@smake/eigen 1.0.2 → 1.1.0

package/eigen/Eigen/src/SparseCore/SparseMatrix.h

@@ -21,7 +21,7 @@ namespace Eigen {
   * This class implements a more versatile variants of the common \em compressed row/column storage format.
   * Each colmun's (resp. row) non zeros are stored as a pair of value with associated row (resp. colmiun) index.
   * All the non zeros are stored in a single large buffer. Unlike the \em compressed format, there might be extra
-  * space inbetween the nonzeros of two successive colmuns (resp. rows) such that insertion of new non-zero
+  * space in between the nonzeros of two successive colmuns (resp. rows) such that insertion of new non-zero
   * can be done with limited memory reallocation and copies.
   *
   * A call to the function makeCompressed() turns the matrix into the standard \em compressed format
@@ -99,6 +99,8 @@ class SparseMatrix
     typedef SparseCompressedBase<SparseMatrix> Base;
     using Base::convert_index;
     friend class SparseVector<_Scalar,0,_StorageIndex>;
+    template<typename, typename, typename, typename, typename>
+    friend struct internal::Assignment;
   public:
     using Base::isCompressed;
     using Base::nonZeros;
@@ -503,8 +505,8 @@ class SparseMatrix
         m_innerNonZeros[i] = m_outerIndex[i+1] - m_outerIndex[i]; 
       }
     }
-    
-    /** Suppresses all nonzeros which are \b much \b smaller \b than \a reference under the tolerence \a epsilon */
+
+    /** Suppresses all nonzeros which are \b much \b smaller \b than \a reference under the tolerance \a epsilon */
     void prune(const Scalar& reference, const RealScalar& epsilon = NumTraits<RealScalar>::dummy_precision())
     {
       prune(default_prunning_func(reference,epsilon));
@@ -577,10 +579,12 @@ class SparseMatrix
       else if (innerChange < 0) 
       {
         // Inner size decreased: allocate a new m_innerNonZeros
-        m_innerNonZeros = static_cast<StorageIndex*>(std::malloc((m_outerSize+outerChange+1) * sizeof(StorageIndex)));
+        m_innerNonZeros = static_cast<StorageIndex*>(std::malloc((m_outerSize + outerChange) * sizeof(StorageIndex)));
         if (!m_innerNonZeros) internal::throw_std_bad_alloc();
-        for(Index i = 0; i < m_outerSize; i++)
+        for(Index i = 0; i < m_outerSize + (std::min)(outerChange, Index(0)); i++)
           m_innerNonZeros[i] = m_outerIndex[i+1] - m_outerIndex[i];
+        for(Index i = m_outerSize; i < m_outerSize + outerChange; i++)
+          m_innerNonZeros[i] = 0;
       }
       
       // Change the m_innerNonZeros in case of a decrease of inner size
@@ -605,9 +609,9 @@ class SparseMatrix
       m_outerIndex = newOuterIndex;
       if (outerChange > 0)
       {
-        StorageIndex last = m_outerSize == 0 ? 0 : m_outerIndex[m_outerSize];
+        StorageIndex lastIdx = m_outerSize == 0 ? 0 : m_outerIndex[m_outerSize];
         for(Index i=m_outerSize; i<m_outerSize+outerChange+1; i++)          
-          m_outerIndex[i] = last; 
+          m_outerIndex[i] = lastIdx; 
       }
       m_outerSize += outerChange;
     }
@@ -781,6 +785,9 @@ class SparseMatrix
     template<typename OtherDerived>
     inline SparseMatrix& operator=(const EigenBase<OtherDerived>& other)
     { return Base::operator=(other.derived()); }
+
+    template<typename Lhs, typename Rhs>
+    inline SparseMatrix& operator=(const Product<Lhs,Rhs,AliasFreeProduct>& other);
 #endif // EIGEN_PARSED_BY_DOXYGEN
 
