inline void ComposePivots ( const DistMatrix<Int,VC,STAR>& p, std::vector<Int>& image, std::vector<Int>& preimage ) { #ifndef RELEASE CallStackEntry entry("ComposePivots"); #endif DistMatrix<Int,STAR,STAR> p_STAR_STAR( p ); ComposePivots( p_STAR_STAR, image, preimage ); }
inline void ApplyColumnPivots( DistMatrix<F>& A, const DistMatrix<int,STAR,STAR>& p ) { #ifndef RELEASE PushCallStack("ApplyColumnPivots"); #endif std::vector<int> image, preimage; ComposePivots( p, image, preimage ); ApplyColumnPivots( A, image, preimage ); #ifndef RELEASE PopCallStack(); #endif }
inline void LU( DistMatrix<F>& A, DistMatrix<int,VC,STAR>& p ) { #ifndef RELEASE CallStackEntry entry("LU"); if( A.Grid() != p.Grid() ) throw std::logic_error("{A,p} must be distributed over the same grid"); if( p.Viewing() && (std::min(A.Height(),A.Width()) != p.Height() || p.Width() != 1) ) throw std::logic_error ("p must be a vector of the same height as the min dimension of A."); #endif const Grid& g = A.Grid(); if( !p.Viewing() ) p.ResizeTo( std::min(A.Height(),A.Width()), 1 ); // Matrix views DistMatrix<F> ATL(g), ATR(g), A00(g), A01(g), A02(g), AB(g), ABL(g), ABR(g), A10(g), A11(g), A12(g), A20(g), A21(g), A22(g); DistMatrix<int,VC,STAR> pT(g), p0(g), pB(g), p1(g), p2(g); // Temporary distributions DistMatrix<F, STAR,STAR> A11_STAR_STAR(g); DistMatrix<F, MC, STAR> A21_MC_STAR(g); DistMatrix<F, STAR,VR > A12_STAR_VR(g); DistMatrix<F, STAR,MR > A12_STAR_MR(g); DistMatrix<int,STAR,STAR> p1_STAR_STAR(g); // Pivot composition std::vector<int> image, preimage; // Start the algorithm PartitionDownDiagonal ( A, ATL, ATR, ABL, ABR, 0 ); PartitionDown ( p, pT, pB, 0 ); while( ATL.Height() < A.Height() && ATL.Width() < A.Width() ) { RepartitionDownDiagonal ( ATL, /**/ ATR, A00, /**/ A01, A02, /*************/ /******************/ /**/ A10, /**/ A11, A12, ABL, /**/ ABR, A20, /**/ A21, A22 ); RepartitionDown ( pT, p0, /**/ /**/ p1, pB, p2 ); View1x2( AB, ABL, ABR ); const int pivotOffset = A01.Height(); A12_STAR_VR.AlignWith( A22 ); A12_STAR_MR.AlignWith( A22 ); A21_MC_STAR.AlignWith( A22 ); A11_STAR_STAR.ResizeTo( A11.Height(), A11.Width() ); p1_STAR_STAR.ResizeTo( p1.Height(), 1 ); //--------------------------------------------------------------------// A21_MC_STAR = A21; A11_STAR_STAR = A11; lu::Panel( A11_STAR_STAR, A21_MC_STAR, p1_STAR_STAR, pivotOffset ); ComposePivots( p1_STAR_STAR, pivotOffset, image, preimage ); ApplyRowPivots( AB, image, preimage ); // Perhaps we should give up perfectly distributing this operation since // it's total contribution is only O(n^2) A12_STAR_VR = A12; LocalTrsm ( LEFT, LOWER, NORMAL, UNIT, F(1), A11_STAR_STAR, A12_STAR_VR ); A12_STAR_MR = A12_STAR_VR; LocalGemm( NORMAL, NORMAL, F(-1), A21_MC_STAR, A12_STAR_MR, F(1), A22 ); A11 = A11_STAR_STAR; A12 = A12_STAR_MR; A21 = A21_MC_STAR; p1 = p1_STAR_STAR; //--------------------------------------------------------------------// A12_STAR_VR.FreeAlignments(); A12_STAR_MR.FreeAlignments(); A21_MC_STAR.FreeAlignments(); SlidePartitionDownDiagonal ( ATL, /**/ ATR, A00, A01, /**/ A02, /**/ A10, A11, /**/ A12, /*************/ /******************/ ABL, /**/ ABR, A20, A21, /**/ A22 ); SlidePartitionDown ( pT, p0, p1, /**/ /**/ pB, p2 ); } }
