inline void HermitianTridiagU( Matrix<R>& A )
{
#ifndef RELEASE
PushCallStack("HermitianTridiagU");
if( A.Height() != A.Width() )
throw std::logic_error( "A must be square." );
#endif
// Matrix views
Matrix<R>
ATL, ATR, A00, a01, A02, a01T,
ABL, ABR, a10, alpha11, a12, alpha01B,
A20, a21, A22;
// Temporary matrices
Matrix<R> w01;
PushBlocksizeStack( 1 );
PartitionUpDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ABR.Height()+1 < A.Height() )
{
RepartitionUpDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
PartitionUp
( a01, a01T,
alpha01B, 1 );
w01.ResizeTo( a01.Height(), 1 );
//--------------------------------------------------------------------//
const R tau = Reflector( alpha01B, a01T );
const R epsilon1 = alpha01B.Get(0,0);
alpha01B.Set(0,0,R(1));
Symv( UPPER, tau, A00, a01, R(0), w01 );
const R alpha = -tau*Dot( w01, a01 )/R(2);
Axpy( alpha, a01, w01 );
Syr2( UPPER, R(-1), a01, w01, A00 );
alpha01B.Set(0,0,epsilon1);
//--------------------------------------------------------------------//
SlidePartitionUpDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelQR( Matrix<Real>& A )
{
#ifndef RELEASE
PushCallStack("internal::PanelQR");
#endif
Matrix<Real>
ATL, ATR, A00, a01, A02, aLeftCol, ARightPan,
ABL, ABR, a10, alpha11, a12,
A20, a21, A22;
Matrix<Real> z;
PushBlocksizeStack( 1 );
PartitionDownLeftDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
aLeftCol.View2x1( alpha11,
a21 );
ARightPan.View2x1( a12,
A22 );
Zeros( ARightPan.Width(), 1, z );
//--------------------------------------------------------------------//
const Real tau = Reflector( alpha11, a21 );
const Real alpha = alpha11.Get(0,0);
alpha11.Set(0,0,1);
Gemv( TRANSPOSE, Real(1), ARightPan, aLeftCol, Real(0), z );
Ger( -tau, aLeftCol, z, ARightPan );
alpha11.Set(0,0,alpha);
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelLQ( Matrix<Real>& A )
{
#ifndef RELEASE
PushCallStack("internal::PanelLQ");
#endif
Matrix<Real>
ATL, ATR, A00, a01, A02, aTopRow, ABottomPan,
ABL, ABR, a10, alpha11, a12,
A20, a21, A22;
Matrix<Real> z;
PushBlocksizeStack( 1 );
PartitionDownLeftDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
aTopRow.View1x2( alpha11, a12 );
ABottomPan.View1x2( a21, A22 );
Zeros( ABottomPan.Height(), 1, z );
//--------------------------------------------------------------------//
const Real tau = Reflector( alpha11, a12 );
const Real alpha = alpha11.Get(0,0);
alpha11.Set(0,0,1);
Gemv( NORMAL, Real(1), ABottomPan, aTopRow, Real(0), z );
Ger( -tau, z, aTopRow, ABottomPan );
alpha11.Set(0,0,alpha);
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LU( Matrix<F>& A )
{
#ifndef RELEASE
PushCallStack("LU");
#endif
// Matrix views
Matrix<F>
ATL, ATR, A00, a01, A02, alpha21T,
ABL, ABR, a10, alpha11, a12, a21B,
A20, a21, A22;
PushBlocksizeStack( 1 );
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
//--------------------------------------------------------------------//
F alpha = alpha11.Get(0,0);
if( alpha == static_cast<F>(0) )
throw SingularMatrixException();
Scal( static_cast<F>(1)/alpha, a21 );
Geru( (F)-1, a21, a12, A22 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void UnblockedBidiagU( DistMatrix<R>& A )
{
#ifndef RELEASE
PushCallStack("bidiag::UnblockedBidiagU");
if( A.Height() < A.Width() )
throw std::logic_error("A must be at least as tall as it is wide");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<R>
ATL(g), ATR(g), A00(g), a01(g), A02(g), alpha12L(g), a12R(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g), aB1(g), AB2(g),
A20(g), a21(g), A22(g);
// Temporary matrices
DistMatrix<R,STAR,MR > a12_STAR_MR(g);
DistMatrix<R,MC, STAR> aB1_MC_STAR(g);
DistMatrix<R,MR, STAR> x12Trans_MR_STAR(g);
DistMatrix<R,MC, STAR> w21_MC_STAR(g);
PushBlocksizeStack( 1 );
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
View2x1
( aB1, alpha11,
a21 );
View2x1
( AB2, a12,
A22 );
aB1_MC_STAR.AlignWith( aB1 );
a12_STAR_MR.AlignWith( a12 );
x12Trans_MR_STAR.AlignWith( AB2 );
w21_MC_STAR.AlignWith( A22 );
Zeros( a12.Width(), 1, x12Trans_MR_STAR );
Zeros( a21.Height(), 1, w21_MC_STAR );
const bool thisIsMyRow = ( g.Row() == alpha11.ColAlignment() );
const bool thisIsMyCol = ( g.Col() == alpha11.RowAlignment() );
const bool nextIsMyCol = ( g.Col() == a12.RowAlignment() );
//--------------------------------------------------------------------//
// Find tauQ, u, and epsilonQ such that
// I - tauQ | 1 | | 1, u^T | | alpha11 | = | epsilonQ |
// | u | | a21 | = | 0 |
const R tauQ = Reflector( alpha11, a21 );
R epsilonQ=0;
if( thisIsMyCol && thisIsMyRow )
epsilonQ = alpha11.GetLocal(0,0);
// Set aB1 = | 1 | and form x12^T := (aB1^T AB2)^T = AB2^T aB1
// | u |
alpha11.Set(0,0,R(1));
aB1_MC_STAR = aB1;
internal::LocalGemv
( TRANSPOSE, R(1), AB2, aB1_MC_STAR, R(0), x12Trans_MR_STAR );
x12Trans_MR_STAR.SumOverCol();
// Update AB2 := AB2 - tauQ aB1 x12
// = AB2 - tauQ aB1 aB1^T AB2
// = (I - tauQ aB1 aB1^T) AB2
internal::LocalGer( -tauQ, aB1_MC_STAR, x12Trans_MR_STAR, AB2 );
// Put epsilonQ back instead of the temporary value, 1
if( thisIsMyCol && thisIsMyRow )
alpha11.SetLocal(0,0,epsilonQ);
if( A22.Width() != 0 )
{
// Expose the subvector we seek to zero, a12R
PartitionRight( a12, alpha12L, a12R );
// Find tauP, v, and epsilonP such that
// I - tauP | 1 | | 1, v^T | | alpha12L | = | epsilonP |
// | v | | a12R^T | = | 0 |
const R tauP = Reflector( alpha12L, a12R );
R epsilonP=0;
if( nextIsMyCol && thisIsMyRow )
epsilonP = alpha12L.GetLocal(0,0);
// Set a12^T = | 1 | and form w21 := A22 a12^T = A22 | 1 |
// | v | | v |
alpha12L.Set(0,0,R(1));
a12_STAR_MR = a12;
internal::LocalGemv
( NORMAL, R(1), A22, a12_STAR_MR, R(0), w21_MC_STAR );
w21_MC_STAR.SumOverRow();
// A22 := A22 - tauP w21 a12
// = A22 - tauP A22 a12^T a12
// = A22 (I - tauP a12^T a12)
internal::LocalGer( -tauP, w21_MC_STAR, a12_STAR_MR, A22 );
// Put epsilonP back instead of the temporary value, 1
if( nextIsMyCol && thisIsMyRow )
alpha12L.SetLocal(0,0,epsilonP);
}
//--------------------------------------------------------------------//
aB1_MC_STAR.FreeAlignments();
