tomoto 0.2.2 → 0.3.0

data/vendor/eigen/Eigen/src/LU/Determinant.h

@@ -15,6 +15,7 @@ namespace Eigen {
 namespace internal {
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 inline const typename Derived::Scalar bruteforce_det3_helper
 (const MatrixBase<Derived>& matrix, int a, int b, int c)
 {
@@ -22,14 +23,6 @@ inline const typename Derived::Scalar bruteforce_det3_helper
          * (matrix.coeff(1,b) * matrix.coeff(2,c) - matrix.coeff(1,c) * matrix.coeff(2,b));
 }
 
-template<typename Derived>
-const typename Derived::Scalar bruteforce_det4_helper
-(const MatrixBase<Derived>& matrix, int j, int k, int m, int n)
-{
-  return (matrix.coeff(j,0) * matrix.coeff(k,1) - matrix.coeff(k,0) * matrix.coeff(j,1))
-       * (matrix.coeff(m,2) * matrix.coeff(n,3) - matrix.coeff(n,2) * matrix.coeff(m,3));
-}
-
 template<typename Derived,
          int DeterminantType = Derived::RowsAtCompileTime
 > struct determinant_impl
@@ -44,7 +37,8 @@ template<typename Derived,
 
 template<typename Derived> struct determinant_impl<Derived, 1>
 {
-  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  static inline EIGEN_DEVICE_FUNC
+  typename traits<Derived>::Scalar run(const Derived& m)
   {
     return m.coeff(0,0);
   }
@@ -52,7 +46,8 @@ template<typename Derived> struct determinant_impl<Derived, 1>
 
 template<typename Derived> struct determinant_impl<Derived, 2>
 {
-  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  static inline EIGEN_DEVICE_FUNC
+  typename traits<Derived>::Scalar run(const Derived& m)
   {
     return m.coeff(0,0) * m.coeff(1,1) - m.coeff(1,0) * m.coeff(0,1);
   }
@@ -60,7 +55,8 @@ template<typename Derived> struct determinant_impl<Derived, 2>
 
 template<typename Derived> struct determinant_impl<Derived, 3>
 {
-  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  static inline EIGEN_DEVICE_FUNC
+  typename traits<Derived>::Scalar run(const Derived& m)
   {
     return bruteforce_det3_helper(m,0,1,2)
           - bruteforce_det3_helper(m,1,0,2)
@@ -70,15 +66,34 @@ template<typename Derived> struct determinant_impl<Derived, 3>
 
 template<typename Derived> struct determinant_impl<Derived, 4>
 {
-  static typename traits<Derived>::Scalar run(const Derived& m)
+  typedef typename traits<Derived>::Scalar Scalar;
+  static EIGEN_DEVICE_FUNC
+  Scalar run(const Derived& m)
+  {
+    Scalar d2_01 = det2(m, 0, 1);
+    Scalar d2_02 = det2(m, 0, 2);
+    Scalar d2_03 = det2(m, 0, 3);
+    Scalar d2_12 = det2(m, 1, 2);
+    Scalar d2_13 = det2(m, 1, 3);
+    Scalar d2_23 = det2(m, 2, 3);
+    Scalar d3_0 = det3(m, 1,d2_23, 2,d2_13, 3,d2_12);
+    Scalar d3_1 = det3(m, 0,d2_23, 2,d2_03, 3,d2_02);
+    Scalar d3_2 = det3(m, 0,d2_13, 1,d2_03, 3,d2_01);
+    Scalar d3_3 = det3(m, 0,d2_12, 1,d2_02, 2,d2_01);
+    return internal::pmadd(-m(0,3),d3_0, m(1,3)*d3_1) +
+           internal::pmadd(-m(2,3),d3_2, m(3,3)*d3_3);
+  }
+protected:
+  static EIGEN_DEVICE_FUNC
+  Scalar det2(const Derived& m, Index i0, Index i1)
+  {
+    return m(i0,0) * m(i1,1) - m(i1,0) * m(i0,1);
+  }
+
+  static EIGEN_DEVICE_FUNC
+  Scalar det3(const Derived& m, Index i0, const Scalar& d0, Index i1, const Scalar& d1, Index i2, const Scalar& d2)
   {
-    // trick by Martin Costabel to compute 4x4 det with only 30 muls
-    return bruteforce_det4_helper(m,0,1,2,3)
-          - bruteforce_det4_helper(m,0,2,1,3)
-          + bruteforce_det4_helper(m,0,3,1,2)
-          + bruteforce_det4_helper(m,1,2,0,3)
-          - bruteforce_det4_helper(m,1,3,0,2)
-          + bruteforce_det4_helper(m,2,3,0,1);
+    return internal::pmadd(m(i0,2), d0, internal::pmadd(-m(i1,2), d1, m(i2,2)*d2));
   }
 };
 
