コード例 #1
0
ファイル: AllGather.hpp プロジェクト: elemental/Elemental
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() );
}
コード例 #2
0
ファイル: Scatter.hpp プロジェクト: YingzhouLi/Elemental
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() );
    }
}
コード例 #3
0
ファイル: Gather.cpp プロジェクト: andreasnoack/Elemental
void Gather
( const BlockMatrix<T>& A,
        DistMatrix<T,CIRC,CIRC,BLOCK>& B )
{
    DEBUG_ONLY(CSE cse("copy::Gather"))
    AssertSameGrids( A, B );
    if( A.DistSize() == 1 && A.CrossSize() == 1 )
    {
        B.Resize( A.Height(), A.Width() );
        if( B.CrossRank() == B.Root() )
            Copy( A.LockedMatrix(), B.Matrix() );
        return;
    }

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

    // Gather the colShifts and rowShifts
    // ==================================
    Int myShifts[2];
    myShifts[0] = A.ColShift();
    myShifts[1] = A.RowShift();
    vector<Int> shifts;
    const Int crossSize = B.CrossSize();
    if( B.CrossRank() == B.Root() )
        shifts.resize( 2*crossSize );
    mpi::Gather( myShifts, 2, shifts.data(), 2, B.Root(), B.CrossComm() );

    // Gather the payload data
    // =======================
    const bool irrelevant = ( A.RedundantRank()!=0 || A.CrossRank()!=A.Root() );
    int totalSend = ( irrelevant ? 0 : A.LocalHeight()*A.LocalWidth() );
    vector<int> recvCounts, recvOffsets;
    if( B.CrossRank() == B.Root() )
        recvCounts.resize( crossSize );
    mpi::Gather( &totalSend, 1, recvCounts.data(), 1, B.Root(), B.CrossComm() );
    int totalRecv = Scan( recvCounts, recvOffsets );
    //vector<T> sendBuf(totalSend), recvBuf(totalRecv);
    vector<T> sendBuf, recvBuf;
    sendBuf.reserve( totalSend );
    recvBuf.reserve( totalRecv );
    if( !irrelevant )
        copy::util::InterleaveMatrix
        ( A.LocalHeight(), A.LocalWidth(),
          A.LockedBuffer(), 1, A.LDim(),
          sendBuf.data(),   1, A.LocalHeight() );
    mpi::Gather
    ( sendBuf.data(), totalSend,
      recvBuf.data(), recvCounts.data(), recvOffsets.data(), 
      B.Root(), B.CrossComm() );

    // Unpack
    // ======
    const Int mb = A.BlockHeight();
    const Int nb = A.BlockWidth();
    const Int colCut = A.ColCut();
    const Int rowCut = A.RowCut();
    if( B.Root() == B.CrossRank() )
    {
        for( Int q=0; q<crossSize; ++q )
        {
            if( recvCounts[q] == 0 )
                continue;
            const Int colShift = shifts[2*q+0];
            const Int rowShift = shifts[2*q+1];
            const Int colStride = A.ColStride();
            const Int rowStride = A.RowStride();
            const Int localHeight =
              BlockedLength( height, colShift, mb, colCut, colStride );
            const Int localWidth = 
              BlockedLength( width, rowShift, nb, rowCut, rowStride );
            const T* data = &recvBuf[recvOffsets[q]];
            for( Int jLoc=0; jLoc<localWidth; ++jLoc )
            {
                const Int jBefore = rowShift*nb - rowCut;
                const Int jLocAdj = ( rowShift==0 ? jLoc+rowCut : jLoc );
                const Int numFilledLocalBlocks = jLocAdj / nb;
                const Int jMid = numFilledLocalBlocks*nb*rowStride;
                const Int jPost = jLocAdj-numFilledLocalBlocks*nb;
                const Int j = jBefore + jMid + jPost;
                const T* sourceCol = &data[jLoc*localHeight];
                for( Int iLoc=0; iLoc<localHeight; ++iLoc )
                {
                    const Int iBefore = colShift*mb - colCut;
                    const Int iLocAdj = (colShift==0 ? iLoc+colCut : iLoc);
                    const Int numFilledLocalBlocks = iLocAdj / mb;
                    const Int iMid = numFilledLocalBlocks*mb*colStride;
                    const Int iPost = iLocAdj-numFilledLocalBlocks*mb;
                    const Int i = iBefore + iMid + iPost;
                    B.SetLocal(i,j,sourceCol[iLoc]);
                }
            }
        }
    }
}
コード例 #4
0
ファイル: Gather.cpp プロジェクト: andreasnoack/Elemental
void Gather
( const ElementalMatrix<T>& A,
        DistMatrix<T,CIRC,CIRC>& B )
{
    DEBUG_ONLY(CSE cse("copy::Gather"))
    AssertSameGrids( A, B );
    if( A.DistSize() == 1 && A.CrossSize() == 1 )
    {
        B.Resize( A.Height(), A.Width() );
        if( B.CrossRank() == B.Root() )
            Copy( A.LockedMatrix(), B.Matrix() );
        return;
    }

