@smake/eigen 1.0.2 → 1.1.0

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

@@ -83,7 +83,7 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src)
     // eval without temporary
     dst.resize(src.rows(), src.cols());
     dst.setZero();
-    dst.reserve((std::max)(src.rows(),src.cols())*2);
+    dst.reserve((std::min)(src.rows()*src.cols(), (std::max)(src.rows(),src.cols())*2));
     for (Index j=0; j<outerEvaluationSize; ++j)
     {
       dst.startVec(j);
@@ -107,7 +107,7 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src)
     
     DstXprType temp(src.rows(), src.cols());
 
-    temp.reserve((std::max)(src.rows(),src.cols())*2);
+    temp.reserve((std::min)(src.rows()*src.cols(), (std::max)(src.rows(),src.cols())*2));
     for (Index j=0; j<outerEvaluationSize; ++j)
     {
       temp.startVec(j);
@@ -134,8 +134,8 @@ struct Assignment<DstXprType, SrcXprType, Functor, Sparse2Sparse>
 };
 
 // Generic Sparse to Dense assignment
-template< typename DstXprType, typename SrcXprType, typename Functor>
-struct Assignment<DstXprType, SrcXprType, Functor, Sparse2Dense>
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Weak>
+struct Assignment<DstXprType, SrcXprType, Functor, Sparse2Dense, Weak>
 {
   static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
   {
@@ -153,6 +153,73 @@ struct Assignment<DstXprType, SrcXprType, Functor, Sparse2Dense>
   }
 };
 
+// Specialization for dense ?= dense +/- sparse and dense ?= sparse +/- dense
+template<typename DstXprType, typename Func1, typename Func2>
+struct assignment_from_dense_op_sparse
+{
+  template<typename SrcXprType, typename InitialFunc>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const InitialFunc& /*func*/)
+  {
+    #ifdef EIGEN_SPARSE_ASSIGNMENT_FROM_DENSE_OP_SPARSE_PLUGIN
+    EIGEN_SPARSE_ASSIGNMENT_FROM_DENSE_OP_SPARSE_PLUGIN
+    #endif
+
+    call_assignment_no_alias(dst, src.lhs(), Func1());
+    call_assignment_no_alias(dst, src.rhs(), Func2());
+  }
+
+  // Specialization for dense1 = sparse + dense2; -> dense1 = dense2; dense1 += sparse;
+  template<typename Lhs, typename Rhs, typename Scalar>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::is_same<typename internal::evaluator_traits<Rhs>::Shape,DenseShape>::value>::type
+  run(DstXprType &dst, const CwiseBinaryOp<internal::scalar_sum_op<Scalar,Scalar>, const Lhs, const Rhs> &src,
+      const internal::assign_op<typename DstXprType::Scalar,Scalar>& /*func*/)
+  {
+    #ifdef EIGEN_SPARSE_ASSIGNMENT_FROM_SPARSE_ADD_DENSE_PLUGIN
+    EIGEN_SPARSE_ASSIGNMENT_FROM_SPARSE_ADD_DENSE_PLUGIN
+    #endif
+
+    // Apply the dense matrix first, then the sparse one.
+    call_assignment_no_alias(dst, src.rhs(), Func1());
+    call_assignment_no_alias(dst, src.lhs(), Func2());
+  }
+
+  // Specialization for dense1 = sparse - dense2; -> dense1 = -dense2; dense1 += sparse;
+  template<typename Lhs, typename Rhs, typename Scalar>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::is_same<typename internal::evaluator_traits<Rhs>::Shape,DenseShape>::value>::type
+  run(DstXprType &dst, const CwiseBinaryOp<internal::scalar_difference_op<Scalar,Scalar>, const Lhs, const Rhs> &src,
+      const internal::assign_op<typename DstXprType::Scalar,Scalar>& /*func*/)
+  {
+    #ifdef EIGEN_SPARSE_ASSIGNMENT_FROM_SPARSE_SUB_DENSE_PLUGIN
+    EIGEN_SPARSE_ASSIGNMENT_FROM_SPARSE_SUB_DENSE_PLUGIN
+    #endif
+
+    // Apply the dense matrix first, then the sparse one.
+    call_assignment_no_alias(dst, -src.rhs(), Func1());
+    call_assignment_no_alias(dst,  src.lhs(), add_assign_op<typename DstXprType::Scalar,typename Lhs::Scalar>());
+  }
+};
+
+#define EIGEN_CATCH_ASSIGN_DENSE_OP_SPARSE(ASSIGN_OP,BINOP,ASSIGN_OP2) \
+  template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> \
+  struct Assignment<DstXprType, CwiseBinaryOp<internal::BINOP<Scalar,Scalar>, const Lhs, const Rhs>, internal::ASSIGN_OP<typename DstXprType::Scalar,Scalar>, \
+                    Sparse2Dense, \
+                    typename internal::enable_if<   internal::is_same<typename internal::evaluator_traits<Lhs>::Shape,DenseShape>::value \
+                                                 || internal::is_same<typename internal::evaluator_traits<Rhs>::Shape,DenseShape>::value>::type> \
+    : assignment_from_dense_op_sparse<DstXprType, internal::ASSIGN_OP<typename DstXprType::Scalar,typename Lhs::Scalar>, internal::ASSIGN_OP2<typename DstXprType::Scalar,typename Rhs::Scalar> > \
+  {}
+
+EIGEN_CATCH_ASSIGN_DENSE_OP_SPARSE(assign_op,    scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_DENSE_OP_SPARSE(add_assign_op,scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_DENSE_OP_SPARSE(sub_assign_op,scalar_sum_op,sub_assign_op);
+
+EIGEN_CATCH_ASSIGN_DENSE_OP_SPARSE(assign_op,    scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_DENSE_OP_SPARSE(add_assign_op,scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_DENSE_OP_SPARSE(sub_assign_op,scalar_difference_op,add_assign_op);
+
+
 // Specialization for "dst = dec.solve(rhs)"
 // NOTE we need to specialize it for Sparse2Sparse to avoid ambiguous specialization error
 template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
@@ -179,35 +246,22 @@ struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Sparse>
 {
   typedef typename DstXprType::StorageIndex StorageIndex;
   typedef typename DstXprType::Scalar Scalar;
-  typedef Array<StorageIndex,Dynamic,1> ArrayXI;
-  typedef Array<Scalar,Dynamic,1> ArrayXS;
-  template<int Options>
-  static void run(SparseMatrix<Scalar,Options,StorageIndex> &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
-  {
-    Index dstRows = src.rows();
-    Index dstCols = src.cols();
-    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
-      dst.resize(dstRows, dstCols);
 