a12_STAR_MR.FreeAlignments();
x12Trans_MR_STAR.FreeAlignments();
w21_MC_STAR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelLU
( DistMatrix<F, STAR,STAR>& A,
DistMatrix<F, MC, STAR>& B,
DistMatrix<int,STAR,STAR>& p,
int pivotOffset )
{
#ifndef RELEASE
PushCallStack("internal::PanelLU");
if( A.Grid() != p.Grid() || p.Grid() != B.Grid() )
throw std::logic_error
("Matrices must be distributed over the same grid");
if( A.Width() != B.Width() )
throw std::logic_error("A and B must be the same width");
if( A.Height() != p.Height() || p.Width() != 1 )
throw std::logic_error("p must be a vector that conforms with A");
#endif
const Grid& g = A.Grid();
const int r = g.Height();
const int colShift = B.ColShift();
const int colAlignment = B.ColAlignment();
// Matrix views
DistMatrix<F,STAR,STAR>
ATL(g), ATR(g), A00(g), a01(g), A02(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g),
A20(g), a21(g), A22(g);
DistMatrix<F,MC,STAR>
BL(g), BR(g),
B0(g), b1(g), B2(g);
const int width = A.Width();
const int numBytes = (width+1)*sizeof(F)+sizeof(int);
std::vector<byte> sendData(numBytes);
std::vector<byte> recvData(numBytes);
// Extract pointers to send and recv data
// TODO: Think of how to make this safer with respect to alignment issues
F* sendBufFloat = (F*)&sendData[0];
F* recvBufFloat = (F*)&recvData[0];
int* sendBufInt = (int*)&sendData[(width+1)*sizeof(F)];
int* recvBufInt = (int*)&recvData[(width+1)*sizeof(F)];
// Start the algorithm
PushBlocksizeStack( 1 );
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionRight( B, BL, BR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
RepartitionRight
( BL, /**/ BR,
B0, /**/ b1, B2 );
//--------------------------------------------------------------------//
const int currentRow = a01.Height();
// Store the index/value of the pivot candidate in A
F pivot = alpha11.GetLocal(0,0);
int pivotRow = currentRow;
for( int i=0; i<a21.Height(); ++i )
{
F value = a21.GetLocal(i,0);
if( FastAbs(value) > FastAbs(pivot) )
{
pivot = value;
pivotRow = currentRow + i + 1;
}
}
// Update the pivot candidate to include local data from B
for( int i=0; i<B.LocalHeight(); ++i )
{
F value = b1.GetLocal(i,0);
if( FastAbs(value) > FastAbs(pivot) )
{
pivot = value;
pivotRow = A.Height() + colShift + i*r;
}
}
// Fill the send buffer with:
// [ pivotValue | pivot row data | pivotRow ]
if( pivotRow < A.Height() )
{
sendBufFloat[0] = A.GetLocal(pivotRow,a10.Width());
const int ALDim = A.LocalLDim();
const F* ABuffer = A.LocalBuffer(pivotRow,0);
for( int j=0; j<width; ++j )
sendBufFloat[j+1] = ABuffer[j*ALDim];
}
else
{
const int localRow = ((pivotRow-A.Height())-colShift)/r;
sendBufFloat[0] = b1.GetLocal(localRow,0);
const int BLDim = B.LocalLDim();
const F* BBuffer = B.LocalBuffer(localRow,0);
for( int j=0; j<width; ++j )
sendBufFloat[j+1] = BBuffer[j*BLDim];
}
*sendBufInt = pivotRow;
// Communicate to establish the pivot information
mpi::AllReduce
( &sendData[0], &recvData[0], numBytes, PivotOp<F>(), g.ColComm() );
// Update the pivot vector
pivotRow = *recvBufInt;
p.SetLocal(currentRow,0,pivotRow+pivotOffset);
// Copy the current row into the pivot row
if( pivotRow < A.Height() )
{
const int ALDim = A.LocalLDim();
F* ASetBuffer = A.LocalBuffer(pivotRow,0);
const F* AGetBuffer = A.LocalBuffer(currentRow,0);
for( int j=0; j<width; ++j )
ASetBuffer[j*ALDim] = AGetBuffer[j*ALDim];
}
else
{
const int ownerRank = (colAlignment+(pivotRow-A.Height())) % r;
if( g.Row() == ownerRank )
{
const int localRow = ((pivotRow-A.Height())-colShift) / r;
const int ALDim = A.LocalLDim();
const int BLDim = B.LocalLDim();
F* BBuffer = B.LocalBuffer(localRow,0);
const F* ABuffer = A.LocalBuffer(currentRow,0);
for( int j=0; j<width; ++j )
BBuffer[j*BLDim] = ABuffer[j*ALDim];
}
}
// Copy the pivot row into the current row
{
F* ABuffer = A.LocalBuffer(currentRow,0);
const int ALDim = A.LocalLDim();
for( int j=0; j<width; ++j )
ABuffer[j*ALDim] = recvBufFloat[j+1];
}
// Now we can perform the update of the current panel
const F alpha = alpha11.GetLocal(0,0);
if( alpha == F(0) )
throw SingularMatrixException();
const F alpha11Inv = F(1) / alpha;
Scale( alpha11Inv, a21.LocalMatrix() );
Scale( alpha11Inv, b1.LocalMatrix() );
Geru( F(-1), a21.LocalMatrix(), a12.LocalMatrix(), A22.LocalMatrix() );
Geru( F(-1), b1.LocalMatrix(), a12.LocalMatrix(), B2.LocalMatrix() );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
SlidePartitionRight
( BL, /**/ BR,
B0, b1, /**/ B2 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelLU( Matrix<F>& A, Matrix<int>& p, int pivotOffset )
{
#ifndef RELEASE
PushCallStack("internal::PanelLU");
if( A.Width() != p.Height() || p.Width() != 1 )
throw std::logic_error("p must be a vector that conforms with A");
#endif
// Matrix views
Matrix<F>
ATL, ATR, A00, a01, A02,
ABL, ABR, a10, alpha11, a12,
A20, a21, A22;
const int width = A.Width();
std::vector<F> buffer( width );
// Start the algorithm
PushBlocksizeStack( 1 );
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Height() < A.Height() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
//--------------------------------------------------------------------//
const int currentRow = A00.Height();
// Find the index and value of the pivot candidate
F pivot = alpha11.Get(0,0);
int pivotRow = currentRow;
for( int i=0; i<a21.Height(); ++i )
{
const F value = a21.Get(i,0);
if( FastAbs(value) > FastAbs(pivot) )
{
pivot = value;
pivotRow = currentRow + i + 1;
}
}
p.Set( currentRow, 0, pivotRow+pivotOffset );
// Swap the pivot row and current row
for( int j=0; j<width; ++j )
{
buffer[j] = A.Get(currentRow,j);
A.Set(currentRow,j,A.Get(pivotRow,j));
A.Set(pivotRow,j,buffer[j]);
}
// Now we can perform the update of the current panel
const F alpha = alpha11.Get(0,0);
if( alpha == F(0) )
throw SingularMatrixException();
const F alpha11Inv = F(1) / alpha;
Scale( alpha11Inv, a21 );
Geru( F(-1), a21, a12, A22 );
//--------------------------------------------------------------------//
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LocalTrmmAccumulateRUN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
