- merge ei_xpr_copy and ei_eval_if_needed_before_nesting

- make use of CoeffReadCost to determine when to unroll the loops,
  for now only in Product.h and in OperatorEquals.h
performance remains the same: generally still not as good as before the
big changes.

Eigen/src/Core/ForwardDeclarations.h

@@ -83,27 +83,30 @@ template<typename T> struct ei_eval
 template<typename T> struct ei_unref { typedef T type; };
 template<typename T> struct ei_unref<T&> { typedef T type; };
 
-template<typename T> struct ei_xpr_copy
+template<typename T> struct ei_is_temporary
 {
-  typedef typename ei_meta_if< ei_traits<T>::Flags & EvalBeforeNestingBit,
-      typename ei_eval<T>::type, const T&>::ret type;
+  enum { ret = 0 };
 };
 
-template<typename T> struct ei_xpr_copy<Temporary<T> >
+template<typename T> struct ei_is_temporary<Temporary<T> >
 {
-  typedef Temporary<T> type;
+  enum { ret = 1 };
 };
 
-template<typename T, int n=1> struct ei_eval_if_needed_before_nesting
+template<typename T, int n=1> struct ei_xpr_copy
 {
-  // FIXME should we consider the additional store as well as the creation cost of the temporary ?
-  enum { eval = T::Flags & EvalBeforeNestingBit
-             || (n+1) * NumTraits<typename ei_traits<T>::Scalar>::ReadCost < (n-1) * T::CoeffReadCost };
-  typedef typename ei_meta_if<eval, typename ei_eval<T>::type, T>::ret XprType;
-  typedef typename ei_meta_if<eval, typename ei_eval<T>::type, typename T::XprCopy>::ret CopyType;
+  typedef typename ei_meta_if<
+    ei_is_temporary<T>::ret,
+    T,
+    typename ei_meta_if<
+      ei_traits<T>::Flags & EvalBeforeNestingBit
+      || (n+1) * NumTraits<typename ei_traits<T>::Scalar>::ReadCost < (n-1) * T::CoeffReadCost,
+      typename ei_eval<T>::type,
+      const T&
+    >::ret
+  >::ret type;
 };
 
-
 template<typename T> struct ei_functor_traits
 {
   enum

Eigen/src/Core/OperatorEquals.h

@@ -102,14 +102,15 @@ template<typename OtherDerived>
 Derived& MatrixBase<Derived>
   ::lazyAssign(const MatrixBase<OtherDerived>& other)
 {
+  const bool unroll = SizeAtCompileTime * OtherDerived::CoeffReadCost <= EIGEN_UNROLLING_LIMIT;
   if(IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime)
     // copying a vector expression into a vector
   {
     ei_assert(size() == other.size());
-    if(SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+    if(unroll)
       ei_vector_operator_equals_unroller
         <Derived, OtherDerived,
-        SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT ? SizeAtCompileTime : Dynamic
+        unroll ? SizeAtCompileTime : Dynamic
         >::run(derived(), other.derived());
     else
       for(int i = 0; i < size(); i++)
@@ -118,11 +119,11 @@ Derived& MatrixBase<Derived>
   else // copying a matrix expression into a matrix
   {
     ei_assert(rows() == other.rows() && cols() == other.cols());
-    if(SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+    if(unroll)
     {
       ei_matrix_operator_equals_unroller
         <Derived, OtherDerived,
-        SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT ? SizeAtCompileTime : Dynamic
+        unroll ? SizeAtCompileTime : Dynamic
         >::run(derived(), other.derived());
     }
     else

