Esempio n. 1
0
inline void
ApplyPackedReflectorsRLHF
( Conjugation conjugation, int offset, 
  const DistMatrix<Complex<R> >& H,
  const DistMatrix<Complex<R>,MD,STAR>& t,
        DistMatrix<Complex<R> >& A )
{
#ifndef RELEASE
    PushCallStack("internal::ApplyPackedReflectorsRLHF");
    if( H.Grid() != t.Grid() || t.Grid() != A.Grid() )
        throw std::logic_error
        ("{H,t,A} must be distributed over the same grid");
    if( offset > 0 || offset < -H.Width() )
        throw std::logic_error("Transforms out of bounds");
    if( H.Width() != A.Width() )
        throw std::logic_error
        ("Width of transforms must equal width of target matrix");
    if( t.Height() != H.DiagonalLength( offset ) )
        throw std::logic_error("t must be the same length as H's offset diag");
    if( !t.AlignedWithDiagonal( H, offset ) )
        throw std::logic_error("t must be aligned with H's 'offset' diagonal");
#endif
    typedef Complex<R> C;
    const Grid& g = H.Grid();

    DistMatrix<C>
        HTL(g), HTR(g),  H00(g), H01(g), H02(g),  HPan(g), HPanCopy(g),
        HBL(g), HBR(g),  H10(g), H11(g), H12(g),
                         H20(g), H21(g), H22(g);
    DistMatrix<C> ALeft(g);
    DistMatrix<C,MD,STAR>
        tT(g),  t0(g),
        tB(g),  t1(g),
                t2(g);

    DistMatrix<C,STAR,VR  > HPan_STAR_VR(g);
    DistMatrix<C,STAR,MR  > HPan_STAR_MR(g);
    DistMatrix<C,STAR,STAR> t1_STAR_STAR(g);
    DistMatrix<C,STAR,STAR> SInv_STAR_STAR(g);
    DistMatrix<C,STAR,MC  > ZAdj_STAR_MC(g);
    DistMatrix<C,STAR,VC  > ZAdj_STAR_VC(g);

    LockedPartitionDownDiagonal
    ( H, HTL, HTR,
         HBL, HBR, 0 );
    LockedPartitionDown
    ( t, tT,
         tB, 0 );
    while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
    {
        LockedRepartitionDownDiagonal
        ( HTL, /**/ HTR,  H00, /**/ H01, H02,
         /*************/ /******************/
               /**/       H10, /**/ H11, H12,
          HBL, /**/ HBR,  H20, /**/ H21, H22 );

        const int HPanWidth = H10.Width() + H11.Width();
        const int HPanOffset = 
            std::min( H11.Height(), std::max(-offset-H00.Height(),0) );
        const int HPanHeight = H11.Height()-HPanOffset;
        HPan.LockedView( H, H00.Height()+HPanOffset, 0, HPanHeight, HPanWidth );

        LockedRepartitionDown
        ( tT,  t0,
         /**/ /**/
               t1,
          tB,  t2, HPanHeight );

        ALeft.View( A, 0, 0, A.Height(), HPanWidth );

        HPan_STAR_MR.AlignWith( ALeft );
        ZAdj_STAR_MC.AlignWith( ALeft );
        ZAdj_STAR_VC.AlignWith( ALeft );
        Zeros( HPan.Height(), ALeft.Height(), ZAdj_STAR_MC );
        Zeros( HPan.Height(), HPan.Height(), SInv_STAR_STAR );
        //--------------------------------------------------------------------//
        HPanCopy = HPan;
        MakeTrapezoidal( RIGHT, LOWER, offset, HPanCopy );
        SetDiagonalToOne( RIGHT, offset, HPanCopy );

        HPan_STAR_VR = HPanCopy;
        Herk
        ( UPPER, NORMAL,
          C(1), HPan_STAR_VR.LockedLocalMatrix(),
          C(0), SInv_STAR_STAR.LocalMatrix() );
        SInv_STAR_STAR.SumOverGrid();
        t1_STAR_STAR = t1;
        FixDiagonal( conjugation, t1_STAR_STAR, SInv_STAR_STAR );

