void RowScatter
( T alpha,
const ElementalMatrix<T>& A,
ElementalMatrix<T>& B )
{
DEBUG_ONLY(CSE cse("axpy_contract::RowScatter"))
AssertSameGrids( A, B );
if( A.Height() != B.Height() || A.Width() != B.Width() )
LogicError("Matrix sizes did not match");
if( !B.Participating() )
return;
const Int width = B.Width();
const Int colDiff = B.ColAlign()-A.ColAlign();
if( colDiff == 0 )
{
if( width == 1 )
{
const Int localHeight = B.LocalHeight();
const Int portionSize = mpi::Pad( localHeight );
//vector<T> buffer( portionSize );
vector<T> buffer;
buffer.reserve( portionSize );
// Reduce to rowAlign
const Int rowAlign = B.RowAlign();
mpi::Reduce
( A.LockedBuffer(), buffer.data(), portionSize,
rowAlign, B.RowComm() );
if( B.RowRank() == rowAlign )
{
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, 1,
buffer.data(), 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
else
{
const Int rowStride = B.RowStride();
const Int rowAlign = B.RowAlign();
const Int localHeight = B.LocalHeight();
const Int localWidth = B.LocalWidth();
const Int maxLocalWidth = MaxLength(width,rowStride);
const Int portionSize = mpi::Pad( localHeight*maxLocalWidth );
const Int sendSize = rowStride*portionSize;
// Pack
//vector<T> buffer( sendSize );
vector<T> buffer;
buffer.reserve( sendSize );
copy::util::RowStridedPack
( localHeight, width,
rowAlign, rowStride,
A.LockedBuffer(), A.LDim(),
buffer.data(), portionSize );
// Communicate
mpi::ReduceScatter( buffer.data(), portionSize, B.RowComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
buffer.data(), 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
else
{
#ifdef EL_UNALIGNED_WARNINGS
if( B.Grid().Rank() == 0 )
cerr << "Unaligned RowScatter" << endl;
#endif
const Int colRank = B.ColRank();
const Int colStride = B.ColStride();
const Int sendRow = Mod( colRank+colDiff, colStride );
const Int recvRow = Mod( colRank-colDiff, colStride );
const Int localHeight = B.LocalHeight();
const Int localHeightA = A.LocalHeight();
if( width == 1 )
{
//vector<T> buffer( localHeight+localHeightA );
vector<T> buffer;
buffer.reserve( localHeight+localHeightA );
T* sendBuf = &buffer[0];
T* recvBuf = &buffer[localHeightA];
// Reduce to rowAlign
const Int rowAlign = B.RowAlign();
mpi::Reduce
( A.LockedBuffer(), sendBuf, localHeightA, rowAlign, B.RowComm() );
if( B.RowRank() == rowAlign )
{
// Perform the realignment
mpi::SendRecv
( sendBuf, localHeightA, sendRow,
recvBuf, localHeight, recvRow, B.ColComm() );
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, 1,
recvBuf, 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
else
{
const Int rowStride = B.RowStride();
const Int rowAlign = B.RowAlign();
const Int localWidth = B.LocalWidth();
const Int maxLocalWidth = MaxLength(width,rowStride);
const Int recvSize_RS = mpi::Pad( localHeightA*maxLocalWidth );
const Int sendSize_RS = rowStride * recvSize_RS;
const Int recvSize_SR = localHeight * localWidth;
//vector<T> buffer( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
vector<T> buffer;
buffer.reserve( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
T* firstBuf = &buffer[0];
T* secondBuf = &buffer[recvSize_RS];
// Pack
copy::util::RowStridedPack
( localHeightA, width,
rowAlign, rowStride,
A.LockedBuffer(), A.LDim(),
secondBuf, recvSize_RS );
// Reduce-scatter over each process row
mpi::ReduceScatter( secondBuf, firstBuf, recvSize_RS, B.RowComm() );
// Trade reduced data with the appropriate process row
mpi::SendRecv
( firstBuf, localHeightA*localWidth, sendRow,
secondBuf, localHeight*localWidth, recvRow, B.ColComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
secondBuf, 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}
}
void ColScatter
( T alpha,
const ElementalMatrix<T>& A,
ElementalMatrix<T>& B )
{
DEBUG_ONLY(CSE cse("axpy_contract::ColScatter"))
AssertSameGrids( A, B );
if( A.Height() != B.Height() || A.Width() != B.Width() )
LogicError("A and B must be the same size");
#ifdef EL_VECTOR_WARNINGS
if( A.Width() == 1 && B.Grid().Rank() == 0 )
{
cerr <<
"The vector version of ColScatter does not"
" yet have a vector version implemented, but it would only "
"require a modification of the vector version of RowScatter"
<< endl;
}
#endif
#ifdef EL_CACHE_WARNINGS
if( A.Width() != 1 && B.Grid().Rank() == 0 )
{
cerr <<
"axpy_contract::ColScatter potentially causes a large "
"amount of cache-thrashing. If possible, avoid it by forming the "
"(conjugate-)transpose of the [* ,V] matrix instead." << endl;
}
#endif
if( !B.Participating() )
return;
const Int height = B.Height();
const Int localHeight = B.LocalHeight();
const Int localWidth = B.LocalWidth();
const Int colAlign = B.ColAlign();
const Int colStride = B.ColStride();
const Int rowDiff = B.RowAlign()-A.RowAlign();
// TODO: Allow for modular equivalence if possible
if( rowDiff == 0 )
{
const Int maxLocalHeight = MaxLength(height,colStride);
const Int recvSize = mpi::Pad( maxLocalHeight*localWidth );
const Int sendSize = colStride*recvSize;
//vector<T> buffer( sendSize );
vector<T> buffer;
buffer.reserve( sendSize );
// Pack
copy::util::ColStridedPack
( height, localWidth,
colAlign, colStride,
A.LockedBuffer(), A.LDim(),
buffer.data(), recvSize );
// Communicate
mpi::ReduceScatter( buffer.data(), recvSize, B.ColComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
buffer.data(), 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
else
{
#ifdef EL_UNALIGNED_WARNINGS
if( B.Grid().Rank() == 0 )
cerr << "Unaligned ColScatter" << endl;
#endif
const Int localWidthA = A.LocalWidth();
const Int maxLocalHeight = MaxLength(height,colStride);
const Int recvSize_RS = mpi::Pad( maxLocalHeight*localWidthA );
const Int sendSize_RS = colStride*recvSize_RS;
const Int recvSize_SR = localHeight*localWidth;
//vector<T> buffer( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
vector<T> buffer;
buffer.reserve( recvSize_RS + Max(sendSize_RS,recvSize_SR) );
T* firstBuf = &buffer[0];
T* secondBuf = &buffer[recvSize_RS];
// Pack
copy::util::ColStridedPack
( height, localWidth,
colAlign, colStride,
A.LockedBuffer(), A.LDim(),
secondBuf, recvSize_RS );
// Reduce-scatter over each col
mpi::ReduceScatter( secondBuf, firstBuf, recvSize_RS, B.ColComm() );
// Trade reduced data with the appropriate col
const Int sendCol = Mod( B.RowRank()+rowDiff, B.RowStride() );
const Int recvCol = Mod( B.RowRank()-rowDiff, B.RowStride() );
mpi::SendRecv
( firstBuf, localHeight*localWidthA, sendCol,
secondBuf, localHeight*localWidth, recvCol, B.RowComm() );
// Update with our received data
axpy::util::InterleaveMatrixUpdate
( alpha, localHeight, localWidth,
secondBuf, 1, localHeight,
B.Buffer(), 1, B.LDim() );
}
}