-    Index size = src.diagonal().size();
-    dst.makeCompressed();
-    dst.resizeNonZeros(size);
-    Map<ArrayXI>(dst.innerIndexPtr(), size).setLinSpaced(0,StorageIndex(size)-1);
-    Map<ArrayXI>(dst.outerIndexPtr(), size+1).setLinSpaced(0,StorageIndex(size));
-    Map<ArrayXS>(dst.valuePtr(), size) = src.diagonal();
-  }
+  template<int Options, typename AssignFunc>
+  static void run(SparseMatrix<Scalar,Options,StorageIndex> &dst, const SrcXprType &src, const AssignFunc &func)
+  { dst.assignDiagonal(src.diagonal(), func); }
   
   template<typename DstDerived>
   static void run(SparseMatrixBase<DstDerived> &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
-  {
-    dst.diagonal() = src.diagonal();
-  }
+  { dst.derived().diagonal() = src.diagonal(); }
   
-  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
-  { dst.diagonal() += src.diagonal(); }
+  template<typename DstDerived>
+  static void run(SparseMatrixBase<DstDerived> &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.derived().diagonal() += src.diagonal(); }
   
-  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
-  { dst.diagonal() -= src.diagonal(); }
+  template<typename DstDerived>
+  static void run(SparseMatrixBase<DstDerived> &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.derived().diagonal() -= src.diagonal(); }
 };
 } // end namespace internal
 

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

@@ -164,7 +164,7 @@ public:
       }
       else
       {
-        if(m_matrix.isCompressed())
+        if(m_matrix.isCompressed() && nnz!=block_size)
         {
           // no need to realloc, simply copy the tail at its respective position and insert tmp
           matrix.data().resize(start + nnz + tail_size);
@@ -326,46 +326,6 @@ private:
 