const DistMatrix<T,MC, MR >& U,
const DistMatrix<T,STAR,MC >& X_STAR_MC,
DistMatrix<T,MR, STAR>& ZTrans_MR_STAR )
{
#ifndef RELEASE
CallStackEntry entry("internal::LocalTrmmAccumulateRUN");
if( U.Grid() != X_STAR_MC.Grid() ||
X_STAR_MC.Grid() != ZTrans_MR_STAR.Grid() )
throw std::logic_error
("{U,X,Z} must be distributed over the same grid");
if( U.Height() != U.Width() ||
U.Height() != X_STAR_MC.Width() ||
U.Height() != ZTrans_MR_STAR.Height() )
{
std::ostringstream msg;
msg << "Nonconformal LocalTrmmAccumulateRUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X[* ,MC] ~ " << X_STAR_MC.Height() << " x "
<< X_STAR_MC.Width() << "\n"
<< " Z^H/T[MR,* ] ~ " << ZTrans_MR_STAR.Height() << " x "
<< ZTrans_MR_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( X_STAR_MC.RowAlignment() != U.ColAlignment() ||
ZTrans_MR_STAR.ColAlignment() != U.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<T> D11(g);
DistMatrix<T,STAR,MC>
XL_STAR_MC(g), XR_STAR_MC(g),
X0_STAR_MC(g), X1_STAR_MC(g), X2_STAR_MC(g);
DistMatrix<T,MR,STAR>
ZTTrans_MR_STAR(g), Z0Trans_MR_STAR(g),
ZBTrans_MR_STAR(g), Z1Trans_MR_STAR(g),
Z2Trans_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
LockedPartitionRight( X_STAR_MC, XL_STAR_MC, XR_STAR_MC, 0 );
PartitionDown
( ZTrans_MR_STAR, ZTTrans_MR_STAR,
ZBTrans_MR_STAR, 0 );
while( UTL.Height() < U.Height() )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
LockedRepartitionRight
( XL_STAR_MC, /**/ XR_STAR_MC,
X0_STAR_MC, /**/ X1_STAR_MC, X2_STAR_MC );
RepartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
/***************/ /***************/
Z1Trans_MR_STAR,
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
D11.AlignWith( U11 );
//--------------------------------------------------------------------//
D11 = U11;
MakeTriangular( UPPER, D11 );
if( diag == UNIT )
SetDiagonal( D11, T(1) );
LocalGemm
( orientation, orientation,
alpha, D11, X1_STAR_MC, T(1), Z1Trans_MR_STAR );
LocalGemm
( orientation, orientation,
alpha, U01, X0_STAR_MC, T(1), Z1Trans_MR_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlideLockedPartitionRight
( XL_STAR_MC, /**/ XR_STAR_MC,
X0_STAR_MC, X1_STAR_MC, /**/ X2_STAR_MC );
SlidePartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
Z1Trans_MR_STAR,
/***************/ /***************/
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
}
PopBlocksizeStack();
}
inline void
PanelLQ( DistMatrix<Real>& A )
{
#ifndef RELEASE
PushCallStack("internal::PanelLQ");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<Real>
ATL(g), ATR(g), A00(g), a01(g), A02(g), aTopRow(g), ABottomPan(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g),
A20(g), a21(g), A22(g);
// Temporary distributions
DistMatrix<Real,STAR,MR> aTopRow_STAR_MR(g);
DistMatrix<Real,MC,STAR> z_MC_STAR(g);
PushBlocksizeStack( 1 );
PartitionDownLeftDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
aTopRow.View1x2( alpha11, a12 );
ABottomPan.View1x2( a21, A22 );
aTopRow_STAR_MR.AlignWith( ABottomPan );
z_MC_STAR.AlignWith( ABottomPan );
Zeros( ABottomPan.Height(), 1, z_MC_STAR );
//--------------------------------------------------------------------//
const Real tau = Reflector( alpha11, a12 );
const bool myDiagonalEntry = ( g.Row() == alpha11.ColAlignment() &&
g.Col() == alpha11.RowAlignment() );
Real alpha = 0;
if( myDiagonalEntry )
{
alpha = alpha11.GetLocal(0,0);
alpha11.SetLocal(0,0,1);
}
aTopRow_STAR_MR = aTopRow;
Gemv
( NORMAL,
Real(1), ABottomPan.LockedLocalMatrix(),
aTopRow_STAR_MR.LockedLocalMatrix(),
Real(0), z_MC_STAR.LocalMatrix() );
z_MC_STAR.SumOverRow();
Ger
( -tau,
z_MC_STAR.LockedLocalMatrix(),
aTopRow_STAR_MR.LockedLocalMatrix(),
ABottomPan.LocalMatrix() );
if( myDiagonalEntry )
alpha11.SetLocal(0,0,alpha);
//--------------------------------------------------------------------//
aTopRow_STAR_MR.FreeAlignments();
z_MC_STAR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LocalSymmetricAccumulateLU
( Orientation orientation, T alpha,
const DistMatrix<T>& A,
const DistMatrix<T,MC, STAR>& B_MC_STAR,
const DistMatrix<T,STAR,MR >& BAdjOrTrans_STAR_MR,
DistMatrix<T,MC, STAR>& Z_MC_STAR,
DistMatrix<T,MR, STAR>& Z_MR_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalSymmetricAccumulateLU");
if( A.Grid() != B_MC_STAR.Grid() ||
B_MC_STAR.Grid() != BAdjOrTrans_STAR_MR.Grid() ||
BAdjOrTrans_STAR_MR.Grid() != Z_MC_STAR.Grid() ||
Z_MC_STAR.Grid() != Z_MR_STAR.Grid() )
throw std::logic_error
("{A,B,Z} must be distributed over the same grid");
if( A.Height() != A.Width() ||
A.Height() != B_MC_STAR.Height() ||
A.Height() != BAdjOrTrans_STAR_MR.Width() ||
A.Height() != Z_MC_STAR.Height() ||
A.Height() != Z_MR_STAR.Height() ||
B_MC_STAR.Width() != BAdjOrTrans_STAR_MR.Height() ||
BAdjOrTrans_STAR_MR.Height() != Z_MC_STAR.Width() ||
Z_MC_STAR.Width() != Z_MR_STAR.Width() )
{
std::ostringstream msg;
msg << "Nonconformal LocalSymmetricAccumulateLU: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B[MC,* ] ~ " << B_MC_STAR.Height() << " x "
<< B_MC_STAR.Width() << "\n"
<< " B^H/T[* ,MR] ~ " << BAdjOrTrans_STAR_MR.Height() << " x "
<< BAdjOrTrans_STAR_MR.Width() << "\n"
<< " Z[MC,* ] ~ " << Z_MC_STAR.Height() << " x "
<< Z_MC_STAR.Width() << "\n"
<< " Z[MR,* ] ` " << Z_MR_STAR.Height() << " x "
<< Z_MR_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( B_MC_STAR.ColAlignment() != A.ColAlignment() ||
BAdjOrTrans_STAR_MR.RowAlignment() != A.RowAlignment() ||
Z_MC_STAR.ColAlignment() != A.ColAlignment() ||
Z_MR_STAR.ColAlignment() != A.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = A.Grid();
DistMatrix<T>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<T> D11(g);
DistMatrix<T,MC,STAR>
BT_MC_STAR(g), B0_MC_STAR(g),
BB_MC_STAR(g), B1_MC_STAR(g),
B2_MC_STAR(g);
DistMatrix<T,STAR,MR>
BLAdjOrTrans_STAR_MR(g), BRAdjOrTrans_STAR_MR(g),
B0AdjOrTrans_STAR_MR(g), B1AdjOrTrans_STAR_MR(g),
B2AdjOrTrans_STAR_MR(g);
DistMatrix<T,MC,STAR>
ZT_MC_STAR(g), Z0_MC_STAR(g),
ZB_MC_STAR(g), Z1_MC_STAR(g),
Z2_MC_STAR(g);
DistMatrix<T,MR,STAR>
ZT_MR_STAR(g), Z0_MR_STAR(g),