        HPan_STAR_MR = HPan_STAR_VR;
        LocalGemm
        ( NORMAL, ADJOINT,
          C(1), HPan_STAR_MR, ALeft, C(0), ZAdj_STAR_MC );
        ZAdj_STAR_VC.SumScatterFrom( ZAdj_STAR_MC );

        LocalTrsm
        ( LEFT, UPPER, ADJOINT, NON_UNIT,
          C(1), SInv_STAR_STAR, ZAdj_STAR_VC );

        ZAdj_STAR_MC = ZAdj_STAR_VC;
        LocalGemm
        ( ADJOINT, NORMAL,
          C(-1), ZAdj_STAR_MC, HPan_STAR_MR, C(1), ALeft );
        //--------------------------------------------------------------------//
        HPan_STAR_MR.FreeAlignments();
        ZAdj_STAR_MC.FreeAlignments();
        ZAdj_STAR_VC.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( HTL, /**/ HTR,  H00, H01, /**/ H02,
               /**/       H10, H11, /**/ H12,
         /*************/ /******************/
          HBL, /**/ HBR,  H20, H21, /**/ H22 );

        SlideLockedPartitionDown
        ( tT,  t0,
               t1,
         /**/ /**/
          tB,  t2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
inline void
internal::ApplyPackedReflectorsLUVF
( Conjugation conjugation, int offset, 
  const DistMatrix<Complex<R>,MC,MR  >& H,
  const DistMatrix<Complex<R>,MD,STAR>& t,
        DistMatrix<Complex<R>,MC,MR  >& A )
{
#ifndef RELEASE
    PushCallStack("internal::ApplyPackedReflectorsLUVF");
    if( H.Grid() != t.Grid() || t.Grid() != A.Grid() )
        throw std::logic_error
        ("{H,t,A} must be distributed over the same grid");
    if( offset > H.Height() )
        throw std::logic_error("Transforms cannot extend above matrix");
    if( offset < 0 )
        throw std::logic_error("Transforms cannot extend below matrix");
    if( H.Width() != A.Height() )
        throw std::logic_error
        ("Width of transforms must equal height of target matrix");
    if( t.Height() != H.DiagonalLength( offset ) )
        throw std::logic_error("t must be the same length as H's offset diag");
    if( !t.AlignedWithDiagonal( H, offset ) )
        throw std::logic_error("t must be aligned with H's 'offset' diagonal");
#endif
    typedef Complex<R> C;
    const Grid& g = H.Grid();

    // Matrix views    
    DistMatrix<C,MC,MR>
        HTL(g), HTR(g),  H00(g), H01(g), H02(g),  HPan(g),
        HBL(g), HBR(g),  H10(g), H11(g), H12(g),
                         H20(g), H21(g), H22(g);
    DistMatrix<C,MC,MR>
        AT(g),  A0(g),  ATop(g),
        AB(g),  A1(g),
                A2(g);
    DistMatrix<C,MD,STAR>
        tT(g),  t0(g),
        tB(g),  t1(g),
                t2(g);

    DistMatrix<C,MC,  MR  > HPanCopy(g);
    DistMatrix<C,VC,  STAR> HPan_VC_STAR(g);
    DistMatrix<C,MC,  STAR> HPan_MC_STAR(g);
    DistMatrix<C,STAR,STAR> t1_STAR_STAR(g);
    DistMatrix<C,STAR,STAR> SInv_STAR_STAR(g);
    DistMatrix<C,STAR,MR  > Z_STAR_MR(g);
    DistMatrix<C,STAR,VR  > Z_STAR_VR(g);

    LockedPartitionDownDiagonal
    ( H, HTL, HTR,
         HBL, HBR, 0 );
    LockedPartitionDown
    ( t, tT,
         tB, 0 );
    PartitionDown
    ( A, AT,
         AB, 0 );
    while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
    {
        LockedRepartitionDownDiagonal
        ( HTL, /**/ HTR,  H00, /**/ H01, H02,
         /*************/ /******************/
               /**/       H10, /**/ H11, H12,
          HBL, /**/ HBR,  H20, /**/ H21, H22 );

        const int HPanHeight = H01.Height() + H11.Height();
        const int HPanOffset = 
            std::min( H11.Width(), std::max(offset-H00.Width(),0) );
        const int HPanWidth = H11.Width()-HPanOffset;
        HPan.LockedView( H, 0, H00.Width()+HPanOffset, HPanHeight, HPanWidth );