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

    // Gather the colShifts and rowShifts
    // ==================================
    Int myShifts[2];
    myShifts[0] = A.ColShift();
    myShifts[1] = A.RowShift();
    vector<Int> shifts;
    const Int crossSize = B.CrossSize();
    if( B.CrossRank() == B.Root() )
        shifts.resize( 2*crossSize );
    mpi::Gather( myShifts, 2, shifts.data(), 2, B.Root(), B.CrossComm() );

    // Gather the payload data
    // =======================
    const bool irrelevant = ( A.RedundantRank()!=0 || A.CrossRank()!=A.Root() );
    int totalSend = ( irrelevant ? 0 : A.LocalHeight()*A.LocalWidth() );
    vector<int> recvCounts, recvOffsets;
    if( B.CrossRank() == B.Root() )
        recvCounts.resize( crossSize );
    mpi::Gather( &totalSend, 1, recvCounts.data(), 1, B.Root(), B.CrossComm() );
    int totalRecv = Scan( recvCounts, recvOffsets );
    //vector<T> sendBuf(totalSend), recvBuf(totalRecv);
    vector<T> sendBuf, recvBuf;
    sendBuf.reserve( totalSend );
    recvBuf.reserve( totalRecv );
    if( !irrelevant )
        copy::util::InterleaveMatrix
        ( A.LocalHeight(), A.LocalWidth(),
          A.LockedBuffer(), 1, A.LDim(),
          sendBuf.data(),   1, A.LocalHeight() );
    mpi::Gather
    ( sendBuf.data(), totalSend,
      recvBuf.data(), recvCounts.data(), recvOffsets.data(), 
      B.Root(), B.CrossComm() );

    // Unpack
    // ======
    if( B.Root() == B.CrossRank() )
    {
        for( Int q=0; q<crossSize; ++q )
        {
            if( recvCounts[q] == 0 )
                continue;
            const Int colShift = shifts[2*q+0];
            const Int rowShift = shifts[2*q+1];
            const Int colStride = A.ColStride();
            const Int rowStride = A.RowStride();
            const Int localHeight = Length( height, colShift, colStride );
            const Int localWidth = Length( width, rowShift, rowStride );
            copy::util::InterleaveMatrix
            ( localHeight, localWidth,
              &recvBuf[recvOffsets[q]],    1,         localHeight,
              B.Buffer(colShift,rowShift), colStride, rowStride*B.LDim() );
        }
    }
}
コード例 #5
0
ファイル: AllGather.cpp プロジェクト: bluehope/Elemental
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() );
    }
}
コード例 #6
0
ファイル: Scatter.hpp プロジェクト: YingzhouLi/Elemental
void Scatter
( const DistMatrix<T,CIRC,CIRC>& A,
        ElementalMatrix<T>& B )
{
    DEBUG_CSE
    AssertSameGrids( A, B );

    const Int m = A.Height();
    const Int n = A.Width();
    const Int colStride = B.ColStride();
    const Int rowStride = B.RowStride();
    B.Resize( m, n );
    if( B.CrossSize() != 1 || B.RedundantSize() != 1 )
    {
        // TODO:
        // Broadcast over the redundant communicator and use mpi::Translate
        // rank to determine whether a process is the root of the broadcast.
        GeneralPurpose( A, B ); 
        return;
    }

    const Int pkgSize = mpi::Pad(MaxLength(m,colStride)*MaxLength(n,rowStride));
    const Int recvSize = pkgSize;
    const Int sendSize = B.DistSize()*pkgSize;

    // Translate the root of A into the DistComm of B (if possible)
    const Int root = A.Root();
    const Int target = mpi::Translate( A.CrossComm(), root, B.DistComm() ); 
    if( target == mpi::UNDEFINED )
        return;

    if( B.DistSize() == 1 )
    {
        Copy( A.LockedMatrix(), B.Matrix() );
        return;
    }

    vector<T> buffer;
    T* recvBuf=0; // some compilers (falsely) warn otherwise
    if( A.CrossRank() == root )
    {
        FastResize( buffer, sendSize+recvSize );
        T* sendBuf = &buffer[0];
        recvBuf    = &buffer[sendSize];

        // Pack the send buffer
        copy::util::StridedPack
        ( m, n,
          B.ColAlign(), colStride,
          B.RowAlign(), rowStride,
          A.LockedBuffer(), A.LDim(),
          sendBuf,          pkgSize );

        // Scatter from the root
        mpi::Scatter
        ( sendBuf, pkgSize, recvBuf, pkgSize, target, B.DistComm() );
    }
    else
    {
        FastResize( buffer, recvSize );
        recvBuf = &buffer[0];

        // Perform the receiving portion of the scatter from the non-root
        mpi::Scatter
        ( static_cast<T*>(0), pkgSize,
          recvBuf,            pkgSize, target, B.DistComm() );
    }

    // Unpack
    copy::util::InterleaveMatrix
    ( B.LocalHeight(), B.LocalWidth(),
      recvBuf,    1, B.LocalHeight(),
      B.Buffer(), 1, B.LDim() );
}