ZB_MR_STAR(g), Z1_MR_STAR(g),
Z2_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionDown
( B_MC_STAR, BT_MC_STAR,
BB_MC_STAR, 0 );
LockedPartitionRight
( BAdjOrTrans_STAR_MR, BLAdjOrTrans_STAR_MR, BRAdjOrTrans_STAR_MR, 0 );
PartitionDown
( Z_MC_STAR, ZT_MC_STAR,
ZB_MC_STAR, 0 );
PartitionDown
( Z_MR_STAR, ZT_MR_STAR,
ZB_MR_STAR, 0 );
while( ATL.Height() < A.Height() )
{
LockedRepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionDown
( BT_MC_STAR, B0_MC_STAR,
/**********/ /**********/
B1_MC_STAR,
BB_MC_STAR, B2_MC_STAR );
LockedRepartitionRight
( BLAdjOrTrans_STAR_MR, /**/ BRAdjOrTrans_STAR_MR,
B0AdjOrTrans_STAR_MR, /**/ B1AdjOrTrans_STAR_MR,
B2AdjOrTrans_STAR_MR );
RepartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
/**********/ /**********/
Z1_MC_STAR,
ZB_MC_STAR, Z2_MC_STAR );
RepartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
/**********/ /**********/
Z1_MR_STAR,
ZB_MR_STAR, Z2_MR_STAR );
D11.AlignWith( A11 );
//--------------------------------------------------------------------//
D11 = A11;
MakeTrapezoidal( LEFT, UPPER, 0, D11 );
LocalGemm
( NORMAL, orientation,
alpha, D11, B1AdjOrTrans_STAR_MR, T(1), Z1_MC_STAR );
MakeTrapezoidal( LEFT, UPPER, 1, D11 );
LocalGemm
( orientation, NORMAL, alpha, D11, B1_MC_STAR, T(1), Z1_MR_STAR );
LocalGemm
( NORMAL, orientation,
alpha, A12, B2AdjOrTrans_STAR_MR, T(1), Z1_MC_STAR );
LocalGemm
( orientation, NORMAL, alpha, A12, B1_MC_STAR, T(1), Z2_MR_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionDown
( BT_MC_STAR, B0_MC_STAR,
B1_MC_STAR,
/**********/ /**********/
BB_MC_STAR, B2_MC_STAR );
SlideLockedPartitionRight
( BLAdjOrTrans_STAR_MR, /**/ BRAdjOrTrans_STAR_MR,
B0AdjOrTrans_STAR_MR, B1AdjOrTrans_STAR_MR, /**/ B2AdjOrTrans_STAR_MR );
SlidePartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
Z1_MC_STAR,
/**********/ /**********/
ZB_MC_STAR, Z2_MC_STAR );
SlidePartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
Z1_MR_STAR,
/**********/ /**********/
ZB_MR_STAR, Z2_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void HermitianTridiagU
( Matrix<Complex<R> >& A, Matrix<Complex<R> >& t )
{
#ifndef RELEASE
PushCallStack("HermitianTridiagU");
#endif
const int tHeight = std::max(A.Height()-1,0);
#ifndef RELEASE
if( A.Height() != A.Width() )
throw std::logic_error("A must be square");
if( t.Viewing() && (t.Height() != tHeight || t.Width() != 1) )
throw std::logic_error("t is of the wrong size");
#endif
typedef Complex<R> C;
if( !t.Viewing() )
t.ResizeTo( tHeight, 1 );
// Matrix views
Matrix<C>
ATL, ATR, A00, a01, A02, a01T,
ABL, ABR, a10, alpha11, a12, alpha01B,
A20, a21, A22;
// Temporary matrices
Matrix<C> w01;
PushBlocksizeStack( 1 );
PartitionUpDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ABR.Height()+1 < A.Height() )
{
RepartitionUpDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
PartitionUp
( a01, a01T,
alpha01B, 1 );
w01.ResizeTo( a01.Height(), 1 );
//--------------------------------------------------------------------//
const C tau = Reflector( alpha01B, a01T );
const R epsilon1 = alpha01B.GetRealPart(0,0);
t.Set(t.Height()-1-A22.Height(),0,tau);
alpha01B.Set(0,0,C(1));
Hemv( UPPER, tau, A00, a01, C(0), w01 );
const C alpha = -tau*Dot( w01, a01 )/C(2);
Axpy( alpha, a01, w01 );
Her2( UPPER, C(-1), a01, w01, A00 );
alpha01B.Set(0,0,epsilon1);
//--------------------------------------------------------------------//
SlidePartitionUpDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LocalSymvRowAccumulateU
( T alpha,
const DistMatrix<T>& A,
const DistMatrix<T,STAR,MC>& x_STAR_MC,
const DistMatrix<T,STAR,MR>& x_STAR_MR,
DistMatrix<T,STAR,MC>& z_STAR_MC,
DistMatrix<T,STAR,MR>& z_STAR_MR )
{
#ifndef RELEASE
PushCallStack("internal::LocalSymvRowAccumulateU");
if( A.Grid() != x_STAR_MC.Grid() ||
x_STAR_MC.Grid() != x_STAR_MR.Grid() ||
x_STAR_MR.Grid() != z_STAR_MC.Grid() ||
z_STAR_MC.Grid() != z_STAR_MR.Grid() )
throw std::logic_error
("{A,x,z} must be distributed over the same grid");
if( x_STAR_MC.Height() != 1 || x_STAR_MR.Height() != 1 ||
z_STAR_MC.Height() != 1 || z_STAR_MR.Height() != 1 )
throw std::logic_error("Expected x and z to be row vectors");
if( A.Height() != A.Width() ||
A.Height() != x_STAR_MC.Width() ||
A.Height() != x_STAR_MR.Width() ||
A.Height() != z_STAR_MC.Width() ||
A.Height() != z_STAR_MR.Width() )
{
std::ostringstream msg;
msg << "Nonconformal LocalSymvRowAccumulateU: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " x[* ,MC] ~ " << x_STAR_MC.Height() << " x "
<< x_STAR_MC.Width() << "\n"
<< " x[* ,MR] ~ " << x_STAR_MR.Height() << " x "
<< x_STAR_MR.Width() << "\n"
<< " z[* ,MC] ~ " << z_STAR_MC.Height() << " x "
<< z_STAR_MC.Width() << "\n"
<< " z[* ,MR] ~ " << z_STAR_MR.Height() << " x "
<< z_STAR_MR.Width() << "\n";
throw std::logic_error( msg.str() );
}
if( x_STAR_MC.RowAlignment() != A.ColAlignment() ||
x_STAR_MR.RowAlignment() != A.RowAlignment() ||
z_STAR_MC.RowAlignment() != A.ColAlignment() ||
z_STAR_MR.RowAlignment() != A.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<T> A11(g), A12(g);
DistMatrix<T> D11(g);
DistMatrix<T,STAR,MC> x1_STAR_MC(g);
DistMatrix<T,STAR,MR>
xL_STAR_MR(g), xR_STAR_MR(g),
x0_STAR_MR(g), x1_STAR_MR(g), x2_STAR_MR(g);
DistMatrix<T,STAR,MC> z1_STAR_MC(g);
DistMatrix<T,STAR,MR> z1_STAR_MR(g), z2_STAR_MR(g);
// We want our local gemvs to be of width blocksize, so we will
// temporarily change to max(r,c) times the current blocksize
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*LocalSymvBlocksize<T>() );
LockedPartitionRight( x_STAR_MR, xL_STAR_MR, xR_STAR_MR, 0 );
while( xL_STAR_MR.Width() < x_STAR_MR.Width() )
{
LockedRepartitionRight
( xL_STAR_MR, /**/ xR_STAR_MR,
x0_STAR_MR, /**/ x1_STAR_MR, x2_STAR_MR );
const int n0 = x0_STAR_MR.Width();
const int n1 = x1_STAR_MR.Width();
const int n2 = x2_STAR_MR.Width();
LockedView( A11, A, n0, n0, n1, n1 );
LockedView( A12, A, n0, n0+n1, n1, n2 );
LockedView( x1_STAR_MC, x_STAR_MC, 0, n0, 1, n1 );
View( z1_STAR_MC, z_STAR_MC, 0, n0, 1, n1 );
View( z1_STAR_MR, z_STAR_MR, 0, n0, 1, n1 );
View( z2_STAR_MR, z_STAR_MR, 0, n0+n1, 1, n2 );