@@ -89,6 +104,7 @@ template<typename Derived> struct determinant_impl<Derived, 4>
   * \returns the determinant of this matrix
   */
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 inline typename internal::traits<Derived>::Scalar MatrixBase<Derived>::determinant() const
 {
   eigen_assert(rows() == cols());

data/vendor/eigen/Eigen/src/LU/FullPivLU.h

@@ -18,6 +18,7 @@ template<typename _MatrixType> struct traits<FullPivLU<_MatrixType> >
 {
   typedef MatrixXpr XprKind;
   typedef SolverStorage StorageKind;
+  typedef int StorageIndex;
   enum { Flags = 0 };
 };
 
@@ -48,12 +49,12 @@ template<typename _MatrixType> struct traits<FullPivLU<_MatrixType> >
   * The data of the LU decomposition can be directly accessed through the methods matrixLU(),
   * permutationP(), permutationQ().
   *
-  * As an exemple, here is how the original matrix can be retrieved:
+  * As an example, here is how the original matrix can be retrieved:
   * \include class_FullPivLU.cpp
   * Output: \verbinclude class_FullPivLU.out
   *
   * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
-  * 
+  *
   * \sa MatrixBase::fullPivLu(), MatrixBase::determinant(), MatrixBase::inverse()
   */
 template<typename _MatrixType> class FullPivLU
@@ -62,9 +63,9 @@ template<typename _MatrixType> class FullPivLU
   public:
     typedef _MatrixType MatrixType;
     typedef SolverBase<FullPivLU> Base;
+    friend class SolverBase<FullPivLU>;
 
     EIGEN_GENERIC_PUBLIC_INTERFACE(FullPivLU)
-    // FIXME StorageIndex defined in EIGEN_GENERIC_PUBLIC_INTERFACE should be int
     enum {
       MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
       MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
@@ -218,6 +219,7 @@ template<typename _MatrixType> class FullPivLU
       return internal::image_retval<FullPivLU>(*this, originalMatrix);
     }
 
+    #ifdef EIGEN_PARSED_BY_DOXYGEN
     /** \return a solution x to the equation Ax=b, where A is the matrix of which
       * *this is the LU decomposition.
       *
@@ -237,14 +239,10 @@ template<typename _MatrixType> class FullPivLU
       *
       * \sa TriangularView::solve(), kernel(), inverse()
       */
-    // FIXME this is a copy-paste of the base-class member to add the isInitialized assertion.
     template<typename Rhs>
     inline const Solve<FullPivLU, Rhs>
-    solve(const MatrixBase<Rhs>& b) const
-    {
-      eigen_assert(m_isInitialized && "LU is not initialized.");
-      return Solve<FullPivLU, Rhs>(*this, b.derived());
-    }
+    solve(const MatrixBase<Rhs>& b) const;
+    #endif
 
     /** \returns an estimate of the reciprocal condition number of the matrix of which \c *this is
         the LU decomposition.
@@ -320,7 +318,7 @@ template<typename _MatrixType> class FullPivLU
       return m_usePrescribedThreshold ? m_prescribedThreshold
       // this formula comes from experimenting (see "LU precision tuning" thread on the list)
       // and turns out to be identical to Higham's formula used already in LDLt.
-                                      : NumTraits<Scalar>::epsilon() * m_lu.diagonalSize();
+          : NumTraits<Scalar>::epsilon() * RealScalar(m_lu.diagonalSize());
     }
 
