Fix a couple issues introduced in the previous commit: * removed DirectAccessBit from Part * use a template specialization in inverseProduct() to transform a Part xpr to a Flagged xpr
diff --git a/Eigen/src/Core/InverseProduct.h b/Eigen/src/Core/InverseProduct.h index 87f426a..cfab542 100755 --- a/Eigen/src/Core/InverseProduct.h +++ b/Eigen/src/Core/InverseProduct.h
@@ -25,8 +25,12 @@ #ifndef EIGEN_INVERSEPRODUCT_H #define EIGEN_INVERSEPRODUCT_H +template<typename XprType> struct ei_is_part { enum {value=false}; }; +template<typename XprType, unsigned int Mode> struct ei_is_part<Part<XprType,Mode> > { enum {value=true}; }; + template<typename Lhs, typename Rhs, - int TriangularPart = (int(Lhs::Flags) & LowerTriangularBit) + int TriangularPart = ei_is_part<Lhs>::value ? -1 // this is to solve ambiguous specializations + : (int(Lhs::Flags) & LowerTriangularBit) ? Lower : (int(Lhs::Flags) & UpperTriangularBit) ? Upper @@ -35,6 +39,16 @@ > struct ei_trisolve_selector; +// transform a Part xpr to a Flagged xpr +template<typename Lhs, unsigned int LhsMode, typename Rhs, int TriangularPart, int StorageOrder> +struct ei_trisolve_selector<Part<Lhs,LhsMode>,Rhs,TriangularPart,StorageOrder> +{ + static void run(const Part<Lhs,LhsMode>& lhs, Rhs& other) + { + ei_trisolve_selector<Flagged<Lhs,LhsMode,0>,Rhs>::run(lhs._expression(), other); + } +}; + // forward substitution, row-major template<typename Lhs, typename Rhs> struct ei_trisolve_selector<Lhs,Rhs,Lower,RowMajor> @@ -102,12 +116,12 @@ int blockyEnd = (std::max(size-5,0)/4)*4; for(int i=0; i<blockyEnd;) { - int startBlock = i; - int endBlock = startBlock+4; - Matrix<Scalar,4,1> btmp; /* Let's process the 4x4 sub-matrix as usual. * btmp stores the diagonal coefficients used to update the remaining part of the result. */ + int startBlock = i; + int endBlock = startBlock+4; + Matrix<Scalar,4,1> btmp; for (;i<endBlock;++i) { if(!(Lhs::Flags & UnitDiagBit)) @@ -135,8 +149,10 @@ { if(!(Lhs::Flags & UnitDiagBit)) other.coeffRef(i,c) /= lhs.coeff(i,i); - // NOTE we cannot use lhs.col(i).end(size-i-1) because Part::coeffRef gets called by .col() to - // get the address of the start of the row + + /* NOTE we cannot use lhs.col(i).end(size-i-1) because Part::coeffRef gets called by .col() to + * get the address of the start of the row + */ other.col(c).end(size-i-1) -= other.coeffRef(i,c) * Block<Lhs,Dynamic,1>(lhs, i+1,i, size-i-1,1); } if(!(Lhs::Flags & UnitDiagBit)) @@ -146,6 +162,7 @@ }; // backward substitution, col-major +// see the previous specialization for details on the algorithm template<typename Lhs, typename Rhs> struct ei_trisolve_selector<Lhs,Rhs,Upper,ColMajor> {
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index dbc2511..1e9d0d9 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h
@@ -32,7 +32,7 @@ * This class is the base that is inherited by all matrix, vector, and expression * types. Most of the Eigen API is contained in this class. Other important classes for * the Eigen API are Matrix, Cwise, and PartialRedux. - * + * * Note that some methods are defined in the \ref Array module. * * \param Derived is the derived type, e.g. a matrix type, or an expression, etc. @@ -550,7 +550,7 @@ template<typename OtherDerived> typename ei_eval<Derived>::type cross(const MatrixBase<OtherDerived>& other) const; - typename ei_eval<Derived>::type perpendicular(void) const; + typename ei_eval<Derived>::type someOrthogonal(void) const; }; #endif // EIGEN_MATRIXBASE_H
