void GatherSubdiagonal
( const DistMatrix<F,MC,MR,BLOCK>& H,
const IR& winInd,
DistMatrix<Base<F>,STAR,STAR>& hSubWin )
{
DEBUG_CSE
const Int winSize = winInd.end - winInd.beg;
const Int blockSize = H.BlockHeight();
const Grid& grid = H.Grid();
const auto& HLoc = H.LockedMatrix();
DEBUG_ONLY(
if( H.BlockHeight() != H.BlockWidth() )
LogicError("Assumed square distribution blocks");
if( H.ColCut() != H.RowCut() )
LogicError("Assumed symmetric cuts");
if( blockSize < 2 )
LogicError("Assumed blocks of size at least two");
)
void TransformRows
( const Matrix<F>& Z,
DistMatrix<F,MC,MR,BLOCK>& H )
{
DEBUG_CSE
const Int height = H.Height();
const Grid& grid = H.Grid();
const Int blockHeight = H.BlockHeight();
const Int firstBlockHeight = blockHeight - H.ColCut();
if( height <= firstBlockHeight || grid.Height() == 1 )
{
if( grid.Row() == H.RowOwner(0) )
{
// This process row can locally update its portion of H
Matrix<F> HLocCopy( H.Matrix() );
Gemm( ADJOINT, NORMAL, F(1), Z, HLocCopy, H.Matrix() );
}
}
else if( height <= firstBlockHeight + blockHeight )
{
const bool firstRow = H.RowOwner( 0 );
const bool secondRow = H.RowOwner( firstBlockHeight );
if( grid.Row() == firstRow )
{
//
// Replace H with
//
// | ZLeft, ZRight |' | HTop |,
// | HBottom |
//
// where HTop is owned by this process row and HBottom by the next.
//
auto ZLeft = Z( ALL, IR(0,firstBlockHeight) );
// Partition space for the combined matrix
Matrix<F> HCombine( height, H.LocalWidth() );
auto HTop = HCombine( IR(0,firstBlockHeight), ALL );
auto HBottom = HCombine( IR(firstBlockHeight,END), ALL );
// Copy our portion into the combined matrix
HTop = H.LockedMatrix();
// Exchange the data
El::SendRecv( HTop, HBottom, H.ColComm(), secondRow, secondRow );
// Form our portion of the result
Gemm( ADJOINT, NORMAL, F(1), ZLeft, HCombine, H.Matrix() );
}
else if( grid.Row() == secondRow )
{
//
// Replace H with
//
// | ZLeft, ZRight |' | HTop |,
// | HBottom |
//
// where HTop is owned by the previous process row and HBottom by
// this one.
//
auto ZRight = Z( ALL, IR(firstBlockHeight,END) );
// Partition space for the combined matrix
Matrix<F> HCombine( height, H.LocalWidth() );
auto HTop = HCombine( IR(0,firstBlockHeight), ALL );
auto HBottom = HCombine( IR(firstBlockHeight,END), ALL );
// Copy our portion into the combined matrix
HBottom = H.LockedMatrix();
// Exchange the data
El::SendRecv( HBottom, HTop, H.ColComm(), firstRow, firstRow );
// Form our portion of the result
Gemm( ADJOINT, NORMAL, F(1), ZRight, HCombine, H.Matrix() );
}
}
else
{
// Fall back to the entire process column interacting.
// TODO(poulson): Only form the subset of the result that we need.
DistMatrix<F,STAR,MR,BLOCK> H_STAR_MR( H );
Matrix<F> HLocCopy( H_STAR_MR.Matrix() );
Gemm( ADJOINT, NORMAL, F(1), Z, HLocCopy, H_STAR_MR.Matrix() );
H = H_STAR_MR;
}
}
void TransformRows
( const Matrix<F>& V,
DistMatrix<F,MC,MR,BLOCK>& A )
{
DEBUG_CSE
const Int height = A.Height();
const Grid& grid = A.Grid();
const Int blockHeight = A.BlockHeight();
const Int firstBlockHeight = blockHeight - A.ColCut();
if( height <= firstBlockHeight || grid.Height() == 1 )
{
if( grid.Row() == A.RowOwner(0) )
{
// This process row can locally update its portion of A
TransformRows( V, A.Matrix() );
}
}
else if( height <= firstBlockHeight + blockHeight )
{
const int firstRow = A.RowOwner( 0 );
const int secondRow = A.RowOwner( firstBlockHeight );
if( grid.Row() == firstRow )
{
//
// Replace A with
//
// | VLeft, VRight |' | ATop |,
// | ABottom |
//
// where ATop is owned by this process row and ABottom by the next.
//
auto VLeft = V( ALL, IR(0,firstBlockHeight) );
// Partition space for the combined matrix
Matrix<F> ACombine( height, A.LocalWidth() );
auto ATop = ACombine( IR(0,firstBlockHeight), ALL );
auto ABottom = ACombine( IR(firstBlockHeight,END), ALL );
// Copy our portion into the combined matrix
ATop = A.LockedMatrix();
// Exchange the data
El::SendRecv( ATop, ABottom, A.ColComm(), secondRow, secondRow );
// Form our portion of the result
Gemm( ADJOINT, NORMAL, F(1), VLeft, ACombine, A.Matrix() );
}
else if( grid.Row() == secondRow )
{
//
// Replace A with
//
// | VLeft, VRight |' | ATop |,
// | ABottom |
//
// where ATop is owned by the previous process row and ABottom by
// this one.
//
auto VRight = V( ALL, IR(firstBlockHeight,END) );
// Partition space for the combined matrix
Matrix<F> ACombine( height, A.LocalWidth() );
auto ATop = ACombine( IR(0,firstBlockHeight), ALL );
auto ABottom = ACombine( IR(firstBlockHeight,END), ALL );
// Copy our portion into the combined matrix
ABottom = A.LockedMatrix();
// Exchange the data
El::SendRecv( ABottom, ATop, A.ColComm(), firstRow, firstRow );
// Form our portion of the result
Gemm( ADJOINT, NORMAL, F(1), VRight, ACombine, A.Matrix() );
}
}
else
{
// Fall back to the entire process column interacting.
// TODO(poulson): Only form the subset of the result that we need.
DistMatrix<F,STAR,MR,BLOCK> A_STAR_MR( A );
Matrix<F> ALocCopy( A_STAR_MR.Matrix() );
Gemm( ADJOINT, NORMAL, F(1), V, ALocCopy, A_STAR_MR.Matrix() );
A = A_STAR_MR;
}
}