        LockedRepartitionDown
        ( tT,  t0,
         /**/ /**/
               t1,
          tB,  t2, HPanWidth );

        RepartitionDown
        ( AT,  A0,
         /**/ /**/
               A1,
          AB,  A2 );

        ATop.View2x1( A0,
                      A1 );

        HPan_MC_STAR.AlignWith( ATop );
        Z_STAR_MR.AlignWith( ATop );
        Z_STAR_VR.AlignWith( ATop );
        Z_STAR_MR.ResizeTo( HPan.Width(), ATop.Width() );
        SInv_STAR_STAR.ResizeTo( HPan.Width(), HPan.Width() );
        Zero( SInv_STAR_STAR );
        //--------------------------------------------------------------------//
        HPanCopy = HPan;
        MakeTrapezoidal( RIGHT, UPPER, offset, HPanCopy );
        SetDiagonalToOne( RIGHT, offset, HPanCopy );
        HPan_VC_STAR = HPanCopy;
        Herk
        ( UPPER, ADJOINT, 
          (C)1, HPan_VC_STAR.LockedLocalMatrix(),
          (C)0, SInv_STAR_STAR.LocalMatrix() ); 
        SInv_STAR_STAR.SumOverGrid();
        t1_STAR_STAR = t1;
        FixDiagonal( conjugation, t1_STAR_STAR, SInv_STAR_STAR );

        HPan_MC_STAR = HPanCopy;
        internal::LocalGemm
        ( ADJOINT, NORMAL, (C)1, HPan_MC_STAR, ATop, (C)0, Z_STAR_MR );
        Z_STAR_VR.SumScatterFrom( Z_STAR_MR );
        
        internal::LocalTrsm
        ( LEFT, UPPER, ADJOINT, NON_UNIT, (C)1, SInv_STAR_STAR, Z_STAR_VR );

        Z_STAR_MR = Z_STAR_VR;
        internal::LocalGemm
        ( NORMAL, NORMAL, (C)-1, HPan_MC_STAR, Z_STAR_MR, (C)1, ATop );
        //--------------------------------------------------------------------//
        HPan_MC_STAR.FreeAlignments();
        Z_STAR_MR.FreeAlignments();
        Z_STAR_VR.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( HTL, /**/ HTR,  H00, H01, /**/ H02,
               /**/       H10, H11, /**/ H12,
         /*************/ /******************/
          HBL, /**/ HBR,  H20, H21, /**/ H22 );

        SlideLockedPartitionDown
        ( tT,  t0,
               t1,
         /**/ /**/
          tB,  t2 );

        SlidePartitionDown
        ( AT,  A0,
               A1,
         /**/ /**/
          AB,  A2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
inline void
internal::ApplyPackedReflectorsLUVF
( int offset, 
  const DistMatrix<R,MC,MR>& H,
        DistMatrix<R,MC,MR>& A )
{
#ifndef RELEASE
    PushCallStack("internal::ApplyPackedReflectorsLUVF");
    if( H.Grid() != A.Grid() )
        throw std::logic_error("{H,A} must be distributed over the same grid");
    if( offset > H.Height() )
        throw std::logic_error("Transforms cannot extend above matrix");
    if( offset < 0 )
        throw std::logic_error("Transforms cannot extend below matrix");
    if( H.Width() != A.Height() )
        throw std::logic_error
        ("Width of transforms must equal height of target matrix");
#endif
    const Grid& g = H.Grid();

    // Matrix views    
    DistMatrix<R,MC,MR>
        HTL(g), HTR(g),  H00(g), H01(g), H02(g),  HPan(g),
        HBL(g), HBR(g),  H10(g), H11(g), H12(g),
                         H20(g), H21(g), H22(g);
    DistMatrix<R,MC,MR>
        AT(g),  A0(g),  ATop(g),
        AB(g),  A1(g),
                A2(g);