 //----------
 
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major).
-  */
-template<typename Derived>
-typename SparseMatrixBase<Derived>::InnerVectorReturnType SparseMatrixBase<Derived>::innerVector(Index outer)
-{ return InnerVectorReturnType(derived(), outer); }
-
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major). Read-only.
-  */
-template<typename Derived>
-const typename SparseMatrixBase<Derived>::ConstInnerVectorReturnType SparseMatrixBase<Derived>::innerVector(Index outer) const
-{ return ConstInnerVectorReturnType(derived(), outer); }
-
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major).
-  */
-template<typename Derived>
-typename SparseMatrixBase<Derived>::InnerVectorsReturnType
-SparseMatrixBase<Derived>::innerVectors(Index outerStart, Index outerSize)
-{
-  return Block<Derived,Dynamic,Dynamic,true>(derived(),
-                                             IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
-                                             IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
-
-}
-
-/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
-  * is col-major (resp. row-major). Read-only.
-  */
-template<typename Derived>
-const typename SparseMatrixBase<Derived>::ConstInnerVectorsReturnType
-SparseMatrixBase<Derived>::innerVectors(Index outerStart, Index outerSize) const
-{
-  return Block<const Derived,Dynamic,Dynamic,true>(derived(),
-                                                  IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
-                                                  IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
-
-}
-
 /** Generic implementation of sparse Block expression.
   * Real-only.
   */
@@ -486,9 +446,13 @@ struct unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBa
     {}
 
     inline Index nonZerosEstimate() const {
-      Index nnz = m_block.nonZeros();
-      if(nnz<0)
-        return m_argImpl.nonZerosEstimate() * m_block.size() / m_block.nestedExpression().size();
+      const Index nnz = m_block.nonZeros();
+      if(nnz < 0) {
+        // Scale the non-zero estimate for the underlying expression linearly with block size.
+        // Return zero if the underlying block is empty.
+        const Index nested_sz = m_block.nestedExpression().size();        
+        return nested_sz == 0 ? 0 : m_argImpl.nonZerosEstimate() * m_block.size() / nested_sz;
+      }
       return nnz;
     }
 
@@ -503,22 +467,25 @@ template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
 class unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBased>::InnerVectorInnerIterator
  : public EvalIterator
 {
-  enum { IsRowMajor = unary_evaluator::IsRowMajor };
+  // NOTE MSVC fails to compile if we don't explicitely "import" IsRowMajor from unary_evaluator
+  //      because the base class EvalIterator has a private IsRowMajor enum too. (bug #1786)
+  // NOTE We cannot call it IsRowMajor because it would shadow unary_evaluator::IsRowMajor
+  enum { XprIsRowMajor = unary_evaluator::IsRowMajor };
   const XprType& m_block;
   Index m_end;
 public:
 
   EIGEN_STRONG_INLINE InnerVectorInnerIterator(const unary_evaluator& aEval, Index outer)
-    : EvalIterator(aEval.m_argImpl, outer + (IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol())),
+    : EvalIterator(aEval.m_argImpl, outer + (XprIsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol())),
       m_block(aEval.m_block),
-      m_end(IsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows())
+      m_end(XprIsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows())
   {
-    while( (EvalIterator::operator bool()) && (EvalIterator::index() < (IsRowMajor ? m_block.startCol() : m_block.startRow())) )
+    while( (EvalIterator::operator bool()) && (EvalIterator::index() < (XprIsRowMajor ? m_block.startCol() : m_block.startRow())) )
       EvalIterator::operator++();
   }
 
-  inline StorageIndex index() const { return EvalIterator::index() - convert_index<StorageIndex>(IsRowMajor ? m_block.startCol() : m_block.startRow()); }
-  inline Index outer()  const { return EvalIterator::outer() - (IsRowMajor ? m_block.startRow() : m_block.startCol()); }
+  inline StorageIndex index() const { return EvalIterator::index() - convert_index<StorageIndex>(XprIsRowMajor ? m_block.startCol() : m_block.startRow()); }
+  inline Index outer()  const { return EvalIterator::outer() - (XprIsRowMajor ? m_block.startRow() : m_block.startCol()); }
   inline Index row()    const { return EvalIterator::row()   - m_block.startRow(); }
   inline Index col()    const { return EvalIterator::col()   - m_block.startCol(); }
 