     /** \returns the rank of the matrix of which *this is the LU decomposition.
@@ -406,16 +404,16 @@ template<typename _MatrixType> class FullPivLU
 
     MatrixType reconstructedMatrix() const;
 
-    EIGEN_DEVICE_FUNC inline Index rows() const { return m_lu.rows(); }
-    EIGEN_DEVICE_FUNC inline Index cols() const { return m_lu.cols(); }
+    EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+    inline Index rows() const EIGEN_NOEXCEPT { return m_lu.rows(); }
+    EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR
+    inline Index cols() const EIGEN_NOEXCEPT { return m_lu.cols(); }
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl(const RhsType &rhs, DstType &dst) const;
 
     template<bool Conjugate, typename RhsType, typename DstType>
-    EIGEN_DEVICE_FUNC
     void _solve_impl_transposed(const RhsType &rhs, DstType &dst) const;
     #endif
 
@@ -531,8 +529,8 @@ void FullPivLU<MatrixType>::computeInPlace()
       m_nonzero_pivots = k;
       for(Index i = k; i < size; ++i)
       {
-        m_rowsTranspositions.coeffRef(i) = i;
-        m_colsTranspositions.coeffRef(i) = i;
+        m_rowsTranspositions.coeffRef(i) = internal::convert_index<StorageIndex>(i);
+        m_colsTranspositions.coeffRef(i) = internal::convert_index<StorageIndex>(i);
       }
       break;
     }
@@ -543,8 +541,8 @@ void FullPivLU<MatrixType>::computeInPlace()
     // Now that we've found the pivot, we need to apply the row/col swaps to
     // bring it to the location (k,k).
 
-    m_rowsTranspositions.coeffRef(k) = row_of_biggest_in_corner;
-    m_colsTranspositions.coeffRef(k) = col_of_biggest_in_corner;
+    m_rowsTranspositions.coeffRef(k) = internal::convert_index<StorageIndex>(row_of_biggest_in_corner);
+    m_colsTranspositions.coeffRef(k) = internal::convert_index<StorageIndex>(col_of_biggest_in_corner);
     if(k != row_of_biggest_in_corner) {
       m_lu.row(k).swap(m_lu.row(row_of_biggest_in_corner));
       ++number_of_transpositions;
@@ -757,7 +755,6 @@ void FullPivLU<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const
   const Index rows = this->rows(),
               cols = this->cols(),
               nonzero_pivots = this->rank();
-  eigen_assert(rhs.rows() == rows);
   const Index smalldim = (std::min)(rows, cols);
 
   if(nonzero_pivots == 0)
@@ -807,7 +804,6 @@ void FullPivLU<_MatrixType>::_solve_impl_transposed(const RhsType &rhs, DstType
 
   const Index rows = this->rows(), cols = this->cols(),
     nonzero_pivots = this->rank();
-   eigen_assert(rhs.rows() == cols);
   const Index smalldim = (std::min)(rows, cols);
 
   if(nonzero_pivots == 0)
@@ -821,29 +817,19 @@ void FullPivLU<_MatrixType>::_solve_impl_transposed(const RhsType &rhs, DstType
   // Step 1
   c = permutationQ().inverse() * rhs;
 
-  if (Conjugate) {
-    // Step 2
-    m_lu.topLeftCorner(nonzero_pivots, nonzero_pivots)
-        .template triangularView<Upper>()
-        .adjoint()
-        .solveInPlace(c.topRows(nonzero_pivots));
-    // Step 3
-    m_lu.topLeftCorner(smalldim, smalldim)
-        .template triangularView<UnitLower>()
-        .adjoint()
-        .solveInPlace(c.topRows(smalldim));
-  } else {
-    // Step 2
-    m_lu.topLeftCorner(nonzero_pivots, nonzero_pivots)
-        .template triangularView<Upper>()
-        .transpose()
-        .solveInPlace(c.topRows(nonzero_pivots));
-    // Step 3
-    m_lu.topLeftCorner(smalldim, smalldim)
-        .template triangularView<UnitLower>()
-        .transpose()
-        .solveInPlace(c.topRows(smalldim));
-  }
+  // Step 2
+  m_lu.topLeftCorner(nonzero_pivots, nonzero_pivots)
+      .template triangularView<Upper>()
+      .transpose()
+      .template conjugateIf<Conjugate>()
+      .solveInPlace(c.topRows(nonzero_pivots));
+
+  // Step 3
+  m_lu.topLeftCorner(smalldim, smalldim)
+      .template triangularView<UnitLower>()
+      .transpose()
+      .template conjugateIf<Conjugate>()
+      .solveInPlace(c.topRows(smalldim));
 