    DistMatrix<R,MC,  MR  > HPanCopy(g);
    DistMatrix<R,VC,  STAR> HPan_VC_STAR(g);
    DistMatrix<R,MC,  STAR> HPan_MC_STAR(g);
    DistMatrix<R,STAR,STAR> SInv_STAR_STAR(g);
    DistMatrix<R,STAR,MR  > Z_STAR_MR(g);
    DistMatrix<R,STAR,VR  > Z_STAR_VR(g);

    LockedPartitionDownDiagonal
    ( H, HTL, HTR,
         HBL, HBR, 0 );
    PartitionDown
    ( A, AT,
         AB, 0 );
    while( HTL.Height() < H.Height() && HTL.Width() < H.Width() )
    {
        LockedRepartitionDownDiagonal
        ( HTL, /**/ HTR,  H00, /**/ H01, H02,
         /*************/ /******************/
               /**/       H10, /**/ H11, H12,
          HBL, /**/ HBR,  H20, /**/ H21, H22 );

        const int HPanHeight = H01.Height() + H11.Height();
        const int HPanOffset = 
            std::min( H11.Width(), std::max(offset-H00.Width(),0) );
        const int HPanWidth = H11.Width()-HPanOffset;
        HPan.LockedView( H, 0, H00.Width()+HPanOffset, HPanHeight, HPanWidth );

        RepartitionDown
        ( AT,  A0,
         /**/ /**/
               A1,
          AB,  A2 );

        ATop.View2x1( A0, 
                      A1 );

        HPan_MC_STAR.AlignWith( ATop );
        Z_STAR_MR.AlignWith( ATop );
        Z_STAR_VR.AlignWith( ATop );
        Z_STAR_MR.ResizeTo( HPan.Width(), ATop.Width() );
        SInv_STAR_STAR.ResizeTo( HPan.Width(), HPan.Width() );
        Zero( SInv_STAR_STAR );
        //--------------------------------------------------------------------//
        HPanCopy = HPan;
        MakeTrapezoidal( RIGHT, UPPER, offset, HPanCopy );
        SetDiagonalToOne( RIGHT, offset, HPanCopy );
        HPan_VC_STAR = HPanCopy;
        Syrk
        ( UPPER, TRANSPOSE, 
          (R)1, HPan_VC_STAR.LockedLocalMatrix(),
          (R)0, SInv_STAR_STAR.LocalMatrix() ); 
        SInv_STAR_STAR.SumOverGrid();
        HalveMainDiagonal( SInv_STAR_STAR );

        HPan_MC_STAR = HPanCopy;
        internal::LocalGemm
        ( TRANSPOSE, NORMAL, 
          (R)1, HPan_MC_STAR, ATop, (R)0, Z_STAR_MR );
        Z_STAR_VR.SumScatterFrom( Z_STAR_MR );
        
        internal::LocalTrsm
        ( LEFT, UPPER, TRANSPOSE, NON_UNIT, 
          (R)1, SInv_STAR_STAR, Z_STAR_VR );

        Z_STAR_MR = Z_STAR_VR;
        internal::LocalGemm
        ( NORMAL, NORMAL, 
          (R)-1, HPan_MC_STAR, Z_STAR_MR, (R)1, ATop );
        //--------------------------------------------------------------------//
        HPan_MC_STAR.FreeAlignments();
        Z_STAR_MR.FreeAlignments();
        Z_STAR_VR.FreeAlignments();

        SlideLockedPartitionDownDiagonal
        ( HTL, /**/ HTR,  H00, H01, /**/ H02,
               /**/       H10, H11, /**/ H12,
         /*************/ /******************/
          HBL, /**/ HBR,  H20, H21, /**/ H22 );