     template<typename OtherDerived>
@@ -896,6 +903,113 @@ public:
       m_data.index(p) = convert_index(inner);
       return (m_data.value(p) = Scalar(0));
     }
+protected:
+    struct IndexPosPair {
+      IndexPosPair(Index a_i, Index a_p) : i(a_i), p(a_p) {}
+      Index i;
+      Index p;
+    };
+
+    /** \internal assign \a diagXpr to the diagonal of \c *this
+      * There are different strategies:
+      *   1 - if *this is overwritten (Func==assign_op) or *this is empty, then we can work treat *this as a dense vector expression.
+      *   2 - otherwise, for each diagonal coeff,
+      *     2.a - if it already exists, then we update it,
+      *     2.b - otherwise, if *this is uncompressed and that the current inner-vector has empty room for at least 1 element, then we perform an in-place insertion.
+      *     2.c - otherwise, we'll have to reallocate and copy everything, so instead of doing so for each new element, it is recorded in a std::vector.
+      *   3 - at the end, if some entries failed to be inserted in-place, then we alloc a new buffer, copy each chunk at the right position, and insert the new elements.
+      * 
+      * TODO: some piece of code could be isolated and reused for a general in-place update strategy.
+      * TODO: if we start to defer the insertion of some elements (i.e., case 2.c executed once),
+      *       then it *might* be better to disable case 2.b since they will have to be copied anyway.
+      */
+    template<typename DiagXpr, typename Func>
+    void assignDiagonal(const DiagXpr diagXpr, const Func& assignFunc)
+    {
+      Index n = diagXpr.size();
+
+      const bool overwrite = internal::is_same<Func, internal::assign_op<Scalar,Scalar> >::value;
+      if(overwrite)
+      {
+        if((this->rows()!=n) || (this->cols()!=n))
+          this->resize(n, n);
+      }
+
+      if(m_data.size()==0 || overwrite)
+      {
+        typedef Array<StorageIndex,Dynamic,1> ArrayXI;  
+        this->makeCompressed();
+        this->resizeNonZeros(n);
+        Eigen::Map<ArrayXI>(this->innerIndexPtr(), n).setLinSpaced(0,StorageIndex(n)-1);
+        Eigen::Map<ArrayXI>(this->outerIndexPtr(), n+1).setLinSpaced(0,StorageIndex(n));
+        Eigen::Map<Array<Scalar,Dynamic,1> > values = this->coeffs();
+        values.setZero();
+        internal::call_assignment_no_alias(values, diagXpr, assignFunc);
+      }
+      else
+      {
+        bool isComp = isCompressed();
+        internal::evaluator<DiagXpr> diaEval(diagXpr);
+        std::vector<IndexPosPair> newEntries;
+
+        // 1 - try in-place update and record insertion failures
+        for(Index i = 0; i<n; ++i)
+        {
+          internal::LowerBoundIndex lb = this->lower_bound(i,i);
+          Index p = lb.value;
+          if(lb.found)
+          {
+            // the coeff already exists
+            assignFunc.assignCoeff(m_data.value(p), diaEval.coeff(i));
+          }
+          else if((!isComp) && m_innerNonZeros[i] < (m_outerIndex[i+1]-m_outerIndex[i]))
+          {
+            // non compressed mode with local room for inserting one element
+            m_data.moveChunk(p, p+1, m_outerIndex[i]+m_innerNonZeros[i]-p);
+            m_innerNonZeros[i]++;
+            m_data.value(p) = Scalar(0);
+            m_data.index(p) = StorageIndex(i);
+            assignFunc.assignCoeff(m_data.value(p), diaEval.coeff(i));
+          }
+          else
+          {
+            // defer insertion
+            newEntries.push_back(IndexPosPair(i,p));
+          }
+        }
+        // 2 - insert deferred entries
+        Index n_entries = Index(newEntries.size());
+        if(n_entries>0)
+        {
+          Storage newData(m_data.size()+n_entries);
+          Index prev_p = 0;
+          Index prev_i = 0;
+          for(Index k=0; k<n_entries;++k)
+          {
+            Index i = newEntries[k].i;
+            Index p = newEntries[k].p;
+            internal::smart_copy(m_data.valuePtr()+prev_p, m_data.valuePtr()+p, newData.valuePtr()+prev_p+k);
+            internal::smart_copy(m_data.indexPtr()+prev_p, m_data.indexPtr()+p, newData.indexPtr()+prev_p+k);
+            for(Index j=prev_i;j<i;++j)
+              m_outerIndex[j+1] += k;
+            if(!isComp)
+              m_innerNonZeros[i]++;
+            prev_p = p;
+            prev_i = i;
+            newData.value(p+k) = Scalar(0);
+            newData.index(p+k) = StorageIndex(i);
+            assignFunc.assignCoeff(newData.value(p+k), diaEval.coeff(i));
+          }
+          {
+            internal::smart_copy(m_data.valuePtr()+prev_p, m_data.valuePtr()+m_data.size(), newData.valuePtr()+prev_p+n_entries);
+            internal::smart_copy(m_data.indexPtr()+prev_p, m_data.indexPtr()+m_data.size(), newData.indexPtr()+prev_p+n_entries);
+            for(Index j=prev_i+1;j<=m_outerSize;++j)
+              m_outerIndex[j] += n_entries;
+          }
+          m_data.swap(newData);
+        }
+      }
+    }
 