   // Step 4
   PermutationPType invp = permutationP().inverse().eval();

data/vendor/eigen/Eigen/src/LU/InverseImpl.h

@@ -77,10 +77,11 @@ inline void compute_inverse_size2_helper(
     const MatrixType& matrix, const typename ResultType::Scalar& invdet,
     ResultType& result)
 {
+  typename ResultType::Scalar temp = matrix.coeff(0,0);
   result.coeffRef(0,0) =  matrix.coeff(1,1) * invdet;
   result.coeffRef(1,0) = -matrix.coeff(1,0) * invdet;
   result.coeffRef(0,1) = -matrix.coeff(0,1) * invdet;
-  result.coeffRef(1,1) =  matrix.coeff(0,0) * invdet;
+  result.coeffRef(1,1) =  temp * invdet;
 }
 
 template<typename MatrixType, typename ResultType>
@@ -143,13 +144,18 @@ inline void compute_inverse_size3_helper(
     const Matrix<typename ResultType::Scalar,3,1>& cofactors_col0,
     ResultType& result)
 {
-  result.row(0) = cofactors_col0 * invdet;
-  result.coeffRef(1,0) =  cofactor_3x3<MatrixType,0,1>(matrix) * invdet;
-  result.coeffRef(1,1) =  cofactor_3x3<MatrixType,1,1>(matrix) * invdet;
+  // Compute cofactors in a way that avoids aliasing issues.
+  typedef typename ResultType::Scalar Scalar;
+  const Scalar c01 = cofactor_3x3<MatrixType,0,1>(matrix) * invdet;
+  const Scalar c11 = cofactor_3x3<MatrixType,1,1>(matrix) * invdet;
+  const Scalar c02 = cofactor_3x3<MatrixType,0,2>(matrix) * invdet;
   result.coeffRef(1,2) =  cofactor_3x3<MatrixType,2,1>(matrix) * invdet;
-  result.coeffRef(2,0) =  cofactor_3x3<MatrixType,0,2>(matrix) * invdet;
   result.coeffRef(2,1) =  cofactor_3x3<MatrixType,1,2>(matrix) * invdet;
   result.coeffRef(2,2) =  cofactor_3x3<MatrixType,2,2>(matrix) * invdet;
+  result.coeffRef(1,0) =  c01;
+  result.coeffRef(1,1) =  c11;
+  result.coeffRef(2,0) =  c02;  
+  result.row(0) = cofactors_col0 * invdet;
 }
 
 template<typename MatrixType, typename ResultType>
@@ -181,14 +187,13 @@ struct compute_inverse_and_det_with_check<MatrixType, ResultType, 3>
     bool& invertible
   )
   {
-    using std::abs;
     typedef typename ResultType::Scalar Scalar;
     Matrix<Scalar,3,1> cofactors_col0;
     cofactors_col0.coeffRef(0) =  cofactor_3x3<MatrixType,0,0>(matrix);
     cofactors_col0.coeffRef(1) =  cofactor_3x3<MatrixType,1,0>(matrix);
     cofactors_col0.coeffRef(2) =  cofactor_3x3<MatrixType,2,0>(matrix);
     determinant = (cofactors_col0.cwiseProduct(matrix.col(0))).sum();
-    invertible = abs(determinant) > absDeterminantThreshold;
+    invertible = Eigen::numext::abs(determinant) > absDeterminantThreshold;
     if(!invertible) return;
     const Scalar invdet = Scalar(1) / determinant;
     compute_inverse_size3_helper(matrix, invdet, cofactors_col0, inverse);
@@ -273,7 +278,13 @@ struct compute_inverse_and_det_with_check<MatrixType, ResultType, 4>
     using std::abs;
     determinant = matrix.determinant();
     invertible = abs(determinant) > absDeterminantThreshold;
-    if(invertible) compute_inverse<MatrixType, ResultType>::run(matrix, inverse);
+    if(invertible && extract_data(matrix) != extract_data(inverse)) {
+      compute_inverse<MatrixType, ResultType>::run(matrix, inverse);
+    }
+    else if(invertible) {
+      MatrixType matrix_t = matrix;
+      compute_inverse<MatrixType, ResultType>::run(matrix_t, inverse);
+    }
   }
 };
 