@@ -528,7 +495,8 @@ public:
 template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
 class unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBased>::OuterVectorInnerIterator
 {
-  enum { IsRowMajor = unary_evaluator::IsRowMajor };
+  // NOTE see above
+  enum { XprIsRowMajor = unary_evaluator::IsRowMajor };
   const unary_evaluator& m_eval;
   Index m_outerPos;
   const Index m_innerIndex;
@@ -538,9 +506,9 @@ public:
 
   EIGEN_STRONG_INLINE OuterVectorInnerIterator(const unary_evaluator& aEval, Index outer)
     : m_eval(aEval),
-      m_outerPos( (IsRowMajor ? aEval.m_block.startCol() : aEval.m_block.startRow()) ),
-      m_innerIndex(IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol()),
-      m_end(IsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows()),
+      m_outerPos( (XprIsRowMajor ? aEval.m_block.startCol() : aEval.m_block.startRow()) ),
+      m_innerIndex(XprIsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol()),
+      m_end(XprIsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows()),
       m_it(m_eval.m_argImpl, m_outerPos)
   {
     EIGEN_UNUSED_VARIABLE(outer);
@@ -551,10 +519,10 @@ public:
       ++(*this);
   }
 
-  inline StorageIndex index() const { return convert_index<StorageIndex>(m_outerPos - (IsRowMajor ? m_eval.m_block.startCol() : m_eval.m_block.startRow())); }
+  inline StorageIndex index() const { return convert_index<StorageIndex>(m_outerPos - (XprIsRowMajor ? m_eval.m_block.startCol() : m_eval.m_block.startRow())); }
   inline Index outer()  const { return 0; }
-  inline Index row()    const { return IsRowMajor ? 0 : index(); }
-  inline Index col()    const { return IsRowMajor ? index() : 0; }
+  inline Index row()    const { return XprIsRowMajor ? 0 : index(); }
+  inline Index col()    const { return XprIsRowMajor ? index() : 0; }
 
   inline Scalar value() const { return m_it.value(); }
   inline Scalar& valueRef() { return m_it.valueRef(); }

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

@@ -128,6 +128,28 @@ class SparseCompressedBase
   protected:
     /** Default constructor. Do nothing. */
     SparseCompressedBase() {}
+
+    /** \internal return the index of the coeff at (row,col) or just before if it does not exist.
+      * This is an analogue of std::lower_bound.
+      */
+    internal::LowerBoundIndex lower_bound(Index row, Index col) const
+    {
+      eigen_internal_assert(row>=0 && row<this->rows() && col>=0 && col<this->cols());
+
+      const Index outer = Derived::IsRowMajor ? row : col;
+      const Index inner = Derived::IsRowMajor ? col : row;
+
+      Index start = this->outerIndexPtr()[outer];
+      Index end = this->isCompressed() ? this->outerIndexPtr()[outer+1] : this->outerIndexPtr()[outer] + this->innerNonZeroPtr()[outer];
+      eigen_assert(end>=start && "you are using a non finalized sparse matrix or written coefficient does not exist");
+      internal::LowerBoundIndex p;
+      p.value = std::lower_bound(this->innerIndexPtr()+start, this->innerIndexPtr()+end,inner) - this->innerIndexPtr();
+      p.found = (p.value<end) && (this->innerIndexPtr()[p.value]==inner);
+      return p;
+    }
+
+    friend struct internal::evaluator<SparseCompressedBase<Derived> >;
+
   private:
     template<typename OtherDerived> explicit SparseCompressedBase(const SparseCompressedBase<OtherDerived>&);
 };
@@ -185,6 +207,14 @@ class SparseCompressedBase<Derived>::InnerIterator
     }
 
     inline InnerIterator& operator++() { m_id++; return *this; }
+    inline InnerIterator& operator+=(Index i) { m_id += i ; return *this; }
+
+    inline InnerIterator operator+(Index i) 
+    { 
+        InnerIterator result = *this;
+        result += i;
+        return result;
+    }
 
     inline const Scalar& value() const { return m_values[m_id]; }
     inline Scalar& valueRef() { return const_cast<Scalar&>(m_values[m_id]); }
@@ -245,6 +275,14 @@ class SparseCompressedBase<Derived>::ReverseInnerIterator
     }
 
     inline ReverseInnerIterator& operator--() { --m_id; return *this; }
+    inline ReverseInnerIterator& operator-=(Index i) { m_id -= i; return *this; }
+
+    inline ReverseInnerIterator operator-(Index i) 
+    {
+        ReverseInnerIterator result = *this;
+        result -= i;
+        return result;
+    }
 
     inline const Scalar& value() const { return m_values[m_id-1]; }
     inline Scalar& valueRef() { return const_cast<Scalar&>(m_values[m_id-1]); }
@@ -317,17 +355,8 @@ protected:
 