 private:
   static void check_template_parameters()
@@ -974,7 +1088,7 @@ void set_from_triplets(const InputIterator& begin, const InputIterator& end, Spa
   * \code
     typedef Triplet<double> T;
     std::vector<T> tripletList;
-    triplets.reserve(estimation_of_entries);
+    tripletList.reserve(estimation_of_entries);
     for(...)
     {
       // ...
@@ -987,7 +1101,7 @@ void set_from_triplets(const InputIterator& begin, const InputIterator& end, Spa
   *
   * \warning The list of triplets is read multiple times (at least twice). Therefore, it is not recommended to define
   * an abstract iterator over a complex data-structure that would be expensive to evaluate. The triplets should rather
-  * be explicitely stored into a std::vector for instance.
+  * be explicitly stored into a std::vector for instance.
   */
 template<typename Scalar, int _Options, typename _StorageIndex>
 template<typename InputIterators>
@@ -1233,7 +1347,7 @@ typename SparseMatrix<_Scalar,_Options,_StorageIndex>::Scalar& SparseMatrix<_Sca
     }
     
     m_data.index(p) = convert_index(inner);
-    return (m_data.value(p) = 0);
+    return (m_data.value(p) = Scalar(0));
   }
   
   if(m_data.size() != m_data.allocatedSize())

package/eigen/Eigen/src/SparseCore/SparseMatrixBase.h

@@ -87,6 +87,11 @@ template<typename Derived> class SparseMatrixBase
           * we are dealing with a column-vector (if there is only one column) or with
           * a row-vector (if there is only one row). */
 
+      NumDimensions = int(MaxSizeAtCompileTime) == 1 ? 0 : bool(IsVectorAtCompileTime) ? 1 : 2,
+        /**< This value is equal to Tensor::NumDimensions, i.e. 0 for scalars, 1 for vectors,
+         * and 2 for matrices.
+         */
+
       Flags = internal::traits<Derived>::Flags,
         /**< This stores expression \ref flags flags which may or may not be inherited by new expressions
           * constructed from this one. See the \ref flags "list of flags".
@@ -350,18 +355,6 @@ template<typename Derived> class SparseMatrixBase
     const ConstTransposeReturnType transpose() const { return ConstTransposeReturnType(derived()); }
     const AdjointReturnType adjoint() const { return AdjointReturnType(transpose()); }
 
-    // inner-vector
-    typedef Block<Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true>       InnerVectorReturnType;
-    typedef Block<const Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true> ConstInnerVectorReturnType;
-    InnerVectorReturnType innerVector(Index outer);
-    const ConstInnerVectorReturnType innerVector(Index outer) const;
-
-    // set of inner-vectors
-    typedef Block<Derived,Dynamic,Dynamic,true> InnerVectorsReturnType;
-    typedef Block<const Derived,Dynamic,Dynamic,true> ConstInnerVectorsReturnType;
-    InnerVectorsReturnType innerVectors(Index outerStart, Index outerSize);
-    const ConstInnerVectorsReturnType innerVectors(Index outerStart, Index outerSize) const;
-
     DenseMatrixType toDense() const
     {
       return DenseMatrixType(derived());

package/eigen/Eigen/src/SparseCore/SparseProduct.h

@@ -17,7 +17,7 @@ namespace Eigen {
   * The automatic pruning of the small values can be achieved by calling the pruned() function
   * in which case a totally different product algorithm is employed:
   * \code
-  * C = (A*B).pruned();             // supress numerical zeros (exact)
+  * C = (A*B).pruned();             // suppress numerical zeros (exact)
   * C = (A*B).pruned(ref);
   * C = (A*B).pruned(ref,epsilon);
   * \endcode
@@ -164,6 +164,18 @@ protected:
 
