inline void
TrsvUN( UnitOrNonUnit diag, const DistMatrix<F>& U, DistMatrix<F>& x )
{
#ifndef RELEASE
PushCallStack("internal::TrsvUN");
if( U.Grid() != x.Grid() )
throw std::logic_error("{U,x} must be distributed over the same grid");
if( U.Height() != U.Width() )
throw std::logic_error("U must be square");
if( x.Width() != 1 && x.Height() != 1 )
throw std::logic_error("x must be a vector");
const int xLength = ( x.Width() == 1 ? x.Height() : x.Width() );
if( U.Width() != xLength )
throw std::logic_error("Nonconformal TrsvUN");
#endif
const Grid& g = U.Grid();
if( x.Width() == 1 )
{
// Matrix views
DistMatrix<F> U01(g),
U11(g);
DistMatrix<F>
xT(g), x0(g),
xB(g), x1(g),
x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,MR, STAR> x1_MR_STAR(g);
DistMatrix<F,MC, STAR> z_MC_STAR(g);
// Views of z[MC,* ], which will store updates to x
DistMatrix<F,MC,STAR> z0_MC_STAR(g),
z1_MC_STAR(g);
z_MC_STAR.AlignWith( U );
Zeros( x.Height(), 1, z_MC_STAR );
// Start the algorithm
PartitionUp
( x, xT,
xB, 0 );
while( xT.Height() > 0 )
{
RepartitionUp
( xT, x0,
x1,
/**/ /**/
xB, x2 );
const int n0 = x0.Height();
const int n1 = x1.Height();
LockedView( U01, U, 0, n0, n0, n1 );
LockedView( U11, U, n0, n0, n1, n1 );
View( z0_MC_STAR, z_MC_STAR, 0, 0, n0, 1 );
View( z1_MC_STAR, z_MC_STAR, n0, 0, n1, 1 );
x1_MR_STAR.AlignWith( U01 );
//----------------------------------------------------------------//
if( x2.Height() != 0 )
x1.SumScatterUpdate( F(1), z1_MC_STAR );
x1_STAR_STAR = x1;
U11_STAR_STAR = U11;
Trsv
( UPPER, NORMAL, diag,
U11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_MR_STAR = x1_STAR_STAR;
Gemv
( NORMAL, F(-1),
U01.LockedLocalMatrix(),
x1_MR_STAR.LockedLocalMatrix(),
F(1), z0_MC_STAR.LocalMatrix() );
//----------------------------------------------------------------//
x1_MR_STAR.FreeAlignments();
SlidePartitionUp
( xT, x0,
/**/ /**/
x1,
xB, x2 );
}
}
else
{
// Matrix views
DistMatrix<F> U01(g),
U11(g);
DistMatrix<F>
xL(g), xR(g),
x0(g), x1(g), x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,STAR,MR > x1_STAR_MR(g);
DistMatrix<F,MC, MR > z1(g);
DistMatrix<F,MR, MC > z1_MR_MC(g);
DistMatrix<F,STAR,MC > z_STAR_MC(g);
// Views of z[* ,MC]
DistMatrix<F,STAR,MC> z0_STAR_MC(g),
z1_STAR_MC(g);
z_STAR_MC.AlignWith( U );
Zeros( 1, x.Width(), z_STAR_MC );
// Start the algorithm
PartitionLeft( x, xL, xR, 0 );
while( xL.Width() > 0 )
{
RepartitionLeft
( xL, /**/ xR,
x0, x1, /**/ x2 );
const int n0 = x0.Width();
const int n1 = x1.Width();
LockedView( U01, U, 0, n0, n0, n1 );
LockedView( U11, U, n0, n0, n1, n1 );
View( z0_STAR_MC, z_STAR_MC, 0, 0, 1, n0 );
View( z1_STAR_MC, z_STAR_MC, 0, n0, 1, n1 );
x1_STAR_MR.AlignWith( U01 );
z1.AlignWith( x1 );
//----------------------------------------------------------------//
if( x2.Width() != 0 )
{
z1_MR_MC.SumScatterFrom( z1_STAR_MC );
z1 = z1_MR_MC;
Axpy( F(1), z1, x1 );
}
x1_STAR_STAR = x1;
U11_STAR_STAR = U11;
Trsv
( UPPER, NORMAL, diag,
U11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_STAR_MR = x1_STAR_STAR;
Gemv
( NORMAL, F(-1),
U01.LockedLocalMatrix(),
x1_STAR_MR.LockedLocalMatrix(),
F(1), z0_STAR_MC.LocalMatrix() );
//----------------------------------------------------------------//
x1_STAR_MR.FreeAlignments();
z1.FreeAlignments();
SlidePartitionLeft
( xL, /**/ xR,
x0, /**/ x1, x2 );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
internal::TrsvLN
( UnitOrNonUnit diag,
const DistMatrix<F,MC,MR>& L,
DistMatrix<F,MC,MR>& x )
{
#ifndef RELEASE
PushCallStack("internal::TrsvLN");
if( L.Grid() != x.Grid() )
throw std::logic_error("{L,x} must be distributed over the same grid");
if( L.Height() != L.Width() )
throw std::logic_error("L must be square");
if( x.Width() != 1 && x.Height() != 1 )
throw std::logic_error("x must be a vector");
const int xLength = ( x.Width() == 1 ? x.Height() : x.Width() );
if( L.Width() != xLength )