D11.AlignWith( A11 );
//--------------------------------------------------------------------//
// TODO: These diagonal block updates can be greatly improved
D11 = A11;
MakeTrapezoidal( LEFT, UPPER, 0, D11 );
Gemv
( NORMAL,
alpha, D11.LockedLocalMatrix(),
x1_STAR_MR.LockedLocalMatrix(),
T(1), z1_STAR_MC.LocalMatrix() );
MakeTrapezoidal( LEFT, UPPER, 1, D11 );
Gemv
( TRANSPOSE,
alpha, D11.LockedLocalMatrix(),
x1_STAR_MC.LockedLocalMatrix(),
T(1), z1_STAR_MR.LocalMatrix() );
Gemv
( NORMAL,
alpha, A12.LockedLocalMatrix(),
x2_STAR_MR.LockedLocalMatrix(),
T(1), z1_STAR_MC.LocalMatrix() );
Gemv
( TRANSPOSE,
alpha, A12.LockedLocalMatrix(),
x1_STAR_MC.LockedLocalMatrix(),
T(1), z2_STAR_MR.LocalMatrix() );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionRight
( xL_STAR_MR, /**/ xR_STAR_MR,
x0_STAR_MR, x1_STAR_MR, /**/ x2_STAR_MR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
internal::LocalTrmmAccumulateLUN
( Orientation orientation, UnitOrNonUnit diag, T alpha,
const DistMatrix<T,MC, MR >& U,
const DistMatrix<T,STAR,MR >& XAdjOrTrans_STAR_MR,
DistMatrix<T,MC, STAR>& Z_MC_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalTrmmAccumulateLUN");
if( U.Grid() != XAdjOrTrans_STAR_MR.Grid() ||
XAdjOrTrans_STAR_MR.Grid() != Z_MC_STAR.Grid() )
throw std::logic_error
("{U,X,Z} must be distributed over the same grid");
if( U.Height() != U.Width() ||
U.Height() != XAdjOrTrans_STAR_MR.Width() ||
U.Height() != Z_MC_STAR.Height() ||
XAdjOrTrans_STAR_MR.Height() != Z_MC_STAR.Width() )
{
std::ostringstream msg;
msg << "Nonconformal LocalTrmmAccumulateLUN: \n"
<< " U ~ " << U.Height() << " x " << U.Width() << "\n"
<< " X^H/T[* ,MR] ~ " << XAdjOrTrans_STAR_MR.Height() << " x "
<< XAdjOrTrans_STAR_MR.Width() << "\n"
<< " Z[MC,* ] ~ " << Z_MC_STAR.Height() << " x "
<< Z_MC_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( XAdjOrTrans_STAR_MR.RowAlignment() != U.RowAlignment() ||
Z_MC_STAR.ColAlignment() != U.ColAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = U.Grid();
// Matrix views
DistMatrix<T,MC,MR>
UTL(g), UTR(g), U00(g), U01(g), U02(g),
UBL(g), UBR(g), U10(g), U11(g), U12(g),
U20(g), U21(g), U22(g);
DistMatrix<T,MC,MR> D11(g);
DistMatrix<T,STAR,MR>
XLAdjOrTrans_STAR_MR(g), XRAdjOrTrans_STAR_MR(g),
X0AdjOrTrans_STAR_MR(g), X1AdjOrTrans_STAR_MR(g),
X2AdjOrTrans_STAR_MR(g);
DistMatrix<T,MC,STAR>
ZT_MC_STAR(g), Z0_MC_STAR(g),
ZB_MC_STAR(g), Z1_MC_STAR(g),
Z2_MC_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
LockedPartitionRight
( XAdjOrTrans_STAR_MR, XLAdjOrTrans_STAR_MR, XRAdjOrTrans_STAR_MR, 0 );
PartitionDown
( Z_MC_STAR, ZT_MC_STAR,
ZB_MC_STAR, 0 );
while( UTL.Height() < U.Height() )
{
LockedRepartitionDownDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
LockedRepartitionRight
( XLAdjOrTrans_STAR_MR, /**/ XRAdjOrTrans_STAR_MR,
X0AdjOrTrans_STAR_MR, /**/ X1AdjOrTrans_STAR_MR, X2AdjOrTrans_STAR_MR
);
RepartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
/**********/ /**********/
Z1_MC_STAR,
ZB_MC_STAR, Z2_MC_STAR );
D11.AlignWith( U11 );
//--------------------------------------------------------------------//
D11 = U11;
MakeTrapezoidal( LEFT, UPPER, 0, D11 );
if( diag == UNIT )
SetDiagonalToOne( D11 );
internal::LocalGemm
( NORMAL, orientation, alpha, D11, X1AdjOrTrans_STAR_MR,
(T)1, Z1_MC_STAR );
internal::LocalGemm
( NORMAL, orientation, alpha, U01, X1AdjOrTrans_STAR_MR,
(T)1, Z0_MC_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
SlideLockedPartitionRight
( XLAdjOrTrans_STAR_MR, /**/ XRAdjOrTrans_STAR_MR,
X0AdjOrTrans_STAR_MR, X1AdjOrTrans_STAR_MR, /**/ X2AdjOrTrans_STAR_MR
);
SlidePartitionDown
( ZT_MC_STAR, Z0_MC_STAR,
Z1_MC_STAR,
/**********/ /**********/
ZB_MC_STAR, Z2_MC_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
void Initialize( int& argc, char**& argv )
{
if( ::numElemInits > 0 )
{
++::numElemInits;
return;
}
::args = new Args( argc, argv );
::numElemInits = 1;
if( !mpi::Initialized() )
{
if( mpi::Finalized() )
{
LogicError
("Cannot initialize elemental after finalizing MPI");
}
#ifdef EL_HYBRID
const Int provided =
mpi::InitializeThread
( argc, argv, mpi::THREAD_MULTIPLE );
const int commRank = mpi::Rank( mpi::COMM_WORLD );
if( provided != mpi::THREAD_MULTIPLE && commRank == 0 )
{
cerr << "WARNING: Could not achieve THREAD_MULTIPLE support."
<< endl;
}
#else
mpi::Initialize( argc, argv );
#endif
::elemInitializedMpi = true;
}
else
{
#ifdef EL_HYBRID
const Int provided = mpi::QueryThread();
if( provided != mpi::THREAD_MULTIPLE )
{
throw std::runtime_error
("MPI initialized with inadequate thread support for Elemental");
}
#endif
}
#ifdef EL_HAVE_QT5
InitializeQt5( argc, argv );
#endif
// Queue a default algorithmic blocksize
EmptyBlocksizeStack();
PushBlocksizeStack( 128 );
// Build the default grid
Grid::InitializeDefault();
#ifdef EL_HAVE_QD
InitializeQD();
#endif
InitializeRandom();
// Create the types and ops
// NOTE: mpfr::SetPrecision created the BigFloat types
mpi::CreateCustom();
}
inline void
LocalSymvColAccumulateU
( T alpha,
const DistMatrix<T>& A,
const DistMatrix<T,MC,STAR>& x_MC_STAR,
const DistMatrix<T,MR,STAR>& x_MR_STAR,
DistMatrix<T,MC,STAR>& z_MC_STAR,
DistMatrix<T,MR,STAR>& z_MR_STAR,
bool conjugate=false )
{
#ifndef RELEASE
CallStackEntry entry("internal::LocalSymvColAccumulateU");
if( A.Grid() != x_MC_STAR.Grid() ||
x_MC_STAR.Grid() != x_MR_STAR.Grid() ||
x_MR_STAR.Grid() != z_MC_STAR.Grid() ||
z_MC_STAR.Grid() != z_MR_STAR.Grid() )
LogicError
("{A,x,z} must be distributed over the same grid");
if( x_MC_STAR.Width() != 1 || x_MR_STAR.Width() != 1 ||
z_MC_STAR.Width() != 1 || z_MR_STAR.Width() != 1 )
LogicError("Expected x and z to be column vectors");
if( A.Height() != A.Width() ||
A.Height() != x_MC_STAR.Height() ||
A.Height() != x_MR_STAR.Height() ||
A.Height() != z_MC_STAR.Height() ||
A.Height() != z_MR_STAR.Height() )
{
std::ostringstream msg;
msg << "Nonconformal LocalSymvColAccumulateU: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " x[MC,* ] ~ " << x_MC_STAR.Height() << " x "
<< x_MC_STAR.Width() << "\n"
<< " x[MR,* ] ~ " << x_MR_STAR.Height() << " x "
