Int requiredMemory = 0;
    if( inAGrid )
        requiredMemory += maxSendSize;
    if( inBGrid )
        requiredMemory += maxSendSize;
    vector<T> auxBuf( requiredMemory );
    Int offset = 0;
    T* sendBuf = &auxBuf[offset];
    if( inAGrid )
        offset += maxSendSize;
    T* recvBuf = &auxBuf[offset];

    Int recvRow = 0; // avoid compiler warnings...
    if( inAGrid )
        recvRow = Mod(Mod(A.ColRank()-colAlignA,colStrideA)+colAlign,colStride);
    for( Int colSend=0; colSend<numColSends; ++colSend )
    {
        Int recvCol = 0; // avoid compiler warnings...
        if( inAGrid )
            recvCol=Mod(Mod(A.RowRank()-rowAlignA,rowStrideA)+rowAlign,
                        rowStride);
        for( Int rowSend=0; rowSend<numRowSends; ++rowSend )
        {
            mpi::Request sendRequest;
            // Fire off this round of non-blocking sends
            if( inAGrid )
            {
                // Pack the data
                Int sendHeight = Length(A.LocalHeight(),colSend,numColSends);
                Int sendWidth = Length(A.LocalWidth(),rowSend,numRowSends);
                copy::util::InterleaveMatrix
                ( sendHeight, sendWidth,
                  A.LockedBuffer(colSend,rowSend),
                 numColSends, numRowSends*A.LDim(),
                  sendBuf, 1, sendHeight );
                // Send data
                const Int recvVCRank = recvRow + recvCol*colStride;
                const Int recvViewingRank = B.Grid().VCToViewing( recvVCRank );
                mpi::ISend
                ( sendBuf, sendHeight*sendWidth, recvViewingRank,
                  B.Grid().ViewingComm(), sendRequest );
            }
            // Perform this round of recv's
            if( inBGrid )
            {
                const Int sendColOffset = colAlignA;
                const Int recvColOffset =
                    (colSend*colStrideA+colAlign) % colStride;
                const Int sendRowOffset = rowAlignA;
                const Int recvRowOffset =
                    (rowSend*rowStrideA+rowAlign) % rowStride;

                const Int firstSendRow =
                    Mod( Mod(colRank-recvColOffset,colStride)+sendColOffset,
                         colStrideA );
                const Int firstSendCol =
                    Mod( Mod(rowRank-recvRowOffset,rowStride)+sendRowOffset,
                         rowStrideA );

                const Int colShift = Mod( colRank-recvColOffset, colStride );
                const Int rowShift = Mod( rowRank-recvRowOffset, rowStride );
                const Int numColRecvs = Length( colStrideA, colShift, colStride );
                const Int numRowRecvs = Length( rowStrideA, rowShift, rowStride );

                // Recv data

inline void
Cannon_NN
( T alpha, const DistMatrix<T>& A,
           const DistMatrix<T>& B,
  T beta,        DistMatrix<T>& C )
{
#ifndef RELEASE
    CallStackEntry entry("gemm::Cannon_NN");
    if( A.Grid() != B.Grid() || B.Grid() != C.Grid() )
        LogicError("{A,B,C} must have the same grid");
    if( A.Height() != C.Height() ||
        B.Width()  != C.Width()  ||
        A.Width()  != B.Height() )
    {
        std::ostringstream msg;
        msg << "Nonconformal matrices: \n"
            << "  A ~ " << A.Height() << " x " << A.Width() << "\n"
            << "  B ~ " << B.Height() << " x " << B.Width() << "\n"
            << "  C ~ " << C.Height() << " x " << C.Width() << "\n";
        LogicError( msg.str() );
    }
#endif
    const Grid& g = A.Grid();
    if( g.Height() != g.Width() )
        LogicError("Process grid must be square for Cannon's");
    if( C.ColAlignment() != A.ColAlignment() || 
        C.RowAlignment() != B.RowAlignment() )
        LogicError("C is not properly aligned");

    const Int row = g.Row();
    const Int col = g.Col();
    const Int pSqrt = g.Height();
    mpi::Comm rowComm = g.RowComm();
    mpi::Comm colComm = g.ColComm(); 
    if( A.Width() % pSqrt != 0 )
        LogicError("For now, width(A) must be integer multiple of sqrt(p)");