diff --git a/Eigen/src/Core/Part.h b/Eigen/src/Core/Part.h index 1a7c7f8..871e274 100644 --- a/Eigen/src/Core/Part.h +++ b/Eigen/src/Core/Part.h
@@ -53,7 +53,7 @@ ColsAtCompileTime = MatrixType::ColsAtCompileTime, MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, - Flags = (_MatrixTypeNested::Flags & (HereditaryBits | DirectAccessBit) & (~(PacketAccessBit | LinearAccessBit))) | Mode, + Flags = (_MatrixTypeNested::Flags & (HereditaryBits) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode, CoeffReadCost = _MatrixTypeNested::CoeffReadCost }; }; @@ -109,6 +109,9 @@ return m_matrix.const_cast_derived().coeffRef(row, col); } + /** \internal */ + const MatrixType& _expression() const { return m_matrix; } + /** discard any writes to a row */ const Block<Part, 1, ColsAtCompileTime> row(int i) { return Base::row(i); } const Block<Part, 1, ColsAtCompileTime> row(int i) const { return Base::row(i); }
diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index bf491b7..9a03659 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h
@@ -378,7 +378,7 @@ } }; -// NOTE the 2 following specializations are because taking .col(0) on a vector is a bit slower +// NOTE the 3 following specializations are because taking .col(0) on a vector is a bit slower template<typename Lhs, typename Rhs, int RhsCols> struct ei_product_coeff_vectorized_dyn_selector<Lhs,Rhs,1,RhsCols> { @@ -404,6 +404,18 @@ }; template<typename Lhs, typename Rhs> +struct ei_product_coeff_vectorized_dyn_selector<Lhs,Rhs,1,1> +{ + inline static void run(int row, int /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res) + { + res = ei_dot_impl< + Lhs, + Rhs, + LinearVectorization, NoUnrolling>::run(lhs, rhs); + } +}; + +template<typename Lhs, typename Rhs> struct ei_product_coeff_impl<InnerVectorization, Dynamic, Lhs, Rhs> { inline static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res) @@ -518,7 +530,7 @@ }; // optimized cache friendly colmajor * vector path for matrix with direct access flag -// NOTE this path coul also be enabled for expressions if we add runtime align queries +// NOTE this path could also be enabled for expressions if we add runtime align queries template<typename ProductType, int LhsRows, int RhsOrder, int RhsAccess> struct ei_cache_friendly_product_selector<ProductType,LhsRows,ColMajor,HasDirectAccess,1,RhsOrder,RhsAccess> {
diff --git a/Eigen/src/Geometry/OrthoMethods.h b/Eigen/src/Geometry/OrthoMethods.h index 5955ce2..ea71c12 100644 --- a/Eigen/src/Geometry/OrthoMethods.h +++ b/Eigen/src/Geometry/OrthoMethods.h
@@ -23,8 +23,8 @@ // License and a copy of the GNU General Public License along with // Eigen. If not, see <http://www.gnu.org/licenses/>. -#ifndef EIGEN_CROSS_H -#define EIGEN_CROSS_H +#ifndef EIGEN_ORTHOMETHODS_H +#define EIGEN_ORTHOMETHODS_H /** \geometry_module * \returns the cross product of \c *this and \a other */ @@ -34,7 +34,7 @@ MatrixBase<Derived>::cross(const MatrixBase<OtherDerived>& other) const { EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Derived,3); - + // Note that there is no need for an expression here since the compiler // optimize such a small temporary very well (even within a complex expression) const typename ei_nested<Derived,2>::type lhs(derived()); @@ -107,4 +107,4 @@ return ei_perpendicular_selector<Derived>::run(derived()); } -#endif // EIGEN_CROSS_H +#endif // EIGEN_ORTHOMETHODS_H