@@ -290,6 +301,7 @@ template<typename DstXprType, typename XprType>
 struct Assignment<DstXprType, Inverse<XprType>, internal::assign_op<typename DstXprType::Scalar,typename XprType::Scalar>, Dense2Dense>
 {
   typedef Inverse<XprType> SrcXprType;
+  EIGEN_DEVICE_FUNC
   static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename XprType::Scalar> &)
   {
     Index dstRows = src.rows();
@@ -332,6 +344,7 @@ struct Assignment<DstXprType, Inverse<XprType>, internal::assign_op<typename Dst
   * \sa computeInverseAndDetWithCheck()
   */
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 inline const Inverse<Derived> MatrixBase<Derived>::inverse() const
 {
   EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsInteger,THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES)
@@ -345,6 +358,8 @@ inline const Inverse<Derived> MatrixBase<Derived>::inverse() const
   *
   * This is only for fixed-size square matrices of size up to 4x4.
   *
+  * Notice that it will trigger a copy of input matrix when trying to do the inverse in place.
+  *
   * \param inverse Reference to the matrix in which to store the inverse.
   * \param determinant Reference to the variable in which to store the determinant.
   * \param invertible Reference to the bool variable in which to store whether the matrix is invertible.
@@ -385,6 +400,8 @@ inline void MatrixBase<Derived>::computeInverseAndDetWithCheck(
   *
   * This is only for fixed-size square matrices of size up to 4x4.
   *
+  * Notice that it will trigger a copy of input matrix when trying to do the inverse in place.
+  *
   * \param inverse Reference to the matrix in which to store the inverse.
   * \param invertible Reference to the bool variable in which to store whether the matrix is invertible.
   * \param absDeterminantThreshold Optional parameter controlling the invertibility check.

data/vendor/eigen/Eigen/src/LU/PartialPivLU.h

@@ -19,6 +19,7 @@ template<typename _MatrixType> struct traits<PartialPivLU<_MatrixType> >
 {
   typedef MatrixXpr XprKind;
   typedef SolverStorage StorageKind;
+  typedef int StorageIndex;
   typedef traits<_MatrixType> BaseTraits;
   enum {
     Flags = BaseTraits::Flags & RowMajorBit,
@@ -69,7 +70,7 @@ struct enable_if_ref<Ref<T>,Derived> {
   * The data of the LU decomposition can be directly accessed through the methods matrixLU(), permutationP().
   *
   * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
-  * 
+  *
   * \sa MatrixBase::partialPivLu(), MatrixBase::determinant(), MatrixBase::inverse(), MatrixBase::computeInverse(), class FullPivLU
   */
 template<typename _MatrixType> class PartialPivLU
@@ -79,8 +80,9 @@ template<typename _MatrixType> class PartialPivLU
 
     typedef _MatrixType MatrixType;
     typedef SolverBase<PartialPivLU> Base;
+    friend class SolverBase<PartialPivLU>;
+
     EIGEN_GENERIC_PUBLIC_INTERFACE(PartialPivLU)
-    // FIXME StorageIndex defined in EIGEN_GENERIC_PUBLIC_INTERFACE should be int
     enum {
       MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
       MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
@@ -152,6 +154,7 @@ template<typename _MatrixType> class PartialPivLU
       return m_p;
     }
 