inline void LU( Matrix<F>& A, Matrix<int>& p ) { #ifndef RELEASE CallStackEntry entry("LU"); if( p.Viewing() && (p.Height() != std::min(A.Height(),A.Width()) || p.Width() != 1) ) throw std::logic_error ("p must be a vector of the same height as the min dimension of A"); #endif if( !p.Viewing() ) p.ResizeTo( std::min(A.Height(),A.Width()), 1 ); // Matrix views Matrix<F> ATL, ATR, A00, A01, A02, ABRL, ABRR, ABL, ABR, A10, A11, A12, A20, A21, A22; Matrix<int> pT, p0, pB, p1, p2; // Pivot composition std::vector<int> image, preimage; // Start the algorithm PartitionDownDiagonal ( A, ATL, ATR, ABL, ABR, 0 ); PartitionDown ( p, pT, pB, 0 ); while( ATL.Height() < A.Height() && ATL.Width() < A.Width() ) { RepartitionDownDiagonal ( ATL, /**/ ATR, A00, /**/ A01, A02, /*************/ /******************/ /**/ A10, /**/ A11, A12, ABL, /**/ ABR, A20, /**/ A21, A22 ); RepartitionDown ( pT, p0, /**/ /**/ p1, pB, p2 ); PartitionRight( ABR, ABRL, ABRR, A11.Width() ); const int pivotOffset = A01.Height(); //--------------------------------------------------------------------// lu::Panel( ABRL, p1, pivotOffset ); ComposePivots( p1, pivotOffset, image, preimage ); ApplyRowPivots( ABL, image, preimage ); ApplyRowPivots( ABRR, image, preimage ); Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 ); Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 ); //--------------------------------------------------------------------// SlidePartitionDownDiagonal ( ATL, /**/ ATR, A00, A01, /**/ A02, /**/ A10, A11, /**/ A12, /*************/ /******************/ ABL, /**/ ABR, A20, A21, /**/ A22 ); SlidePartitionDown ( pT, p0, p1, /**/ /**/ pB, p2 ); } }
inline void LU( Matrix<F>& A, Matrix<Int>& p ) { #ifndef RELEASE CallStackEntry entry("LU"); #endif p.ResizeTo( Min(A.Height(),A.Width()), 1 ); // Matrix views Matrix<F> ATL, ATR, A00, A01, A02, ABRL, ABRR, ABL, ABR, A10, A11, A12, A20, A21, A22; Matrix<Int> pT, p0, pB, p1, p2; // Pivot composition std::vector<Int> image, preimage; // Start the algorithm PartitionDownDiagonal ( A, ATL, ATR, ABL, ABR, 0 ); PartitionDown ( p, pT, pB, 0 ); while( ATL.Height() < A.Height() && ATL.Width() < A.Width() ) { RepartitionDownDiagonal ( ATL, /**/ ATR, A00, /**/ A01, A02, /*************/ /******************/ /**/ A10, /**/ A11, A12, ABL, /**/ ABR, A20, /**/ A21, A22 ); RepartitionDown ( pT, p0, /**/ /**/ p1, pB, p2 ); PartitionRight( ABR, ABRL, ABRR, A11.Width() ); const Int pivotOffset = A01.Height(); //--------------------------------------------------------------------// lu::Panel( ABRL, p1, pivotOffset ); ComposePivots( p1, pivotOffset, image, preimage ); ApplyRowPivots( ABL, image, preimage ); ApplyRowPivots( ABRR, image, preimage ); Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 ); Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 ); //--------------------------------------------------------------------// SlidePartitionDownDiagonal ( ATL, /**/ ATR, A00, A01, /**/ A02, /**/ A10, A11, /**/ A12, /*************/ /******************/ ABL, /**/ ABR, A20, A21, /**/ A22 ); SlidePartitionDown ( pT, p0, p1, /**/ /**/ pB, p2 ); } }
inline void ComposePivots ( const DistMatrix<Int,STAR,STAR>& p, std::vector<Int>& image, std::vector<Int>& preimage ) { ComposePivots( p.LockedMatrix(), image, preimage ); }
inline void RowEchelon( DistMatrix<F>& A, DistMatrix<F>& B ) { #ifndef RELEASE CallStackEntry entry("RowEchelon"); if( A.Grid() != B.Grid() ) LogicError("{A,B} must be distributed over the same grid"); if( A.Height() != B.Height() ) LogicError("A and B must be the same height"); #endif const Grid& g = A.Grid(); // Matrix views DistMatrix<F> ATL(g), ATR(g), A00(g), A01(g), A02(g), APan(g), ABL(g), ABR(g), A10(g), A11(g), A12(g), A20(g), A21(g), A22(g); DistMatrix<F> BT(g), B0(g), BB(g), B1(g), B2(g); // Temporary distributions DistMatrix<F,STAR,STAR> A11_STAR_STAR(g); DistMatrix<F,STAR,VR > A12_STAR_VR(g); DistMatrix<F,STAR,MR > A12_STAR_MR(g); DistMatrix<F,MC, STAR> A21_MC_STAR(g); DistMatrix<F,STAR,VR > B1_STAR_VR(g); DistMatrix<F,STAR,MR > B1_STAR_MR(g); DistMatrix<Int,STAR,STAR> p1_STAR_STAR(g); // In case B's columns are not aligned with A's const bool BAligned = ( B.ColShift() == A.ColShift() ); DistMatrix<F,MC,STAR> A21_MC_STAR_B(g); // Pivot composition std::vector<Int> image, preimage; // Start the algorithm PartitionDownDiagonal ( A, ATL, ATR, ABL, ABR, 0 ); PartitionDown ( B, BT, BB, 0 ); while( ATL.Height() < A.Height() && ATL.Width() < A.Width() ) { RepartitionDownDiagonal ( ATL, /**/ ATR, A00, /**/ A01, A02, /*************/ /******************/ /**/ A10, /**/ A11, A12, ABL, /**/ ABR, A20, /**/ A21, A22 ); RepartitionDown ( BT, B0, /**/ /**/ B1, BB, B2 ); View2x1 ( APan, A12, A22 ); A12_STAR_VR.AlignWith( A22 ); A12_STAR_MR.AlignWith( A22 ); A21_MC_STAR.AlignWith( A22 ); B1_STAR_VR.AlignWith( B1 ); B1_STAR_MR.AlignWith( B1 ); if( ! BAligned ) A21_MC_STAR_B.AlignWith( B2 ); p1_STAR_STAR.ResizeTo( A11.Height(), 1 ); //--------------------------------------------------------------------// A11_STAR_STAR = A11; A21_MC_STAR = A21; lu::Panel( A11_STAR_STAR, A21_MC_STAR, p1_STAR_STAR, A00.Height() ); ComposePivots( p1_STAR_STAR, A00.Height(), image, preimage ); ApplyRowPivots( APan, image, preimage ); ApplyRowPivots( BB, image, preimage ); A12_STAR_VR = A12; B1_STAR_VR = B1; LocalTrsm ( LEFT, LOWER, NORMAL, UNIT, F(1), A11_STAR_STAR, A12_STAR_VR ); LocalTrsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11_STAR_STAR, B1_STAR_VR ); A12_STAR_MR = A12_STAR_VR; B1_STAR_MR = B1_STAR_VR; LocalGemm( NORMAL, NORMAL, F(-1), A21_MC_STAR, A12_STAR_MR, F(1), A22 ); if( BAligned ) { LocalGemm ( NORMAL, NORMAL, F(-1), A21_MC_STAR, B1_STAR_MR, F(1), B2 ); } else { A21_MC_STAR_B = A21_MC_STAR; LocalGemm ( NORMAL, NORMAL, F(-1), A21_MC_STAR_B, B1_STAR_MR, F(1), B2 ); } A11 = A11_STAR_STAR; A12 = A12_STAR_MR; B1 = B1_STAR_MR; //--------------------------------------------------------------------// SlidePartitionDownDiagonal ( ATL, /**/ ATR, A00, A01, /**/ A02, /**/ A10, A11, /**/ A12, /*************/ /******************/ ABL, /**/ ABR, A20, A21, /**/ A22 ); SlidePartitionDown ( BT, B0, B1, /**/ /**/ BB, B2 ); } }
inline void RowEchelon( Matrix<F>& A, Matrix<F>& B ) { #ifndef RELEASE CallStackEntry entry("RowEchelon"); if( A.Height() != B.Height() ) LogicError("A and B must be the same height"); #endif // Matrix views Matrix<F> ATL, ATR, A00, A01, A02, APan, ABL, ABR, A10, A11, A12, A20, A21, A22; Matrix<F> BT, B0, BB, B1, B2; Matrix<Int> p1; // Pivot composition std::vector<Int> image, preimage; // Start the algorithm PartitionDownDiagonal ( A, ATL, ATR, ABL, ABR, 0 ); PartitionDown ( B, BT, BB, 0 ); while( ATL.Height() < A.Height() && ATL.Width() < A.Width() ) { RepartitionDownDiagonal ( ATL, /**/ ATR, A00, /**/ A01, A02, /*************/ /******************/ /**/ A10, /**/ A11, A12, ABL, /**/ ABR, A20, /**/ A21, A22 ); RepartitionDown ( BT, B0, /**/ /**/ B1, BB, B2 ); View2x1 ( APan, A12, A22 ); //--------------------------------------------------------------------// lu::Panel( APan, p1, A00.Height() ); ComposePivots( p1, A00.Height(), image, preimage ); ApplyRowPivots( BB, image, preimage ); Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, A12 ); Trsm( LEFT, LOWER, NORMAL, UNIT, F(1), A11, B1 ); Gemm( NORMAL, NORMAL, F(-1), A21, A12, F(1), A22 ); Gemm( NORMAL, NORMAL, F(-1), A21, B1, F(1), B2 ); //--------------------------------------------------------------------// SlidePartitionDownDiagonal ( ATL, /**/ ATR, A00, A01, /**/ A02, /**/ A10, A11, /**/ A12, /*************/ /******************/ ABL, /**/ ABR, A20, A21, /**/ A22 ); SlidePartitionDown ( BT, B0, B1, /**/ /**/ BB, B2 ); } }