throw std::logic_error("Nonconformal TrsvLN");
#endif
const Grid& g = L.Grid();
if( x.Width() == 1 )
{
// Matrix views
DistMatrix<F,MC,MR>
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<F,MC,MR>
xT(g), x0(g),
xB(g), x1(g),
x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,MR, STAR> x1_MR_STAR(g);
DistMatrix<F,MC, STAR> z2_MC_STAR(g);
// Start the algorithm
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionDown
( x, xT,
xB, 0 );
while( xB.Height() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionDown
( xT, x0,
/**/ /**/
x1,
xB, x2 );
x1_MR_STAR.AlignWith( L21 );
z2_MC_STAR.AlignWith( L21 );
z2_MC_STAR.ResizeTo( x2.Height(), 1 );
//----------------------------------------------------------------//
x1_STAR_STAR = x1;
L11_STAR_STAR = L11;
Trsv
( LOWER, NORMAL, diag,
L11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_MR_STAR = x1_STAR_STAR;
Gemv
( NORMAL, (F)-1,
L21.LockedLocalMatrix(),
x1_MR_STAR.LockedLocalMatrix(),
(F)0, z2_MC_STAR.LocalMatrix() );
x2.SumScatterUpdate( (F)1, z2_MC_STAR );
//----------------------------------------------------------------//
x1_MR_STAR.FreeAlignments();
z2_MC_STAR.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionDown
( xT, x0,
x1,
/**/ /**/
xB, x2 );
}
}
else
{
// Matrix views
DistMatrix<F,MC,MR>
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<F,MC,MR>
xL(g), xR(g),
x0(g), x1(g), x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> L11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,STAR,MR > x1_STAR_MR(g);
DistMatrix<F,STAR,MC > z2_STAR_MC(g);
DistMatrix<F,MR, MC > z2_MR_MC(g);
DistMatrix<F,MC, MR > z2(g);
// Start the algorithm
LockedPartitionDownDiagonal
( L, LTL, LTR,
LBL, LBR, 0 );
PartitionRight( x, xL, xR, 0 );
while( xR.Width() > 0 )
{
LockedRepartitionDownDiagonal
( LTL, /**/ LTR, L00, /**/ L01, L02,
/*************/ /******************/
/**/ L10, /**/ L11, L12,
LBL, /**/ LBR, L20, /**/ L21, L22 );
RepartitionRight
( xL, /**/ xR,
x0, /**/ x1, x2 );
x1_STAR_MR.AlignWith( L21 );
z2_STAR_MC.AlignWith( L21 );
z2.AlignWith( x2 );
z2_STAR_MC.ResizeTo( 1, x2.Width() );
//----------------------------------------------------------------//
x1_STAR_STAR = x1;
L11_STAR_STAR = L11;
Trsv
( LOWER, NORMAL, diag,
L11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_STAR_MR = x1_STAR_STAR;
Gemv
( NORMAL, (F)-1,
L21.LockedLocalMatrix(),
x1_STAR_MR.LockedLocalMatrix(),
(F)0, z2_STAR_MC.LocalMatrix() );
z2_MR_MC.SumScatterFrom( z2_STAR_MC );
z2 = z2_MR_MC;
Axpy( (F)1, z2, x2 );
//----------------------------------------------------------------//
x1_STAR_MR.FreeAlignments();
z2_STAR_MC.FreeAlignments();
z2.FreeAlignments();
SlideLockedPartitionDownDiagonal
( LTL, /**/ LTR, L00, L01, /**/ L02,
/**/ L10, L11, /**/ L12,
/*************/ /******************/
LBL, /**/ LBR, L20, L21, /**/ L22 );
SlidePartitionRight
( xL, /**/ xR,
x0, x1, /**/ x2 );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
internal::TrsvUN
( UnitOrNonUnit diag,
const DistMatrix<F,MC,MR>& U,
DistMatrix<F,MC,MR>& x )
{
#ifndef RELEASE
PushCallStack("internal::TrsvUN");
if( U.Grid() != x.Grid() )
throw std::logic_error("{U,x} must be distributed over the same grid");
if( U.Height() != U.Width() )
throw std::logic_error("U must be square");
if( x.Width() != 1 && x.Height() != 1 )
throw std::logic_error("x must be a vector");
const int xLength = ( x.Width() == 1 ? x.Height() : x.Width() );
if( U.Width() != xLength )
throw std::logic_error("Nonconformal TrsvUN");
#endif
const Grid& g = U.Grid();
if( x.Width() == 1 )
{
// Matrix views
DistMatrix<F,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<F,MC,MR>
xT(g), x0(g),
xB(g), x1(g),
x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,MR, STAR> x1_MR_STAR(g);
DistMatrix<F,MC, STAR> z0_MC_STAR(g);
// Start the algorithm
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionUp
( x, xT,
xB, 0 );