+    #ifdef EIGEN_PARSED_BY_DOXYGEN
     /** This method returns the solution x to the equation Ax=b, where A is the matrix of which
       * *this is the LU decomposition.
       *
@@ -169,14 +172,10 @@ template<typename _MatrixType> class PartialPivLU
       *
       * \sa TriangularView::solve(), inverse(), computeInverse()
       */
-    // FIXME this is a copy-paste of the base-class member to add the isInitialized assertion.
     template<typename Rhs>
     inline const Solve<PartialPivLU, Rhs>
-    solve(const MatrixBase<Rhs>& b) const
-    {
-      eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
-      return Solve<PartialPivLU, Rhs>(*this, b.derived());
-    }
+    solve(const MatrixBase<Rhs>& b) const;
+    #endif
 
     /** \returns an estimate of the reciprocal condition number of the matrix of which \c *this is
         the LU decomposition.
@@ -217,8 +216,8 @@ template<typename _MatrixType> class PartialPivLU
 
     MatrixType reconstructedMatrix() const;
 
-    inline Index rows() const { return m_lu.rows(); }
-    inline Index cols() const { return m_lu.cols(); }
+    EIGEN_CONSTEXPR inline Index rows() const EIGEN_NOEXCEPT { return m_lu.rows(); }
+    EIGEN_CONSTEXPR inline Index cols() const EIGEN_NOEXCEPT { return m_lu.cols(); }
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename RhsType, typename DstType>
@@ -231,8 +230,6 @@ template<typename _MatrixType> class PartialPivLU
       * Step 3: replace c by the solution x to Ux = c.
       */
 
-      eigen_assert(rhs.rows() == m_lu.rows());
-
       // Step 1
       dst = permutationP() * rhs;
 
@@ -246,26 +243,21 @@ template<typename _MatrixType> class PartialPivLU
     template<bool Conjugate, typename RhsType, typename DstType>
     EIGEN_DEVICE_FUNC
     void _solve_impl_transposed(const RhsType &rhs, DstType &dst) const {
-     /* The decomposition PA = LU can be rewritten as A = P^{-1} L U.
+     /* The decomposition PA = LU can be rewritten as A^T = U^T L^T P.
       * So we proceed as follows:
-      * Step 1: compute c = Pb.
-      * Step 2: replace c by the solution x to Lx = c.
-      * Step 3: replace c by the solution x to Ux = c.
+      * Step 1: compute c as the solution to L^T c = b
+      * Step 2: replace c by the solution x to U^T x = c.
+      * Step 3: update  c = P^-1 c.
       */
 
       eigen_assert(rhs.rows() == m_lu.cols());
 
-      if (Conjugate) {
-        // Step 1
-        dst = m_lu.template triangularView<Upper>().adjoint().solve(rhs);
-        // Step 2
-        m_lu.template triangularView<UnitLower>().adjoint().solveInPlace(dst);
-      } else {
-        // Step 1
-        dst = m_lu.template triangularView<Upper>().transpose().solve(rhs);
-        // Step 2
-        m_lu.template triangularView<UnitLower>().transpose().solveInPlace(dst);
-      }
+      // Step 1
+      dst = m_lu.template triangularView<Upper>().transpose()
+                .template conjugateIf<Conjugate>().solve(rhs);
+      // Step 2
+      m_lu.template triangularView<UnitLower>().transpose()
+          .template conjugateIf<Conjugate>().solveInPlace(dst);
       // Step 3
       dst = permutationP().transpose() * dst;
     }
@@ -339,17 +331,18 @@ PartialPivLU<MatrixType>::PartialPivLU(EigenBase<InputType>& matrix)
 namespace internal {
 