<< x_MR_STAR.Width() << "\n"
<< " z[MC,* ] ~ " << z_MC_STAR.Height() << " x "
<< z_MC_STAR.Width() << "\n"
<< " z[MR,* ] ~ " << z_MR_STAR.Height() << " x "
<< z_MR_STAR.Width() << "\n";
LogicError( msg.str() );
}
if( x_MC_STAR.ColAlignment() != A.ColAlignment() ||
x_MR_STAR.ColAlignment() != A.RowAlignment() ||
z_MC_STAR.ColAlignment() != A.ColAlignment() ||
z_MR_STAR.ColAlignment() != A.RowAlignment() )
LogicError("Partial matrix distributions are misaligned");
#endif
const Grid& g = A.Grid();
const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );
// Matrix views
DistMatrix<T> A11(g), A12(g);
DistMatrix<T> D11(g);
DistMatrix<T,MC,STAR> x1_MC_STAR(g);
DistMatrix<T,MR,STAR>
xT_MR_STAR(g), x0_MR_STAR(g),
xB_MR_STAR(g), x1_MR_STAR(g),
x2_MR_STAR(g);
DistMatrix<T,MC,STAR> z1_MC_STAR(g);
DistMatrix<T,MR,STAR> z1_MR_STAR(g),
z2_MR_STAR(g);
// We want our local gemvs to be of width blocksize, so we will
// temporarily change to max(r,c) times the current blocksize
const Int ratio = Max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*LocalSymvBlocksize<T>() );
LockedPartitionDown
( x_MR_STAR, xT_MR_STAR,
xB_MR_STAR, 0 );
while( xT_MR_STAR.Height() < x_MR_STAR.Height() )
{
LockedRepartitionDown
( xT_MR_STAR, x0_MR_STAR,
/**********/ /**********/
x1_MR_STAR,
xB_MR_STAR, x2_MR_STAR );
const Int n0 = x0_MR_STAR.Height();
const Int n1 = x1_MR_STAR.Height();
const Int n2 = x2_MR_STAR.Height();
LockedView( A11, A, n0, n0, n1, n1 );
LockedView( A12, A, n0, n0+n1, n1, n2 );
LockedView( x1_MC_STAR, x_MC_STAR, n0, 0, n1, 1 );
View( z1_MC_STAR, z_MC_STAR, n0, 0, n1, 1 );
View( z1_MR_STAR, z_MR_STAR, n0, 0, n1, 1 );
View( z2_MR_STAR, z_MR_STAR, n0+n1, 0, n2, 1 );
D11.AlignWith( A11 );
//--------------------------------------------------------------------//
// TODO: These diagonal block updates can be greatly improved
D11 = A11;
MakeTriangular( UPPER, D11 );
LocalGemv( NORMAL, alpha, D11, x1_MR_STAR, T(1), z1_MC_STAR );
SetDiagonal( D11, T(0) );
LocalGemv( orientation, alpha, D11, x1_MC_STAR, T(1), z1_MR_STAR );
LocalGemv( NORMAL, alpha, A12, x2_MR_STAR, T(1), z1_MC_STAR );
LocalGemv( orientation, alpha, A12, x1_MC_STAR, T(1), z2_MR_STAR );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( xT_MR_STAR, x0_MR_STAR,
x1_MR_STAR,
/**********/ /**********/
xB_MR_STAR, x2_MR_STAR );
}
PopBlocksizeStack();
}
inline void UnblockedBidiagU
( DistMatrix<Complex<R> >& A,
DistMatrix<Complex<R>,MD,STAR>& tP,
DistMatrix<Complex<R>,MD,STAR>& tQ )
{
#ifndef RELEASE
PushCallStack("BidiagU");
#endif
const int tPHeight = std::max(A.Width()-1,0);
const int tQHeight = A.Width();
#ifndef RELEASE
if( A.Grid() != tP.Grid() || tP.Grid() != tQ.Grid() )
throw std::logic_error("Process grids do not match");
if( A.Height() < A.Width() )
throw std::logic_error("A must be at least as tall as it is wide");
if( tP.Viewing() && (tP.Height() != tPHeight || tP.Width() != 1) )
throw std::logic_error("tP is the wrong height");
if( tQ.Viewing() && (tQ.Height() != tQHeight || tQ.Width() != 1) )
throw std::logic_error("tQ is the wrong height");
#endif
typedef Complex<R> C;
const Grid& g = A.Grid();
if( !tP.Viewing() )
tP.ResizeTo( tPHeight, 1 );
if( !tQ.Viewing() )
tQ.ResizeTo( tQHeight, 1 );
// Matrix views
DistMatrix<C>
ATL(g), ATR(g), A00(g), a01(g), A02(g), alpha12L(g), a12R(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g), aB1(g), AB2(g),
A20(g), a21(g), A22(g);
// Temporary matrices
DistMatrix<C,STAR,MR > a12_STAR_MR(g);
DistMatrix<C,MC, STAR> aB1_MC_STAR(g);
DistMatrix<C,MR, STAR> x12Adj_MR_STAR(g);
DistMatrix<C,MC, STAR> w21_MC_STAR(g);
PushBlocksizeStack( 1 );
PartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
while( ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
View2x1
( aB1, alpha11,
a21 );
View2x1
( AB2, a12,
A22 );
aB1_MC_STAR.AlignWith( aB1 );
a12_STAR_MR.AlignWith( a12 );
x12Adj_MR_STAR.AlignWith( AB2 );
w21_MC_STAR.AlignWith( A22 );
Zeros( a12.Width(), 1, x12Adj_MR_STAR );
Zeros( a21.Height(), 1, w21_MC_STAR );
const bool thisIsMyRow = ( g.Row() == alpha11.ColAlignment() );
const bool thisIsMyCol = ( g.Col() == alpha11.RowAlignment() );
const bool nextIsMyCol = ( g.Col() == a12.RowAlignment() );
//--------------------------------------------------------------------//
// Find tauQ, u, and epsilonQ such that
// I - conj(tauQ) | 1 | | 1, u^H | | alpha11 | = | epsilonQ |
// | u | | a21 | | 0 |
const C tauQ = Reflector( alpha11, a21 );
tQ.Set(A00.Height(),0,tauQ );
C epsilonQ=0;
if( thisIsMyCol && thisIsMyRow )
epsilonQ = alpha11.GetLocal(0,0);
// Set aB1 = | 1 | and form x12^H := (aB1^H AB2)^H = AB2^H aB1
// | u |
alpha11.Set(0,0,C(1));
aB1_MC_STAR = aB1;
internal::LocalGemv
( ADJOINT, C(1), AB2, aB1_MC_STAR, C(0), x12Adj_MR_STAR );
x12Adj_MR_STAR.SumOverCol();
// Update AB2 := AB2 - conj(tauQ) aB1 x12
// = AB2 - conj(tauQ) aB1 aB1^H AB2
// = (I - conj(tauQ) aB1 aB1^H) AB2
internal::LocalGer( -Conj(tauQ), aB1_MC_STAR, x12Adj_MR_STAR, AB2 );
// Put epsilonQ back instead of the temporary value, 1
if( thisIsMyCol && thisIsMyRow )
alpha11.SetLocal(0,0,epsilonQ);
if( A22.Width() != 0 )
{
// Due to the deficiencies in the BLAS ?gemv routines, this section
// is easier if we temporarily conjugate a12
Conjugate( a12 );
// Expose the subvector we seek to zero, a12R
PartitionRight( a12, alpha12L, a12R );
// Find tauP, v, and epsilonP such that
// I - conj(tauP) | 1 | | 1, v^H | | alpha12L | = | epsilonP |
// | v | | a12R^T | | 0 |
const C tauP = Reflector( alpha12L, a12R );
tP.Set(A00.Height(),0,tauP);
C epsilonP=0;
if( nextIsMyCol && thisIsMyRow )
epsilonP = alpha12L.GetLocal(0,0);
// Set a12^T = | 1 | and form w21 := A22 a12^T = A22 | 1 |
// | v | | v |
alpha12L.Set(0,0,C(1));
a12_STAR_MR = a12;
internal::LocalGemv
( NORMAL, C(1), A22, a12_STAR_MR, C(0), w21_MC_STAR );
w21_MC_STAR.SumOverRow();
// A22 := A22 - tauP w21 conj(a12)
// = A22 - tauP A22 a12^T conj(a12)
// = A22 (I - tauP a12^T conj(a12))
// = A22 conj(I - conj(tauP) a12^H a12)
// which compensates for the fact that the reflector was generated
// on the conjugated a12.