 } // end namespace internal
 
+// sparse matrix = sparse-product (can be sparse*sparse, sparse*perm, etc.)
+template<typename Scalar, int _Options, typename _StorageIndex>
+template<typename Lhs, typename Rhs>
+SparseMatrix<Scalar,_Options,_StorageIndex>& SparseMatrix<Scalar,_Options,_StorageIndex>::operator=(const Product<Lhs,Rhs,AliasFreeProduct>& src)
+{
+  // std::cout << "in Assignment : " << DstOptions << "\n";
+  SparseMatrix dst(src.rows(),src.cols());
+  internal::generic_product_impl<Lhs, Rhs>::evalTo(dst,src.lhs(),src.rhs());
+  this->swap(dst);
+  return *this;
+}
+
 } // end namespace Eigen
 
 #endif // EIGEN_SPARSEPRODUCT_H

package/eigen/Eigen/src/SparseCore/SparseRef.h

@@ -201,7 +201,7 @@ class Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType
 
     ~Ref() {
       if(m_hasCopy) {
-        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(&m_storage);
         obj->~TPlainObjectType();
       }
     }
@@ -213,7 +213,7 @@ class Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType
     {
       if((Options & int(StandardCompressedFormat)) && (!expr.isCompressed()))
       {
-        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(&m_storage);
         ::new (obj) TPlainObjectType(expr);
         m_hasCopy = true;
         Base::construct(*obj);
@@ -227,14 +227,14 @@ class Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType
     template<typename Expression>
     void construct(const Expression& expr, internal::false_type)
     {
-      TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+      TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(&m_storage);
       ::new (obj) TPlainObjectType(expr);
       m_hasCopy = true;
       Base::construct(*obj);
     }
 
   protected:
-    char m_object_bytes[sizeof(TPlainObjectType)];
+    typename internal::aligned_storage<sizeof(TPlainObjectType), EIGEN_ALIGNOF(TPlainObjectType)>::type m_storage;
     bool m_hasCopy;
 };
 
@@ -319,7 +319,7 @@ class Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType
 
     ~Ref() {
       if(m_hasCopy) {
-        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(&m_storage);
         obj->~TPlainObjectType();
       }
     }
@@ -335,14 +335,14 @@ class Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType
     template<typename Expression>
     void construct(const Expression& expr, internal::false_type)
     {
-      TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+      TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(&m_storage);
       ::new (obj) TPlainObjectType(expr);
       m_hasCopy = true;
       Base::construct(*obj);
     }
 
   protected:
-    char m_object_bytes[sizeof(TPlainObjectType)];
+    typename internal::aligned_storage<sizeof(TPlainObjectType), EIGEN_ALIGNOF(TPlainObjectType)>::type m_storage;
     bool m_hasCopy;
 };
 

package/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h

@@ -142,6 +142,9 @@ template<typename MatrixType, unsigned int _Mode> class SparseSelfAdjointView
       return *this = src.twistedBy(pnull);
     }
 
+    // Since we override the copy-assignment operator, we need to explicitly re-declare the copy-constructor
+    EIGEN_DEFAULT_COPY_CONSTRUCTOR(SparseSelfAdjointView)
+
     template<typename SrcMatrixType,unsigned int SrcMode>
     SparseSelfAdjointView& operator=(const SparseSelfAdjointView<SrcMatrixType,SrcMode>& src)
     {

package/eigen/Eigen/src/SparseCore/SparseUtil.h

@@ -140,6 +140,14 @@ struct SparseSelfAdjointShape { static std::string debugName() { return "SparseS
 template<> struct glue_shapes<SparseShape,SelfAdjointShape> { typedef SparseSelfAdjointShape type;  };
 template<> struct glue_shapes<SparseShape,TriangularShape > { typedef SparseTriangularShape  type;  };
 
+// return type of SparseCompressedBase::lower_bound;
+struct LowerBoundIndex {
+  LowerBoundIndex() : value(-1), found(false) {}
+  LowerBoundIndex(Index val, bool ok) : value(val), found(ok) {}
+  Index value;
+  bool found;
+};
+
 } // end namespace internal
 