 /** \internal This is the blocked version of fullpivlu_unblocked() */
-template<typename Scalar, int StorageOrder, typename PivIndex>
+template<typename Scalar, int StorageOrder, typename PivIndex, int SizeAtCompileTime=Dynamic>
 struct partial_lu_impl
 {
-  // FIXME add a stride to Map, so that the following mapping becomes easier,
-  // another option would be to create an expression being able to automatically
-  // warp any Map, Matrix, and Block expressions as a unique type, but since that's exactly
-  // a Map + stride, why not adding a stride to Map, and convenient ctors from a Matrix,
-  // and Block.
-  typedef Map<Matrix<Scalar, Dynamic, Dynamic, StorageOrder> > MapLU;
-  typedef Block<MapLU, Dynamic, Dynamic> MatrixType;
-  typedef Block<MatrixType,Dynamic,Dynamic> BlockType;
+  static const int UnBlockedBound = 16;
+  static const bool UnBlockedAtCompileTime = SizeAtCompileTime!=Dynamic && SizeAtCompileTime<=UnBlockedBound;
+  static const int ActualSizeAtCompileTime = UnBlockedAtCompileTime ? SizeAtCompileTime : Dynamic;
+  // Remaining rows and columns at compile-time:
+  static const int RRows = SizeAtCompileTime==2 ? 1 : Dynamic;
+  static const int RCols = SizeAtCompileTime==2 ? 1 : Dynamic;
+  typedef Matrix<Scalar, ActualSizeAtCompileTime, ActualSizeAtCompileTime, StorageOrder> MatrixType;
+  typedef Ref<MatrixType> MatrixTypeRef;
+  typedef Ref<Matrix<Scalar, Dynamic, Dynamic, StorageOrder> > BlockType;
   typedef typename MatrixType::RealScalar RealScalar;
 
   /** \internal performs the LU decomposition in-place of the matrix \a lu
@@ -362,19 +355,22 @@ struct partial_lu_impl
     *
     * \returns The index of the first pivot which is exactly zero if any, or a negative number otherwise.
     */
-  static Index unblocked_lu(MatrixType& lu, PivIndex* row_transpositions, PivIndex& nb_transpositions)
+  static Index unblocked_lu(MatrixTypeRef& lu, PivIndex* row_transpositions, PivIndex& nb_transpositions)
   {
     typedef scalar_score_coeff_op<Scalar> Scoring;
     typedef typename Scoring::result_type Score;
     const Index rows = lu.rows();
     const Index cols = lu.cols();
     const Index size = (std::min)(rows,cols);
+    // For small compile-time matrices it is worth processing the last row separately:
+    //  speedup: +100% for 2x2, +10% for others.
+    const Index endk = UnBlockedAtCompileTime ? size-1 : size;
     nb_transpositions = 0;
     Index first_zero_pivot = -1;
-    for(Index k = 0; k < size; ++k)
+    for(Index k = 0; k < endk; ++k)
     {
-      Index rrows = rows-k-1;
-      Index rcols = cols-k-1;
+      int rrows = internal::convert_index<int>(rows-k-1);
+      int rcols = internal::convert_index<int>(cols-k-1);
 
       Index row_of_biggest_in_col;
       Score biggest_in_corner
@@ -391,9 +387,7 @@ struct partial_lu_impl
           ++nb_transpositions;
         }
 
-        // FIXME shall we introduce a safe quotient expression in cas 1/lu.coeff(k,k)
-        // overflow but not the actual quotient?
-        lu.col(k).tail(rrows) /= lu.coeff(k,k);
+        lu.col(k).tail(fix<RRows>(rrows)) /= lu.coeff(k,k);
       }
       else if(first_zero_pivot==-1)
       {
@@ -403,8 +397,18 @@ struct partial_lu_impl
       }
 
       if(k<rows-1)
-        lu.bottomRightCorner(rrows,rcols).noalias() -= lu.col(k).tail(rrows) * lu.row(k).tail(rcols);
+        lu.bottomRightCorner(fix<RRows>(rrows),fix<RCols>(rcols)).noalias() -= lu.col(k).tail(fix<RRows>(rrows)) * lu.row(k).tail(fix<RCols>(rcols));
+    }
+
+    // special handling of the last entry
+    if(UnBlockedAtCompileTime)
+    {
+      Index k = endk;
+      row_transpositions[k] = PivIndex(k);
+      if (Scoring()(lu(k, k)) == Score(0) && first_zero_pivot == -1)
+        first_zero_pivot = k;
     }
+
     return first_zero_pivot;
   }
 