internal::LocalGer( -tauP, w21_MC_STAR, a12_STAR_MR, A22 );
// Put epsilonP back instead of the temporary value, 1
if( nextIsMyCol && thisIsMyRow )
alpha12L.SetLocal(0,0,epsilonP);
// Undue the temporary conjugation
Conjugate( a12 );
}
//--------------------------------------------------------------------//
aB1_MC_STAR.FreeAlignments();
a12_STAR_MR.FreeAlignments();
x12Adj_MR_STAR.FreeAlignments();
w21_MC_STAR.FreeAlignments();
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelLQ
( Matrix<Complex<Real> >& A,
Matrix<Complex<Real> >& t )
{
#ifndef RELEASE
PushCallStack("internal::PanelLQ");
if( t.Height() != std::min(A.Height(),A.Width()) || t.Width() != 1 )
throw std::logic_error
("t must be a vector of height equal to the minimum dimension of A");
#endif
typedef Complex<Real> C;
Matrix<C>
ATL, ATR, A00, a01, A02, aTopRow, ABottomPan,
ABL, ABR, a10, alpha11, a12,
A20, a21, A22;
Matrix<C>
tT, t0,
tB, tau1,
t2;
Matrix<C> z, aTopRowConj;
PushBlocksizeStack( 1 );
PartitionDownLeftDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
RepartitionDown
( tT, t0,
/**/ /****/
tau1,
tB, t2 );
aTopRow.View1x2( alpha11, a12 );
ABottomPan.View1x2( a21, A22 );
Zeros( ABottomPan.Height(), 1, z );
//--------------------------------------------------------------------//
const C tau = Reflector( alpha11, a12 );
tau1.Set( 0, 0, tau );
const C alpha = alpha11.Get(0,0);
alpha11.Set(0,0,1);
Conjugate( aTopRow, aTopRowConj );
Gemv( NORMAL, C(1), ABottomPan, aTopRowConj, C(0), z );
Ger( -Conj(tau), z, aTopRowConj, ABottomPan );
alpha11.Set(0,0,alpha);
//--------------------------------------------------------------------//
SlidePartitionDown
( tT, t0,
tau1,
/**/ /****/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
PanelLQ
( DistMatrix<Complex<Real> >& A,
DistMatrix<Complex<Real>,MD,STAR>& t )
{
#ifndef RELEASE
PushCallStack("internal::PanelLQ");
if( A.Grid() != t.Grid() )
throw std::logic_error("{A,t} must be distributed over the same grid");
if( t.Height() != std::min(A.Height(),A.Width()) || t.Width() != 1 )
throw std::logic_error
("t must be a vector of height equal to the minimum dimension of A");
if( !t.AlignedWithDiagonal( A, 0 ) )
throw std::logic_error("t must be aligned with A's main diagonal");
#endif
typedef Complex<Real> C;
const Grid& g = A.Grid();
// Matrix views
DistMatrix<C>
ATL(g), ATR(g), A00(g), a01(g), A02(g), aTopRow(g), ABottomPan(g),
ABL(g), ABR(g), a10(g), alpha11(g), a12(g),
A20(g), a21(g), A22(g);
DistMatrix<C,MD,STAR>
tT(g), t0(g),
tB(g), tau1(g),
t2(g);
// Temporary distributions
DistMatrix<C> aTopRowConj(g);
DistMatrix<C,STAR,MR > aTopRowConj_STAR_MR(g);
DistMatrix<C,MC, STAR> z_MC_STAR(g);
PushBlocksizeStack( 1 );
PartitionDownLeftDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
PartitionDown
( t, tT,
tB, 0 );
while( ATL.Height() < A.Height() && ATL.Width() < A.Width() )
{
RepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ a01, A02,
/*************/ /**********************/
/**/ a10, /**/ alpha11, a12,
ABL, /**/ ABR, A20, /**/ a21, A22 );
RepartitionDown
( tT, t0,
/**/ /****/
tau1,
tB, t2 );
aTopRow.View1x2( alpha11, a12 );
ABottomPan.View1x2( a21, A22 );
aTopRowConj_STAR_MR.AlignWith( ABottomPan );
z_MC_STAR.AlignWith( ABottomPan );
Zeros( ABottomPan.Height(), 1, z_MC_STAR );
//--------------------------------------------------------------------//
const C tau = Reflector( alpha11, a12 );
tau1.Set( 0, 0, tau );
const bool myDiagonalEntry = ( g.Row() == alpha11.ColAlignment() &&
g.Col() == alpha11.RowAlignment() );
C alpha = 0;
if( myDiagonalEntry )
{
alpha = alpha11.GetLocal(0,0);
alpha11.SetLocal(0,0,1);
}
Conjugate( aTopRow, aTopRowConj );
aTopRowConj_STAR_MR = aTopRowConj;
Gemv
( NORMAL,
C(1), ABottomPan.LockedLocalMatrix(),
aTopRowConj_STAR_MR.LockedLocalMatrix(),
C(0), z_MC_STAR.LocalMatrix() );
z_MC_STAR.SumOverRow();
Ger
( -Conj(tau),
z_MC_STAR.LockedLocalMatrix(),
aTopRowConj_STAR_MR.LockedLocalMatrix(),
ABottomPan.LocalMatrix() );
if( myDiagonalEntry )
alpha11.SetLocal(0,0,alpha);
//--------------------------------------------------------------------//
aTopRowConj_STAR_MR.FreeAlignments();
z_MC_STAR.FreeAlignments();
SlidePartitionDown
( tT, t0,
tau1,
/**/ /****/
tB, t2 );
SlidePartitionDownDiagonal
( ATL, /**/ ATR, A00, a01, /**/ A02,
/**/ a10, alpha11, /**/ a12,
/*************/ /**********************/
ABL, /**/ ABR, A20, a21, /**/ A22 );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LocalTrmmAccumulateLLT
( Orientation orientation, UnitOrNonUnit diag, T alpha,
const DistMatrix<T>& L,
const DistMatrix<T,MC,STAR>& X_MC_STAR,
DistMatrix<T,MR,STAR>& Z_MR_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalTrmmAccumulateLLT");
if( L.Grid() != X_MC_STAR.Grid() ||
X_MC_STAR.Grid() != Z_MR_STAR.Grid() )
throw std::logic_error
("{L,X,Z} must be distributed over the same grid");
if( L.Height() != L.Width() ||
L.Height() != X_MC_STAR.Height() ||
L.Height() != Z_MR_STAR.Height() )
{
std::ostringstream msg;
msg << "Nonconformal LocalTrmmAccumulateLLT: " << "\n"
<< " L ~ " << L.Height() << " x " << L.Width() << "\n"
<< " X[MC,* ] ~ " << X_MC_STAR.Height() << " x "
<< X_MC_STAR.Width() << "\n"
<< " Z[MR,* ] ` " << Z_MR_STAR.Height() << " x "
<< Z_MR_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( X_MC_STAR.ColAlignment() != L.ColAlignment() ||
Z_MR_STAR.ColAlignment() != L.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = L.Grid();
// Matrix views
DistMatrix<T>
LTL(g), LTR(g), L00(g), L01(g), L02(g),
LBL(g), LBR(g), L10(g), L11(g), L12(g),
L20(g), L21(g), L22(g);
DistMatrix<T> D11(g);
DistMatrix<T,MC,STAR>
XT_MC_STAR(g), X0_MC_STAR(g),
XB_MC_STAR(g), X1_MC_STAR(g),
X2_MC_STAR(g);
DistMatrix<T,MR,STAR>
ZT_MR_STAR(g), Z0_MR_STAR(g),
ZB_MR_STAR(g), Z1_MR_STAR(g),
Z2_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
LockedPartitionDown
( X_MC_STAR, XT_MC_STAR,
XB_MC_STAR, 0 );
PartitionDown
( Z_MR_STAR, ZT_MR_STAR,
