inline
DistMatrix<T,MD,STAR,Int>::DistMatrix( const DistMatrix<T,U,V,Int>& A )
: AbstractDistMatrix<T,Int>(0,0,false,false,0,0,
(A.Participating() ? A.ColRank() : 0),0,
0,0,A.Grid()),
diagPath_(A.diagPath_)
{
#ifndef RELEASE
PushCallStack("DistMatrix[MD,* ]::DistMatrix");
#endif
if( MD != U || STAR != V ||
reinterpret_cast<const DistMatrix<T,MD,STAR,Int>*>(&A) != this )
*this = A;
else
throw std::logic_error("Tried to construct [MD,* ] with itself");
#ifndef RELEASE
PopCallStack();
#endif
}
void TranslateBetweenGrids
( const DistMatrix<T,MC,MR>& A, DistMatrix<T,MC,MR>& B )
{
DEBUG_ONLY(CSE cse("copy::TranslateBetweenGrids [MC,MR]"))
B.Resize( A.Height(), A.Width() );
// Just need to ensure that each viewing comm contains the other team's
// owning comm. Congruence is too strong.
// Compute the number of process rows and columns that each process
// needs to send to.
const Int colStride = B.ColStride();
const Int rowStride = B.RowStride();
const Int colRank = B.ColRank();
const Int rowRank = B.RowRank();
const Int colStrideA = A.ColStride();
const Int rowStrideA = A.RowStride();
const Int colGCD = GCD( colStride, colStrideA );
const Int rowGCD = GCD( rowStride, rowStrideA );
const Int colLCM = colStride*colStrideA / colGCD;
const Int rowLCM = rowStride*rowStrideA / rowGCD;
const Int numColSends = colStride / colGCD;
const Int numRowSends = rowStride / rowGCD;
const Int colAlign = B.ColAlign();
const Int rowAlign = B.RowAlign();
const Int colAlignA = A.ColAlign();
const Int rowAlignA = A.RowAlign();
const bool inBGrid = B.Participating();
const bool inAGrid = A.Participating();
if( !inBGrid && !inAGrid )
return;
const Int maxSendSize =
(A.Height()/(colStrideA*numColSends)+1) *
(A.Width()/(rowStrideA*numRowSends)+1);
// Translate the ranks from A's VC communicator to B's viewing so that
// we can match send/recv communicators. Since A's VC communicator is not
// necessarily defined on every process, we instead work with A's owning
// group and account for row-major ordering if necessary.
const int sizeA = A.Grid().Size();
vector<int> rankMap(sizeA), ranks(sizeA);
if( A.Grid().Order() == COLUMN_MAJOR )
{
for( int j=0; j<sizeA; ++j )
ranks[j] = j;
}
else
{
// The (i,j) = i + j*colStrideA rank in the column-major ordering is
// equal to the j + i*rowStrideA rank in a row-major ordering.
// Since we desire rankMap[i+j*colStrideA] to correspond to process
// (i,j) in A's grid's rank in this viewing group, ranks[i+j*colStrideA]
// should correspond to process (i,j) in A's owning group. Since the
// owning group is ordered row-major in this case, its rank is
// j+i*rowStrideA. Note that setting
// ranks[j+i*rowStrideA] = i+j*colStrideA is *NOT* valid.