@@ -420,18 +424,17 @@ struct partial_lu_impl
     * \returns The index of the first pivot which is exactly zero if any, or a negative number otherwise.
     *
     * \note This very low level interface using pointers, etc. is to:
-    *   1 - reduce the number of instanciations to the strict minimum
-    *   2 - avoid infinite recursion of the instanciations with Block<Block<Block<...> > >
+    *   1 - reduce the number of instantiations to the strict minimum
+    *   2 - avoid infinite recursion of the instantiations with Block<Block<Block<...> > >
     */
   static Index blocked_lu(Index rows, Index cols, Scalar* lu_data, Index luStride, PivIndex* row_transpositions, PivIndex& nb_transpositions, Index maxBlockSize=256)
   {
-    MapLU lu1(lu_data,StorageOrder==RowMajor?rows:luStride,StorageOrder==RowMajor?luStride:cols);
-    MatrixType lu(lu1,0,0,rows,cols);
+    MatrixTypeRef lu = MatrixType::Map(lu_data,rows, cols, OuterStride<>(luStride));
 
     const Index size = (std::min)(rows,cols);
 
     // if the matrix is too small, no blocking:
-    if(size<=16)
+    if(UnBlockedAtCompileTime || size<=UnBlockedBound)
     {
       return unblocked_lu(lu, row_transpositions, nb_transpositions);
     }
@@ -457,12 +460,12 @@ struct partial_lu_impl
       //                          A00 | A01 | A02
       // lu  = A_0 | A_1 | A_2 =  A10 | A11 | A12
       //                          A20 | A21 | A22
-      BlockType A_0(lu,0,0,rows,k);
-      BlockType A_2(lu,0,k+bs,rows,tsize);
-      BlockType A11(lu,k,k,bs,bs);
-      BlockType A12(lu,k,k+bs,bs,tsize);
-      BlockType A21(lu,k+bs,k,trows,bs);
-      BlockType A22(lu,k+bs,k+bs,trows,tsize);
+      BlockType A_0 = lu.block(0,0,rows,k);
+      BlockType A_2 = lu.block(0,k+bs,rows,tsize);
+      BlockType A11 = lu.block(k,k,bs,bs);
+      BlockType A12 = lu.block(k,k+bs,bs,tsize);
+      BlockType A21 = lu.block(k+bs,k,trows,bs);
+      BlockType A22 = lu.block(k+bs,k+bs,trows,tsize);
 
       PivIndex nb_transpositions_in_panel;
       // recursively call the blocked LU algorithm on [A11^T A21^T]^T
@@ -501,11 +504,18 @@ struct partial_lu_impl
 template<typename MatrixType, typename TranspositionType>
 void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, typename TranspositionType::StorageIndex& nb_transpositions)
 {
+  // Special-case of zero matrix.
+  if (lu.rows() == 0 || lu.cols() == 0) {
+    nb_transpositions = 0;
+    return;
+  }
   eigen_assert(lu.cols() == row_transpositions.size());
-  eigen_assert((&row_transpositions.coeffRef(1)-&row_transpositions.coeffRef(0)) == 1);
+  eigen_assert(row_transpositions.size() < 2 || (&row_transpositions.coeffRef(1)-&row_transpositions.coeffRef(0)) == 1);
 
   partial_lu_impl
-    <typename MatrixType::Scalar, MatrixType::Flags&RowMajorBit?RowMajor:ColMajor, typename TranspositionType::StorageIndex>
+    < typename MatrixType::Scalar, MatrixType::Flags&RowMajorBit?RowMajor:ColMajor,
+      typename TranspositionType::StorageIndex,
+      EIGEN_SIZE_MIN_PREFER_FIXED(MatrixType::RowsAtCompileTime,MatrixType::ColsAtCompileTime)>
     ::blocked_lu(lu.rows(), lu.cols(), &lu.coeffRef(0,0), lu.outerStride(), &row_transpositions.coeffRef(0), nb_transpositions);
 }
 
@@ -519,7 +529,10 @@ void PartialPivLU<MatrixType>::compute()
   // the row permutation is stored as int indices, so just to be sure:
   eigen_assert(m_lu.rows()<NumTraits<int>::highest());
 
-  m_l1_norm = m_lu.cwiseAbs().colwise().sum().maxCoeff();
+  if(m_lu.cols()>0)
+    m_l1_norm = m_lu.cwiseAbs().colwise().sum().maxCoeff();
+  else
+    m_l1_norm = RealScalar(0);
 
   eigen_assert(m_lu.rows() == m_lu.cols() && "PartialPivLU is only for square (and moreover invertible) matrices");
   const Index size = m_lu.rows();