Eigen/src/Core/Product.h

@@ -84,21 +84,29 @@ template<typename Lhs, typename Rhs, int EvalMode>
 struct ei_traits<Product<Lhs, Rhs, EvalMode> >
 {
   typedef typename Lhs::Scalar Scalar;
+  typedef typename ei_xpr_copy<Lhs,Rhs::ColsAtCompileTime>::type LhsXprCopy;
+  typedef typename ei_xpr_copy<Rhs,Lhs::RowsAtCompileTime>::type RhsXprCopy;
+  typedef typename ei_unref<LhsXprCopy>::type ActualLhs;
+  typedef typename ei_unref<RhsXprCopy>::type ActualRhs;
   enum {
+    LhsCoeffReadCost = ActualLhs::CoeffReadCost,
+    RhsCoeffReadCost = ActualRhs::CoeffReadCost,
+    LhsFlags = ActualLhs::Flags,
+    RhsFlags = ActualRhs::Flags,
     RowsAtCompileTime = Lhs::RowsAtCompileTime,
     ColsAtCompileTime = Rhs::ColsAtCompileTime,
     MaxRowsAtCompileTime = Lhs::MaxRowsAtCompileTime,
     MaxColsAtCompileTime = Rhs::MaxColsAtCompileTime,
     Flags = ( (RowsAtCompileTime == Dynamic || ColsAtCompileTime == Dynamic)
-              ? (unsigned int)(Lhs::Flags | Rhs::Flags)
-              : (unsigned int)(Lhs::Flags | Rhs::Flags) & ~LargeBit )
+              ? (unsigned int)(LhsFlags | RhsFlags)
+              : (unsigned int)(LhsFlags | RhsFlags) & ~LargeBit )
           | EvalBeforeAssigningBit
           | (ei_product_eval_mode<Lhs, Rhs>::value == (int)CacheOptimal ? EvalBeforeNestingBit : 0),
     CoeffReadCost
       = Lhs::ColsAtCompileTime == Dynamic
       ? Dynamic
       : Lhs::ColsAtCompileTime
-        * (NumTraits<Scalar>::MulCost + Lhs::CoeffReadCost + Rhs::CoeffReadCost)
+        * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost)
         + (Lhs::ColsAtCompileTime - 1) * NumTraits<Scalar>::AddCost
   };
 };
@@ -110,10 +118,8 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
 
     EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
 
-    typedef typename ei_eval_if_needed_before_nesting<Lhs,Rhs::ColsAtCompileTime>::CopyType CopyLhs;
-    typedef typename ei_eval_if_needed_before_nesting<Rhs,Lhs::RowsAtCompileTime>::CopyType CopyRhs;
-    typedef typename ei_eval_if_needed_before_nesting<Lhs,Rhs::ColsAtCompileTime>::XprType XprLhs;
-    typedef typename ei_eval_if_needed_before_nesting<Rhs,Lhs::RowsAtCompileTime>::XprType XprRhs;
+    typedef typename ei_traits<Product>::LhsXprCopy LhsXprCopy;
+    typedef typename ei_traits<Product>::RhsXprCopy RhsXprCopy;
 
     Product(const Lhs& lhs, const Rhs& rhs)
       : m_lhs(lhs), m_rhs(rhs)
@@ -133,12 +139,15 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
     const Scalar _coeff(int row, int col) const
     {
       Scalar res;
-      if(Lhs::ColsAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+      const bool unroll = CoeffReadCost <= EIGEN_UNROLLING_LIMIT;
+      if(unroll)
+      {
         ei_product_unroller<Lhs::ColsAtCompileTime-1,
-                            Lhs::ColsAtCompileTime <= EIGEN_UNROLLING_LIMIT
-                              ? Lhs::ColsAtCompileTime : Dynamic,
-                            XprLhs, XprRhs>
+                            unroll ? Lhs::ColsAtCompileTime : Dynamic,
+                            typename ei_unref<LhsXprCopy>::type,
+                            typename ei_unref<RhsXprCopy>::type>
           ::run(row, col, m_lhs, m_rhs, res);
+      }
       else
       {
         res = m_lhs.coeff(row, 0) * m_rhs.coeff(0, col);
@@ -149,8 +158,8 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
     }
 
   protected:
-    const CopyLhs m_lhs;
-    const CopyRhs m_rhs;
+    const LhsXprCopy m_lhs;
+    const RhsXprCopy m_rhs;
 };
 
 /** \returns the matrix product of \c *this and \a other.

Eigen/src/Core/Util.h

@@ -31,7 +31,7 @@
 
 /** Defines the maximal loop size to enable meta unrolling of loops */
 #ifndef EIGEN_UNROLLING_LIMIT
-#define EIGEN_UNROLLING_LIMIT 16
+#define EIGEN_UNROLLING_LIMIT 400
 #endif
 
 #ifdef EIGEN_DEFAULT_TO_ROW_MAJOR

bench/benchmark.cpp

@@ -8,6 +8,10 @@
 using namespace std;
 USING_PART_OF_NAMESPACE_EIGEN
 
+#ifndef REPEAT
+#define REPEAT 40000000
+#endif
+
 int main(int argc, char *argv[])
 {
     Matrix<double,MATSIZE,MATSIZE> I;
@@ -19,7 +23,7 @@ int main(int argc, char *argv[])
             m(i,j) = (i+MATSIZE*j);
         }
     asm("#begin");
-    for(int a = 0; a < 40000000; a++)
+    for(int a = 0; a < REPEAT; a++)
     {
         m = I + 0.00005 * (m + m*m);
     }