    // Begin by scaling our local portion of C
    Scale( beta, C );

    // Load the initial A and B packages (may want to transpose B...)
    const Int localHeightA = A.LocalHeight();
    const Int localHeightB = B.LocalHeight();
    const Int localWidthA = A.LocalWidth();
    const Int localWidthB = B.LocalWidth();
    Matrix<T> pkgA(localHeightA,localWidthA,localHeightA), 
              pkgB(localHeightB,localWidthB,localHeightB);
    for( Int jLoc=0; jLoc<localWidthA; ++jLoc )
        MemCopy
        ( pkgA.Buffer(0,jLoc), A.LockedBuffer(0,jLoc), localHeightA );
    for( Int jLoc=0; jLoc<localWidthB; ++jLoc )
        MemCopy
        ( pkgB.Buffer(0,jLoc), B.LockedBuffer(0,jLoc), localHeightB );

    // Perform the initial circular shifts so that our A and B packages align
    const Int rowShiftA = A.RowShift();
    const Int colShiftB = B.ColShift();
    const Int leftInitA = (col+pSqrt-colShiftB) % pSqrt;
    const Int rightInitA = (col+colShiftB) % pSqrt;
    const Int aboveInitB = (row+pSqrt-rowShiftA) % pSqrt;
    const Int belowInitB = (row+rowShiftA) % pSqrt;
    const Int pkgSizeA = localHeightA*localWidthA;
    const Int pkgSizeB = localHeightB*localWidthB;
    mpi::SendRecv( pkgA.Buffer(), pkgSizeA, leftInitA, rightInitA, rowComm );
    mpi::SendRecv( pkgB.Buffer(), pkgSizeB, aboveInitB, belowInitB, colComm );

    // Now begin the data flow
    const Int aboveRow = (row+pSqrt-1) % pSqrt;
    const Int belowRow = (row+1) % pSqrt;
    const Int leftCol = (col+pSqrt-1) % pSqrt;
    const Int rightCol = (col+1) % pSqrt;
    for( Int q=0; q<pSqrt; ++q )
    {
        Gemm( NORMAL, NORMAL, alpha, pkgA, pkgB, T(1), C.Matrix() );
        if( q != pSqrt-1 )
        {
            mpi::SendRecv
            ( pkgA.Buffer(), pkgSizeA, leftCol, rightCol, rowComm );
            mpi::SendRecv
            ( pkgB.Buffer(), pkgSizeB, aboveRow, belowRow, colComm );
        }
    }
}

void AllGather
( const DistMatrix<T,        U,           V   >& A, 
        DistMatrix<T,Collect<U>(),Collect<V>()>& B ) 
{
    DEBUG_ONLY(CSE cse("copy::AllGather"))
    AssertSameGrids( A, B );

    const Int height = A.Height();
    const Int width = A.Width();
    B.SetGrid( A.Grid() );
    B.Resize( height, width );

    if( A.Participating() )
    {
        const Int colStride = A.ColStride();
        const Int rowStride = A.RowStride();
        const Int distStride = colStride*rowStride;
        const Int maxLocalHeight = MaxLength(height,colStride);
        const Int maxLocalWidth = MaxLength(width,rowStride);
        const Int portionSize = mpi::Pad( maxLocalHeight*maxLocalWidth );
        vector<T> buf( (distStride+1)*portionSize );
        T* sendBuf = &buf[0];
        T* recvBuf = &buf[portionSize];

        // Pack
        util::InterleaveMatrix
        ( A.LocalHeight(), A.LocalWidth(),
          A.LockedBuffer(), 1, A.LDim(),
          sendBuf,          1, A.LocalHeight() );

        // Communicate
        mpi::AllGather
        ( sendBuf, portionSize, recvBuf, portionSize, A.DistComm() );

        // Unpack
        util::StridedUnpack
        ( height, width,
          A.ColAlign(), colStride,
          A.RowAlign(), rowStride,
          recvBuf, portionSize,
          B.Buffer(), B.LDim() );
    }
    if( A.Grid().InGrid() && A.CrossComm() != mpi::COMM_SELF )
    {
        // Pack from the root
        const Int BLocalHeight = B.LocalHeight();
        const Int BLocalWidth = B.LocalWidth();
        vector<T> buf(BLocalHeight*BLocalWidth);
        if( A.CrossRank() == A.Root() )
            util::InterleaveMatrix
            ( BLocalHeight, BLocalWidth,
              B.LockedBuffer(), 1, B.LDim(),
              buf.data(),       1, BLocalHeight ); 