   Index find(Index row, Index col) const
   {
-    eigen_internal_assert(row>=0 && row<m_matrix->rows() && col>=0 && col<m_matrix->cols());
-
-    const Index outer = Derived::IsRowMajor ? row : col;
-    const Index inner = Derived::IsRowMajor ? col : row;
-
-    Index start = m_matrix->outerIndexPtr()[outer];
-    Index end = m_matrix->isCompressed() ? m_matrix->outerIndexPtr()[outer+1] : m_matrix->outerIndexPtr()[outer] + m_matrix->innerNonZeroPtr()[outer];
-    eigen_assert(end>=start && "you are using a non finalized sparse matrix or written coefficient does not exist");
-    const Index p = std::lower_bound(m_matrix->innerIndexPtr()+start, m_matrix->innerIndexPtr()+end,inner) - m_matrix->innerIndexPtr();
-
-    return ((p<end) && (m_matrix->innerIndexPtr()[p]==inner)) ? p : Dynamic;
+    internal::LowerBoundIndex p = m_matrix->lower_bound(row,col);
+    return p.found ? p.value : Dynamic;
   }
 
   const Derived *m_matrix;

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

@@ -101,7 +101,7 @@ public:
       }
       else
       {
-        m_value = 0; // this is to avoid a compilation warning
+        m_value = Scalar(0); // this is to avoid a compilation warning
         m_id = -1;
       }
       return *this;
@@ -126,7 +126,7 @@ public:
   
   
   enum {
-    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    CoeffReadCost = int(evaluator<Lhs>::CoeffReadCost) + int(evaluator<Rhs>::CoeffReadCost) + int(functor_traits<BinaryOp>::Cost),
     Flags = XprType::Flags
   };
   
@@ -211,9 +211,8 @@ public:
 
 
   enum {
-    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
-    // Expose storage order of the sparse expression
-    Flags = (XprType::Flags & ~RowMajorBit) | (int(Rhs::Flags)&RowMajorBit)
+    CoeffReadCost = int(evaluator<Lhs>::CoeffReadCost) + int(evaluator<Rhs>::CoeffReadCost) + int(functor_traits<BinaryOp>::Cost),
+    Flags = XprType::Flags
   };
 
   explicit binary_evaluator(const XprType& xpr)
@@ -299,9 +298,8 @@ public:
 
 
   enum {
-    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
-    // Expose storage order of the sparse expression
-    Flags = (XprType::Flags & ~RowMajorBit) | (int(Lhs::Flags)&RowMajorBit)
+    CoeffReadCost = int(evaluator<Lhs>::CoeffReadCost) + int(evaluator<Rhs>::CoeffReadCost) + int(functor_traits<BinaryOp>::Cost),
+    Flags = XprType::Flags
   };
 
   explicit binary_evaluator(const XprType& xpr)
@@ -459,7 +457,7 @@ public:
   
   
   enum {
-    CoeffReadCost = evaluator<LhsArg>::CoeffReadCost + evaluator<RhsArg>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    CoeffReadCost = int(evaluator<LhsArg>::CoeffReadCost) + int(evaluator<RhsArg>::CoeffReadCost) + int(functor_traits<BinaryOp>::Cost),
     Flags = XprType::Flags
   };
   