ZB_MR_STAR, 0 );
while( LTL.Height() < L.Height() )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
LockedRepartitionDown
( XT_MC_STAR, X0_MC_STAR,
/**********/ /**********/
X1_MC_STAR,
XB_MC_STAR, X2_MC_STAR );
RepartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
/**********/ /**********/
Z1_MR_STAR,
ZB_MR_STAR, Z2_MR_STAR );
D11.AlignWith( L11 );
//--------------------------------------------------------------------//
D11 = L11;
MakeTrapezoidal( LEFT, LOWER, 0, D11 );
if( diag == UNIT )
SetDiagonalToOne( D11 );
LocalGemm
( orientation, NORMAL, alpha, D11, X1_MC_STAR, T(1), Z1_MR_STAR );
LocalGemm
( orientation, NORMAL, alpha, L21, X2_MC_STAR, T(1), Z1_MR_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlideLockedPartitionDown
( XT_MC_STAR, X0_MC_STAR,
X1_MC_STAR,
/**********/ /**********/
XB_MC_STAR, X2_MC_STAR );
SlidePartitionDown
( ZT_MR_STAR, Z0_MR_STAR,
Z1_MR_STAR,
/**********/ /**********/
ZB_MR_STAR, Z2_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
LocalSymmetricAccumulateRU
( Orientation orientation, T alpha,
const DistMatrix<T,MC, MR >& A,
const DistMatrix<T,STAR,MC >& B_STAR_MC,
const DistMatrix<T,MR, STAR>& BTrans_MR_STAR,
DistMatrix<T,MC, STAR>& ZTrans_MC_STAR,
DistMatrix<T,MR, STAR>& ZTrans_MR_STAR )
{
#ifndef RELEASE
PushCallStack("internal::LocalSymmetricAccumulateRU");
if( A.Grid() != B_STAR_MC.Grid() ||
B_STAR_MC.Grid() != BTrans_MR_STAR.Grid() ||
BTrans_MR_STAR.Grid() != ZTrans_MC_STAR.Grid() ||
ZTrans_MC_STAR.Grid() != ZTrans_MR_STAR.Grid() )
throw std::logic_error
("{A,B,C} must be distributed over the same grid");
if( A.Height() != A.Width() ||
A.Height() != B_STAR_MC.Width() ||
A.Height() != BTrans_MR_STAR.Height() ||
A.Height() != ZTrans_MC_STAR.Height() ||
A.Height() != ZTrans_MR_STAR.Height() ||
B_STAR_MC.Height() != BTrans_MR_STAR.Width() ||
BTrans_MR_STAR.Width() != ZTrans_MC_STAR.Width() ||
ZTrans_MC_STAR.Width() != ZTrans_MR_STAR.Width() )
{
std::ostringstream msg;
msg << "Nonconformal LocalSymmetricAccumulateRU: \n"
<< " A ~ " << A.Height() << " x " << A.Width() << "\n"
<< " B[* ,MC] ~ " << B_STAR_MC.Height() << " x "
<< B_STAR_MC.Width() << "\n"
<< " B^H/T[MR,* ] ~ " << BTrans_MR_STAR.Height() << " x "
<< BTrans_MR_STAR.Width() << "\n"
<< " Z^H/T[MC,* ] ~ " << ZTrans_MC_STAR.Height() << " x "
<< ZTrans_MC_STAR.Width() << "\n"
<< " Z^H/T[MR,* ] ~ " << ZTrans_MR_STAR.Height() << " x "
<< ZTrans_MR_STAR.Width() << "\n";
throw std::logic_error( msg.str().c_str() );
}
if( B_STAR_MC.RowAlignment() != A.ColAlignment() ||
BTrans_MR_STAR.ColAlignment() != A.RowAlignment() ||
ZTrans_MC_STAR.ColAlignment() != A.ColAlignment() ||
ZTrans_MR_STAR.ColAlignment() != A.RowAlignment() )
throw std::logic_error("Partial matrix distributions are misaligned");
#endif
const Grid& g = A.Grid();
// Matrix views
DistMatrix<T>
ATL(g), ATR(g), A00(g), A01(g), A02(g),
ABL(g), ABR(g), A10(g), A11(g), A12(g),
A20(g), A21(g), A22(g);
DistMatrix<T> D11(g);
DistMatrix<T,STAR,MC>
BL_STAR_MC(g), BR_STAR_MC(g),
B0_STAR_MC(g), B1_STAR_MC(g), B2_STAR_MC(g);
DistMatrix<T,MR,STAR>
BTTrans_MR_STAR(g), B0Trans_MR_STAR(g),
BBTrans_MR_STAR(g), B1Trans_MR_STAR(g),
B2Trans_MR_STAR(g);
DistMatrix<T,MC,STAR>
ZTTrans_MC_STAR(g), Z0Trans_MC_STAR(g),
ZBTrans_MC_STAR(g), Z1Trans_MC_STAR(g),
Z2Trans_MC_STAR(g);
DistMatrix<T,MR,STAR>
ZBTrans_MR_STAR(g), Z0Trans_MR_STAR(g),
ZTTrans_MR_STAR(g), Z1Trans_MR_STAR(g),
Z2Trans_MR_STAR(g);
const int ratio = std::max( g.Height(), g.Width() );
PushBlocksizeStack( ratio*Blocksize() );
LockedPartitionDownDiagonal
( A, ATL, ATR,
ABL, ABR, 0 );
LockedPartitionRight( B_STAR_MC, BL_STAR_MC, BR_STAR_MC, 0 );
LockedPartitionDown
( BTrans_MR_STAR, BTTrans_MR_STAR,
BBTrans_MR_STAR, 0 );
PartitionDown
( ZTrans_MC_STAR, ZTTrans_MC_STAR,
ZBTrans_MC_STAR, 0 );
PartitionDown
( ZTrans_MR_STAR, ZTTrans_MR_STAR,
ZBTrans_MR_STAR, 0 );
while( ATL.Height() < A.Height() )
{
LockedRepartitionDownDiagonal
( ATL, /**/ ATR, A00, /**/ A01, A02,
/*************/ /******************/
/**/ A10, /**/ A11, A12,
ABL, /**/ ABR, A20, /**/ A21, A22 );
LockedRepartitionRight
( BL_STAR_MC, /**/ BR_STAR_MC,
B0_STAR_MC, /**/ B1_STAR_MC, B2_STAR_MC );
LockedRepartitionDown
( BTTrans_MR_STAR, B0Trans_MR_STAR,
/***************/ /***************/
B1Trans_MR_STAR,
BBTrans_MR_STAR, B2Trans_MR_STAR );
RepartitionDown
( ZTTrans_MC_STAR, Z0Trans_MC_STAR,
/***************/ /***************/
Z1Trans_MC_STAR,
ZBTrans_MC_STAR, Z2Trans_MC_STAR );
RepartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
/***************/ /***************/
Z1Trans_MR_STAR,
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
D11.AlignWith( A11 );
//--------------------------------------------------------------------//
D11 = A11;
MakeTriangular( UPPER, D11 );
LocalGemm
( orientation, orientation,
alpha, D11, B1_STAR_MC, T(1), Z1Trans_MR_STAR );
SetDiagonal( D11, T(0) );
LocalGemm
( NORMAL, NORMAL, alpha, D11, B1Trans_MR_STAR, T(1), Z1Trans_MC_STAR );
LocalGemm
( orientation, orientation,
alpha, A12, B1_STAR_MC, T(1), Z2Trans_MR_STAR );
LocalGemm
( NORMAL, NORMAL, alpha, A12, B2Trans_MR_STAR, T(1), Z1Trans_MC_STAR );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDownDiagonal
( ATL, /**/ ATR, A00, A01, /**/ A02,
/**/ A10, A11, /**/ A12,
/*************/ /******************/
ABL, /**/ ABR, A20, A21, /**/ A22 );
SlideLockedPartitionRight
( BL_STAR_MC, /**/ BR_STAR_MC,
B0_STAR_MC, B1_STAR_MC, /**/ B2_STAR_MC );
SlideLockedPartitionDown
( BTTrans_MR_STAR, B0Trans_MR_STAR,
B1Trans_MR_STAR,
/***************/ /***************/
BBTrans_MR_STAR, B2Trans_MR_STAR );
SlidePartitionDown
( ZTTrans_MC_STAR, Z0Trans_MC_STAR,
Z1Trans_MC_STAR,
/***************/ /***************/
ZBTrans_MC_STAR, Z2Trans_MC_STAR );
SlidePartitionDown
( ZTTrans_MR_STAR, Z0Trans_MR_STAR,
Z1Trans_MR_STAR,
/***************/ /***************/
ZBTrans_MR_STAR, Z2Trans_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}