        // Broadcast from the root
        mpi::Broadcast
        ( buf.data(), BLocalHeight*BLocalWidth, A.Root(), A.CrossComm() );

        // Unpack if not the root
        if( A.CrossRank() != A.Root() )
            util::InterleaveMatrix
            ( BLocalHeight, BLocalWidth,
              buf.data(), 1, BLocalHeight,
              B.Buffer(), 1, B.LDim() );
    }
}

void ColAllToAllDemote
( const DistMatrix<T,Partial<U>(),PartialUnionRow<U,V>()>& A,
        DistMatrix<T,        U,                     V   >& B )
{
    DEBUG_ONLY(CallStackEntry cse("copy::ColAllToAllDemote"))
    AssertSameGrids( A, B );

    const Int height = A.Height();
    const Int width = A.Width();
    B.AlignColsAndResize( A.ColAlign(), height, width, false, false );
    if( !B.Participating() )
        return;

    const Int colAlign = B.ColAlign();
    const Int rowAlignA = A.RowAlign();

    const Int colStride = B.ColStride();
    const Int colStridePart = B.PartialColStride();
    const Int colStrideUnion = B.PartialUnionColStride();
    const Int colRankPart = B.PartialColRank();
    const Int colDiff = (colAlign%colStridePart) - A.ColAlign();

    const Int colShiftA = A.ColShift();

    const Int localHeightB = B.LocalHeight();
    const Int localWidthA = A.LocalWidth();
    const Int maxLocalHeight = MaxLength(height,colStride);
    const Int maxLocalWidth = MaxLength(width,colStrideUnion);
    const Int portionSize = mpi::Pad( maxLocalHeight*maxLocalWidth );

    std::vector<T> buffer( 2*colStrideUnion*portionSize );
    T* firstBuf  = &buffer[0];
    T* secondBuf = &buffer[colStrideUnion*portionSize];

    if( colDiff == 0 )
    {
        // Pack            
        util::PartialColStridedPack
        ( height, localWidthA,
          colAlign, colStride, 
          colStrideUnion, colStridePart, colRankPart,
          colShiftA,
          A.LockedBuffer(), A.LDim(),
          firstBuf,         portionSize );

        // Simultaneously Scatter in columns and Gather in rows
        mpi::AllToAll
        ( firstBuf,  portionSize,
          secondBuf, portionSize, B.PartialUnionColComm() );

        // Unpack
        util::RowStridedUnpack
        ( localHeightB, width,
          rowAlignA, colStrideUnion,
          secondBuf, portionSize,
          B.Buffer(), B.LDim() );
    }
    else
    {
#ifdef EL_UNALIGNED_WARNINGS
        if( B.Grid().Rank() == 0 )
            std::cerr << "Unaligned ColAllToAllDemote" << std::endl;
#endif
        const Int sendColRankPart = Mod( colRankPart+colDiff, colStridePart );
        const Int recvColRankPart = Mod( colRankPart-colDiff, colStridePart );

        // Pack
        util::PartialColStridedPack
        ( height, localWidthA,
          colAlign, colStride, 
          colStrideUnion, colStridePart, sendColRankPart,
          colShiftA,
          A.LockedBuffer(), A.LDim(),
          secondBuf,        portionSize );

        // Simultaneously Scatter in columns and Gather in rows
        mpi::AllToAll
        ( secondBuf, portionSize,
          firstBuf,  portionSize, B.PartialUnionColComm() );

        // Realign the result
        mpi::SendRecv
        ( firstBuf,  colStrideUnion*portionSize, sendColRankPart,
          secondBuf, colStrideUnion*portionSize, recvColRankPart,
          B.PartialColComm() );

        // Unpack
        util::RowStridedUnpack
        ( localHeightB, width,
          rowAlignA, colStrideUnion,
          secondBuf, portionSize,
          B.Buffer(), B.LDim() );
    }
}