 /** \ingroup SparseCore_Module

package/eigen/Eigen/src/SparseCore/SparseVector.h

@@ -281,7 +281,7 @@ class SparseVector
     }
 
     /** Swaps the values of \c *this and \a other.
-      * Overloaded for performance: this version performs a \em shallow swap by swaping pointers and attributes only.
+      * Overloaded for performance: this version performs a \em shallow swap by swapping pointers and attributes only.
       * \sa SparseMatrixBase::swap()
       */
     inline void swap(SparseVector& other)

package/eigen/Eigen/src/SparseLU/SparseLU.h

@@ -18,6 +18,63 @@ template <typename _MatrixType, typename _OrderingType = COLAMDOrdering<typename
 template <typename MappedSparseMatrixType> struct SparseLUMatrixLReturnType;
 template <typename MatrixLType, typename MatrixUType> struct SparseLUMatrixUReturnType;
 
+template <bool Conjugate,class SparseLUType>
+class SparseLUTransposeView : public SparseSolverBase<SparseLUTransposeView<Conjugate,SparseLUType> >
+{
+protected:
+  typedef SparseSolverBase<SparseLUTransposeView<Conjugate,SparseLUType> > APIBase;
+  using APIBase::m_isInitialized;
+public:
+  typedef typename SparseLUType::Scalar Scalar;
+  typedef typename SparseLUType::StorageIndex StorageIndex;
+  typedef typename SparseLUType::MatrixType MatrixType;
+  typedef typename SparseLUType::OrderingType OrderingType;
+
+  enum {
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+  };
+
+  SparseLUTransposeView() : m_sparseLU(NULL) {}
+  SparseLUTransposeView(const SparseLUTransposeView& view) {
+    this->m_sparseLU = view.m_sparseLU;
+  }
+  void setIsInitialized(const bool isInitialized) {this->m_isInitialized = isInitialized;}
+  void setSparseLU(SparseLUType* sparseLU) {m_sparseLU = sparseLU;}
+  using APIBase::_solve_impl;
+  template<typename Rhs, typename Dest>
+  bool _solve_impl(const MatrixBase<Rhs> &B, MatrixBase<Dest> &X_base) const
+  {
+    Dest& X(X_base.derived());
+    eigen_assert(m_sparseLU->info() == Success && "The matrix should be factorized first");
+    EIGEN_STATIC_ASSERT((Dest::Flags&RowMajorBit)==0,
+                        THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES);
+
+
+    // this ugly const_cast_derived() helps to detect aliasing when applying the permutations
+    for(Index j = 0; j < B.cols(); ++j){
+      X.col(j) = m_sparseLU->colsPermutation() * B.const_cast_derived().col(j);
+    }
+    //Forward substitution with transposed or adjoint of U
+    m_sparseLU->matrixU().template solveTransposedInPlace<Conjugate>(X);
+
+    //Backward substitution with transposed or adjoint of L
+    m_sparseLU->matrixL().template solveTransposedInPlace<Conjugate>(X);
+
+    // Permute back the solution
+    for (Index j = 0; j < B.cols(); ++j)
+      X.col(j) = m_sparseLU->rowsPermutation().transpose() * X.col(j);
+    return true;
+  }
+  inline Index rows() const { return m_sparseLU->rows(); }
+  inline Index cols() const { return m_sparseLU->cols(); }
+
+private:
+  SparseLUType *m_sparseLU;
+  SparseLUTransposeView& operator=(const SparseLUTransposeView&);
+};
+
+
 /** \ingroup SparseLU_Module
   * \class SparseLU
   * 
@@ -26,7 +83,7 @@ template <typename MatrixLType, typename MatrixUType> struct SparseLUMatrixURetu
   * This class implements the supernodal LU factorization for general matrices.
   * It uses the main techniques from the sequential SuperLU package 
   * (http://crd-legacy.lbl.gov/~xiaoye/SuperLU/). It handles transparently real 
-  * and complex arithmetics with single and double precision, depending on the 
+  * and complex arithmetic with single and double precision, depending on the 
   * scalar type of your input matrix. 
   * The code has been optimized to provide BLAS-3 operations during supernode-panel updates. 
   * It benefits directly from the built-in high-performant Eigen BLAS routines. 
@@ -97,6 +154,7 @@ class SparseLU : public SparseSolverBase<SparseLU<_MatrixType,_OrderingType> >,
     };
     