for( int i=0; i<colStrideA; ++i )
for( int j=0; j<rowStrideA; ++j )
ranks[i+j*colStrideA] = j+i*rowStrideA;
}
mpi::Translate
( A.Grid().OwningGroup(), sizeA, &ranks[0],
B.Grid().ViewingComm(), &rankMap[0] );
// Have each member of A's grid individually send to all numRow x numCol
// processes in order, while the members of this grid receive from all
// 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
// For now, simply receive sequentially. Until we switch to
// nonblocking recv's, we won't be using much of the
// recvBuf
Int sendRow = firstSendRow;
for( Int colRecv=0; colRecv<numColRecvs; ++colRecv )
{
const Int sendColShift = Shift( sendRow, colAlignA, colStrideA ) + colSend*colStrideA;
const Int sendHeight = Length( A.Height(), sendColShift, colLCM );
const Int localColOffset = (sendColShift-B.ColShift()) / colStride;
Int sendCol = firstSendCol;
for( Int rowRecv=0; rowRecv<numRowRecvs; ++rowRecv )
{
const Int sendRowShift = Shift( sendCol, rowAlignA, rowStrideA ) + rowSend*rowStrideA;
const Int sendWidth = Length( A.Width(), sendRowShift, rowLCM );
const Int localRowOffset = (sendRowShift-B.RowShift()) / rowStride;
const Int sendVCRank = sendRow+sendCol*colStrideA;
mpi::Recv
( recvBuf, sendHeight*sendWidth, rankMap[sendVCRank],
B.Grid().ViewingComm() );
// Unpack the data
copy::util::InterleaveMatrix
( sendHeight, sendWidth,
recvBuf, 1, sendHeight,
B.Buffer(localColOffset,localRowOffset),
colLCM/colStride, (rowLCM/rowStride)*B.LDim() );
// Set up the next send col
sendCol = (sendCol + rowStride) % rowStrideA;
}
// Set up the next send row
sendRow = (sendRow + colStride) % colStrideA;
}
}
// Ensure that this round of non-blocking sends completes
if( inAGrid )
{
mpi::Wait( sendRequest );
recvCol = (recvCol + rowStrideA) % rowStride;
}
}
if( inAGrid )
recvRow = (recvRow + colStrideA) % colStride;
}
}
void TransposeDist( const DistMatrix<T,U,V>& A, DistMatrix<T,V,U>& B )
{
DEBUG_ONLY(CSE cse("copy::TransposeDist"))
AssertSameGrids( A, B );
const Grid& g = B.Grid();
B.Resize( A.Height(), A.Width() );
if( !B.Participating() )
return;
const Int colStrideA = A.ColStride();
const Int rowStrideA = A.RowStride();
const Int distSize = A.DistSize();
if( A.DistSize() == 1 && B.DistSize() == 1 )
{
Copy( A.LockedMatrix(), B.Matrix() );
}
else if( A.Width() == 1 )
{
const Int height = A.Height();
const Int maxLocalHeight = MaxLength(height,distSize);
const Int portionSize = mpi::Pad( maxLocalHeight );
const Int colDiff = Shift(A.DistRank(),A.ColAlign(),distSize) -
Shift(B.DistRank(),B.ColAlign(),distSize);
const Int sendRankB = Mod( B.DistRank()+colDiff, distSize );
const Int recvRankA = Mod( A.DistRank()-colDiff, distSize );
const Int recvRankB =
(recvRankA/colStrideA)+rowStrideA*(recvRankA%colStrideA);
vector<T> buffer;
FastResize( buffer, (colStrideA+rowStrideA)*portionSize );
T* sendBuf = &buffer[0];
T* recvBuf = &buffer[colStrideA*portionSize];
if( A.RowRank() == A.RowAlign() )
{
// Pack
// TODO: Use kernel from copy::util
const Int AColShift = A.ColShift();
const T* ABuf = A.LockedBuffer();
EL_PARALLEL_FOR
for( Int k=0; k<rowStrideA; ++k )
{
T* data = &recvBuf[k*portionSize];
const Int shift =
Shift_(A.ColRank()+colStrideA*k,A.ColAlign(),distSize);
const Int offset = (shift-AColShift) / colStrideA;
const Int thisLocalHeight = Length_(height,shift,distSize);
for( Int iLoc=0; iLoc<thisLocalHeight; ++iLoc )
data[iLoc] = ABuf[offset+iLoc*rowStrideA];
}
}
// (e.g., A[VC,STAR] <- A[MC,MR])
mpi::Scatter
( recvBuf, portionSize,
sendBuf, portionSize, A.RowAlign(), A.RowComm() );
// (e.g., A[VR,STAR] <- A[VC,STAR])
mpi::SendRecv
( sendBuf, portionSize, sendRankB,
recvBuf, portionSize, recvRankB, B.DistComm() );
// (e.g., A[MR,MC] <- A[VR,STAR])
mpi::Gather
( recvBuf, portionSize,
sendBuf, portionSize, B.RowAlign(), B.RowComm() );
if( B.RowRank() == B.RowAlign() )
{
// Unpack
// TODO: Use kernel from copy::util
T* bufB = B.Buffer();
EL_PARALLEL_FOR
for( Int k=0; k<colStrideA; ++k )
{
const T* data = &sendBuf[k*portionSize];
const Int shift =
Shift_(B.ColRank()+rowStrideA*k,B.ColAlign(),distSize);
const Int offset = (shift-B.ColShift()) / rowStrideA;
const Int thisLocalHeight = Length_(height,shift,distSize);
for( Int iLoc=0; iLoc<thisLocalHeight; ++iLoc )
bufB[offset+iLoc*colStrideA] = data[iLoc];
}
}
}