while( xT.Height() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionUp
( xT, x0,
x1,
/**/ /**/
xB, x2 );
x1_MR_STAR.AlignWith( U01 );
z0_MC_STAR.AlignWith( U01 );
z0_MC_STAR.ResizeTo( x0.Height(), 1 );
//----------------------------------------------------------------//
x1_STAR_STAR = x1;
U11_STAR_STAR = U11;
Trsv
( UPPER, NORMAL, diag,
U11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_MR_STAR = x1_STAR_STAR;
Gemv
( NORMAL, (F)-1,
U01.LockedLocalMatrix(),
x1_MR_STAR.LockedLocalMatrix(),
(F)0, z0_MC_STAR.LocalMatrix() );
x0.SumScatterUpdate( (F)1, z0_MC_STAR );
//----------------------------------------------------------------//
x1_MR_STAR.FreeAlignments();
z0_MC_STAR.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionUp
( xT, x0,
/**/ /**/
x1,
xB, x2 );
}
}
else
{
// Matrix views
DistMatrix<F,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<F,MC,MR>
xL(g), xR(g),
x0(g), x1(g), x2(g);
// Temporary distributions
DistMatrix<F,STAR,STAR> U11_STAR_STAR(g);
DistMatrix<F,STAR,STAR> x1_STAR_STAR(g);
DistMatrix<F,STAR,MR > x1_STAR_MR(g);
DistMatrix<F,STAR,MC > z0_STAR_MC(g);
DistMatrix<F,MR, MC > z0_MR_MC(g);
DistMatrix<F,MC, MR > z0(g);
// Start the algorithm
LockedPartitionUpDiagonal
( U, UTL, UTR,
UBL, UBR, 0 );
PartitionLeft( x, xL, xR, 0 );
while( xL.Width() > 0 )
{
LockedRepartitionUpDiagonal
( UTL, /**/ UTR, U00, U01, /**/ U02,
/**/ U10, U11, /**/ U12,
/*************/ /******************/
UBL, /**/ UBR, U20, U21, /**/ U22 );
RepartitionLeft
( xL, /**/ xR,
x0, x1, /**/ x2 );
x1_STAR_MR.AlignWith( U01 );
z0_STAR_MC.AlignWith( U01 );
z0.AlignWith( x0 );
z0_STAR_MC.ResizeTo( 1, x0.Width() );
//----------------------------------------------------------------//
x1_STAR_STAR = x1;
U11_STAR_STAR = U11;
Trsv
( UPPER, NORMAL, diag,
U11_STAR_STAR.LockedLocalMatrix(),
x1_STAR_STAR.LocalMatrix() );
x1 = x1_STAR_STAR;
x1_STAR_MR = x1_STAR_STAR;
Gemv
( NORMAL, (F)-1,
U01.LockedLocalMatrix(),
x1_STAR_MR.LockedLocalMatrix(),
(F)0, z0_STAR_MC.LocalMatrix() );
z0_MR_MC.SumScatterFrom( z0_STAR_MC );
z0 = z0_MR_MC;
Axpy( (F)1, z0, x0 );
//----------------------------------------------------------------//
x1_STAR_MR.FreeAlignments();
z0_STAR_MC.FreeAlignments();
z0.FreeAlignments();
SlideLockedPartitionUpDiagonal
( UTL, /**/ UTR, U00, /**/ U01, U02,
/*************/ /******************/
/**/ U10, /**/ U11, U12,
UBL, /**/ UBR, U20, /**/ U21, U22 );
SlidePartitionLeft
( xL, /**/ xR,
x0, /**/ x1, x2 );
}
}
#ifndef RELEASE
PopCallStack();
#endif
}
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 )
{
#ifndef RELEASE
PushCallStack("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() )
throw std::logic_error
("{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 )
throw std::logic_error("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";
throw std::logic_error( 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() )
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,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 = std::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;
MakeTrapezoidal( LEFT, UPPER, 0, D11 );
Gemv
( NORMAL,
alpha, D11.LockedLocalMatrix(),
x1_MR_STAR.LockedLocalMatrix(),
T(1), z1_MC_STAR.LocalMatrix() );
MakeTrapezoidal( LEFT, UPPER, 1, D11 );
Gemv
( TRANSPOSE,
alpha, D11.LockedLocalMatrix(),
x1_MC_STAR.LockedLocalMatrix(),
T(1), z1_MR_STAR.LocalMatrix() );
Gemv
( NORMAL,
alpha, A12.LockedLocalMatrix(),
x2_MR_STAR.LockedLocalMatrix(),
T(1), z1_MC_STAR.LocalMatrix() );
Gemv
( TRANSPOSE,
alpha, A12.LockedLocalMatrix(),
x1_MC_STAR.LockedLocalMatrix(),
T(1), z2_MR_STAR.LocalMatrix() );
//--------------------------------------------------------------------//
D11.FreeAlignments();
SlideLockedPartitionDown
( xT_MR_STAR, x0_MR_STAR,
x1_MR_STAR,
/**********/ /**********/
xB_MR_STAR, x2_MR_STAR );
}
PopBlocksizeStack();
#ifndef RELEASE
PopCallStack();
#endif
}