   public:
+
     SparseLU():m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1)
     {
       initperfvalues(); 
@@ -128,6 +186,45 @@ class SparseLU : public SparseSolverBase<SparseLU<_MatrixType,_OrderingType> >,
       //Factorize
       factorize(matrix);
     } 
+
+    /** \returns an expression of the transposed of the factored matrix.
+      *
+      * A typical usage is to solve for the transposed problem A^T x = b:
+      * \code
+      * solver.compute(A);
+      * x = solver.transpose().solve(b);
+      * \endcode
+      *
+      * \sa adjoint(), solve()
+      */
+    const SparseLUTransposeView<false,SparseLU<_MatrixType,_OrderingType> > transpose()
+    {
+      SparseLUTransposeView<false,  SparseLU<_MatrixType,_OrderingType> > transposeView;
+      transposeView.setSparseLU(this);
+      transposeView.setIsInitialized(this->m_isInitialized);
+      return transposeView;
+    }
+
+
+    /** \returns an expression of the adjoint of the factored matrix
+      *
+      * A typical usage is to solve for the adjoint problem A' x = b:
+      * \code
+      * solver.compute(A);
+      * x = solver.adjoint().solve(b);
+      * \endcode
+      *
+      * For real scalar types, this function is equivalent to transpose().
+      *
+      * \sa transpose(), solve()
+      */
+    const SparseLUTransposeView<true, SparseLU<_MatrixType,_OrderingType> > adjoint()
+    {
+      SparseLUTransposeView<true,  SparseLU<_MatrixType,_OrderingType> > adjointView;
+      adjointView.setSparseLU(this);
+      adjointView.setIsInitialized(this->m_isInitialized);
+      return adjointView;
+    }
     
     inline Index rows() const { return m_mat.rows(); }
     inline Index cols() const { return m_mat.cols(); }
@@ -193,7 +290,7 @@ class SparseLU : public SparseSolverBase<SparseLU<_MatrixType,_OrderingType> >,
     
     /** \brief Reports whether previous computation was successful.
       *
-      * \returns \c Success if computation was succesful,
+      * \returns \c Success if computation was successful,
       *          \c NumericalIssue if the LU factorization reports a problem, zero diagonal for instance
       *          \c InvalidInput if the input matrix is invalid
       *
@@ -355,6 +452,9 @@ class SparseLU : public SparseSolverBase<SparseLU<_MatrixType,_OrderingType> >,
       return (m_detPermR * m_detPermC) > 0 ? det : -det;
     }
 
+    Index nnzL() const { return m_nnzL; };
+    Index nnzU() const { return m_nnzU; };
+
   protected:
     // Functions 
     void initperfvalues()
@@ -391,7 +491,6 @@ class SparseLU : public SparseSolverBase<SparseLU<_MatrixType,_OrderingType> >,
   private:
     // Disable copy constructor 
     SparseLU (const SparseLU& );
-  
 }; // End class SparseLU
 
 
@@ -501,7 +600,6 @@ void SparseLU<MatrixType, OrderingType>::factorize(const MatrixType& matrix)
   
   m_isInitialized = true;
   
-  
   // Apply the column permutation computed in analyzepattern()
   //   m_mat = matrix * m_perm_c.inverse(); 
   m_mat = matrix;
@@ -585,7 +683,6 @@ void SparseLU<MatrixType, OrderingType>::factorize(const MatrixType& matrix)
   //  (a) a relaxed supernode at the bottom of the etree, or
   //  (b) panel_size contiguous columns, <panel_size> defined by the user
   Index jcol; 
-  IndexVector panel_histo(n);
   Index pivrow; // Pivotal row number in the original row matrix
   Index nseg1; // Number of segments in U-column above panel row jcol
   Index nseg; // Number of segments in each U-column 
@@ -704,13 +801,19 @@ struct SparseLUMatrixLReturnType : internal::no_assignment_operator
   typedef typename MappedSupernodalType::Scalar Scalar;
   explicit SparseLUMatrixLReturnType(const MappedSupernodalType& mapL) : m_mapL(mapL)
   { }
-  Index rows() { return m_mapL.rows(); }
-  Index cols() { return m_mapL.cols(); }
+  Index rows() const { return m_mapL.rows(); }
+  Index cols() const { return m_mapL.cols(); }
   template<typename Dest>
   void solveInPlace( MatrixBase<Dest> &X) const
   {
     m_mapL.solveInPlace(X);
   }
+  template<bool Conjugate, typename Dest>
+  void solveTransposedInPlace( MatrixBase<Dest> &X) const
+  {
+    m_mapL.template solveTransposedInPlace<Conjugate>(X);
+  }
+
   const MappedSupernodalType& m_mapL;
 };
 