@@ -532,9 +530,8 @@ public:
   
   
   enum {
-    CoeffReadCost = evaluator<LhsArg>::CoeffReadCost + evaluator<RhsArg>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
-    // Expose storage order of the sparse expression
-    Flags = (XprType::Flags & ~RowMajorBit) | (int(RhsArg::Flags)&RowMajorBit)
+    CoeffReadCost = int(evaluator<LhsArg>::CoeffReadCost) + int(evaluator<RhsArg>::CoeffReadCost) + int(functor_traits<BinaryOp>::Cost),
+    Flags = XprType::Flags
   };
   
   explicit sparse_conjunction_evaluator(const XprType& xpr)
@@ -607,9 +604,8 @@ public:
   
   
   enum {
-    CoeffReadCost = evaluator<LhsArg>::CoeffReadCost + evaluator<RhsArg>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
-    // Expose storage order of the sparse expression
-    Flags = (XprType::Flags & ~RowMajorBit) | (int(LhsArg::Flags)&RowMajorBit)
+    CoeffReadCost = int(evaluator<LhsArg>::CoeffReadCost) + int(evaluator<RhsArg>::CoeffReadCost) + int(functor_traits<BinaryOp>::Cost),
+    Flags = XprType::Flags
   };
   
   explicit sparse_conjunction_evaluator(const XprType& xpr)

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

@@ -24,7 +24,7 @@ struct unary_evaluator<CwiseUnaryOp<UnaryOp,ArgType>, IteratorBased>
     class InnerIterator;
     
     enum {
-      CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost,
+      CoeffReadCost = int(evaluator<ArgType>::CoeffReadCost) + int(functor_traits<UnaryOp>::Cost),
       Flags = XprType::Flags
     };
     
@@ -79,7 +79,7 @@ struct unary_evaluator<CwiseUnaryView<ViewOp,ArgType>, IteratorBased>
     class InnerIterator;
     
     enum {
-      CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<ViewOp>::Cost,
+      CoeffReadCost = int(evaluator<ArgType>::CoeffReadCost) + int(functor_traits<ViewOp>::Cost),
       Flags = XprType::Flags
     };
     

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

@@ -88,10 +88,11 @@ struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, A
   typedef typename internal::remove_all<SparseLhsType>::type Lhs;
   typedef typename internal::remove_all<DenseRhsType>::type Rhs;
   typedef typename internal::remove_all<DenseResType>::type Res;
-  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename LhsEval::InnerIterator LhsInnerIterator;
   static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha)
   {
-    evaluator<Lhs> lhsEval(lhs);
+    LhsEval lhsEval(lhs);
     for(Index c=0; c<rhs.cols(); ++c)
     {
       for(Index j=0; j<lhs.outerSize(); ++j)
@@ -111,17 +112,38 @@ struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, t
   typedef typename internal::remove_all<SparseLhsType>::type Lhs;
   typedef typename internal::remove_all<DenseRhsType>::type Rhs;
   typedef typename internal::remove_all<DenseResType>::type Res;
-  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename LhsEval::InnerIterator LhsInnerIterator;
   static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha)
   {
-    evaluator<Lhs> lhsEval(lhs);
-    for(Index j=0; j<lhs.outerSize(); ++j)
+    Index n = lhs.rows();
+    LhsEval lhsEval(lhs);
+
+#ifdef EIGEN_HAS_OPENMP
+    Eigen::initParallel();
+    Index threads = Eigen::nbThreads();
+    // This 20000 threshold has been found experimentally on 2D and 3D Poisson problems.
+    // It basically represents the minimal amount of work to be done to be worth it.
+    if(threads>1 && lhsEval.nonZerosEstimate()*rhs.cols() > 20000)
     {
-      typename Res::RowXpr res_j(res.row(j));
-      for(LhsInnerIterator it(lhsEval,j); it ;++it)
-        res_j += (alpha*it.value()) * rhs.row(it.index());
+      #pragma omp parallel for schedule(dynamic,(n+threads*4-1)/(threads*4)) num_threads(threads)
+      for(Index i=0; i<n; ++i)
+        processRow(lhsEval,rhs,res,alpha,i);
+    }
+    else
+#endif
+    {
+      for(Index i=0; i<n; ++i)
+        processRow(lhsEval, rhs, res, alpha, i);
     }
   }
+
+  static void processRow(const LhsEval& lhsEval, const DenseRhsType& rhs, Res& res, const typename Res::Scalar& alpha, Index i)
+  {
+    typename Res::RowXpr res_i(res.row(i));
+    for(LhsInnerIterator it(lhsEval,i); it ;++it)
+      res_i += (alpha*it.value()) * rhs.row(it.index());
+  }
 };
 
 template<typename SparseLhsType, typename DenseRhsType, typename DenseResType>