        SlidePartitionDown
        ( AT,  A0,
               A1,
         /**/ /**/
          AB,  A2 );
    }
#ifndef RELEASE
    PopCallStack();
#endif
}
Esempio n. 4
0
inline void
RUVB
( int offset, 
  const DistMatrix<R>& H,
        DistMatrix<R>& A )
{
#ifndef RELEASE
    CallStackEntry entry("apply_packed_reflectors::RUVB");
    if( H.Grid() != A.Grid() )
        throw std::logic_error("{H,A} must be distributed over the same grid");
    if( offset < 0 || offset > H.Height() )
        throw std::logic_error("Transforms out of bounds");
    if( H.Height() != A.Width() )
        throw std::logic_error
        ("Height of transforms must equal width of target matrix");
#endif
    const Grid& g = H.Grid();

    DistMatrix<R>
        HTL(g), HTR(g),  H00(g), H01(g), H02(g),  HPan(g), HPanCopy(g),
        HBL(g), HBR(g),  H10(g), H11(g), H12(g),
                         H20(g), H21(g), H22(g);
    DistMatrix<R> ALeft(g);

    DistMatrix<R,VC,  STAR> HPan_VC_STAR(g);
    DistMatrix<R,MR,  STAR> HPan_MR_STAR(g);
    DistMatrix<R,STAR,STAR> SInv_STAR_STAR(g);
    DistMatrix<R,STAR,MC  > ZTrans_STAR_MC(g);
    DistMatrix<R,STAR,VC  > ZTrans_STAR_VC(g);

    LockedPartitionUpDiagonal
    ( H, HTL, HTR,
         HBL, HBR, 0 );
    while( HBR.Height() < H.Height() && HBR.Width() < H.Width() )
    {
        LockedRepartitionUpDiagonal
        ( HTL, /**/ HTR,  H00, H01, /**/ H02,
               /**/       H10, H11, /**/ H12,
         /*************/ /******************/
          HBL, /**/ HBR,  H20, H21, /**/ H22 );

        const int HPanHeight = H01.Height() + H11.Height();
        const int HPanOffset = 
            std::min( H11.Width(), std::max(offset-H00.Width(),0) );
        const int HPanWidth = H11.Width()-HPanOffset;
        LockedView( HPan, H, 0, H00.Width()+HPanOffset, HPanHeight, HPanWidth );

        View( ALeft, A, 0, 0, A.Height(), HPanHeight );

        HPan_MR_STAR.AlignWith( ALeft );
        ZTrans_STAR_MC.AlignWith( ALeft );
        ZTrans_STAR_VC.AlignWith( ALeft );
        //--------------------------------------------------------------------//
        HPanCopy = HPan;
        MakeTrapezoidal( RIGHT, UPPER, offset, HPanCopy );
        SetDiagonal( RIGHT, offset, HPanCopy, R(1) );

        HPan_VC_STAR = HPanCopy;
        Zeros( SInv_STAR_STAR, HPan.Width(), HPan.Width() );
        Syrk
        ( LOWER, TRANSPOSE, 
          R(1), HPan_VC_STAR.LockedMatrix(),
          R(0), SInv_STAR_STAR.Matrix() );     
        SInv_STAR_STAR.SumOverGrid();
        HalveMainDiagonal( SInv_STAR_STAR );

        HPan_MR_STAR = HPan_VC_STAR;
        LocalGemm
        ( TRANSPOSE, TRANSPOSE, R(1), HPan_MR_STAR, ALeft, ZTrans_STAR_MC );
        ZTrans_STAR_VC.SumScatterFrom( ZTrans_STAR_MC );
        
        LocalTrsm
        ( LEFT, LOWER, TRANSPOSE, NON_UNIT, 
          R(1), SInv_STAR_STAR, ZTrans_STAR_VC );

        ZTrans_STAR_MC = ZTrans_STAR_VC;
        LocalGemm
        ( TRANSPOSE, TRANSPOSE, 
          R(-1), ZTrans_STAR_MC, HPan_MR_STAR, R(1), ALeft );
        //--------------------------------------------------------------------//
        HPan_MR_STAR.FreeAlignments();
        ZTrans_STAR_MC.FreeAlignments();
        ZTrans_STAR_VC.FreeAlignments();

        SlideLockedPartitionUpDiagonal
        ( HTL, /**/ HTR,  H00, /**/ H01, H02,
         /*************/ /******************/
               /**/       H10, /**/ H11, H12,
          HBL, /**/ HBR,  H20, /**/ H21, H22 );
    }
}