@@ -721,8 +824,8 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator
   SparseLUMatrixUReturnType(const MatrixLType& mapL, const MatrixUType& mapU)
   : m_mapL(mapL),m_mapU(mapU)
   { }
-  Index rows() { return m_mapL.rows(); }
-  Index cols() { return m_mapL.cols(); }
+  Index rows() const { return m_mapL.rows(); }
+  Index cols() const { return m_mapL.cols(); }
 
   template<typename Dest>   void solveInPlace(MatrixBase<Dest> &X) const
   {
@@ -745,8 +848,9 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator
       }
       else
       {
+        // FIXME: the following lines should use Block expressions and not Map!
         Map<const Matrix<Scalar,Dynamic,Dynamic, ColMajor>, 0, OuterStride<> > A( &(m_mapL.valuePtr()[luptr]), nsupc, nsupc, OuterStride<>(lda) );
-        Map< Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor>, 0, OuterStride<> > U (&(X(fsupc,0)), nsupc, nrhs, OuterStride<>(n) );
+        Map< Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor>, 0, OuterStride<> > U (&(X.coeffRef(fsupc,0)), nsupc, nrhs, OuterStride<>(n) );
         U = A.template triangularView<Upper>().solve(U);
       }
 
@@ -764,6 +868,52 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator
       }
     } // End For U-solve
   }
+
+  template<bool Conjugate, typename Dest>   void solveTransposedInPlace(MatrixBase<Dest> &X) const
+  {
+    using numext::conj;
+    Index nrhs = X.cols();
+    Index n    = X.rows();
+    // Forward solve with U
+    for (Index k = 0; k <=  m_mapL.nsuper(); k++)
+    {
+      Index fsupc = m_mapL.supToCol()[k];
+      Index lda = m_mapL.colIndexPtr()[fsupc+1] - m_mapL.colIndexPtr()[fsupc]; // leading dimension
+      Index nsupc = m_mapL.supToCol()[k+1] - fsupc;
+      Index luptr = m_mapL.colIndexPtr()[fsupc];
+
+      for (Index j = 0; j < nrhs; ++j)
+      {
+        for (Index jcol = fsupc; jcol < fsupc + nsupc; jcol++)
+        {
+          typename MatrixUType::InnerIterator it(m_mapU, jcol);
+          for ( ; it; ++it)
+          {
+            Index irow = it.index();
+            X(jcol, j) -= X(irow, j) * (Conjugate? conj(it.value()): it.value());
+          }
+        }
+      }
+      if (nsupc == 1)
+      {
+        for (Index j = 0; j < nrhs; j++)
+        {
+          X(fsupc, j) /= (Conjugate? conj(m_mapL.valuePtr()[luptr]) : m_mapL.valuePtr()[luptr]);
+        }
+      }
+      else
+      {
+        Map<const Matrix<Scalar,Dynamic,Dynamic, ColMajor>, 0, OuterStride<> > A( &(m_mapL.valuePtr()[luptr]), nsupc, nsupc, OuterStride<>(lda) );
+        Map< Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor>, 0, OuterStride<> > U (&(X(fsupc,0)), nsupc, nrhs, OuterStride<>(n) );
+        if(Conjugate)
+          U = A.adjoint().template triangularView<Lower>().solve(U);
+        else
+          U = A.transpose().template triangularView<Lower>().solve(U);
+      }
+    }// End For U-solve
+  }
+
+
   const MatrixLType& m_mapL;
   const MatrixUType& m_mapU;
 };

package/eigen/Eigen/src/SparseLU/SparseLU_Memory.h

@@ -51,7 +51,7 @@ inline Index LUTempSpace(Index&m, Index& w)
 
 
 /** 
-  * Expand the existing storage to accomodate more fill-ins
+  * Expand the existing storage to accommodate more fill-ins
   * \param vec Valid pointer to the vector to allocate or expand
   * \param[in,out] length  At input, contain the current length of the vector that is to be increased. At output, length of the newly allocated vector
   * \param[in] nbElts Current number of elements in the factors