From bffdc491b304cb207ed63cc421f39e597f8a6a9d Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sun, 11 Aug 2013 19:21:43 +0200
Subject: [PATCH] Fix cost evaluation of partial reduxions -> improve
 performance of vectorwise/replicate expressions involving partial reduxions

---
 Eigen/src/Core/VectorwiseOp.h |  5 +++--
 test/vectorwiseop.cpp         | 10 ++++++++++
 2 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h
index 328b6179b..f25ddca17 100644
--- a/Eigen/src/Core/VectorwiseOp.h
+++ b/Eigen/src/Core/VectorwiseOp.h
@@ -50,7 +50,7 @@ struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> >
     MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime,
     Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits,
     Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0),
-    TraversalSize = Direction==Vertical ? RowsAtCompileTime : ColsAtCompileTime
+    TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime :  MatrixType::ColsAtCompileTime
   };
   #if EIGEN_GNUC_AT_LEAST(3,4)
   typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
@@ -58,7 +58,8 @@ struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> >
   typedef typename MemberOp::template Cost<InputScalar,TraversalSize> CostOpType;
   #endif
   enum {
-    CoeffReadCost = TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value)
+    CoeffReadCost = TraversalSize==Dynamic ? Dynamic
+                  : TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value)
   };
 };
 }
diff --git a/test/vectorwiseop.cpp b/test/vectorwiseop.cpp
index 9d60b0388..6cd1acdda 100644
--- a/test/vectorwiseop.cpp
+++ b/test/vectorwiseop.cpp
@@ -101,6 +101,16 @@ template<typename ArrayType> void vectorwiseop_array(const ArrayType& m)
 
   VERIFY_RAISES_ASSERT(m2.rowwise() /= rowvec.transpose());
   VERIFY_RAISES_ASSERT(m1.rowwise() / rowvec.transpose());
+  
+  m2 = m1;
+  // yes, there might be an aliasing issue there but ".rowwise() /="
+  // is suppposed to evaluate " m2.colwise().sum()" into to temporary to avoid
+  // evaluating the reducions multiple times
+  if(ArrayType::RowsAtCompileTime>2 || ArrayType::RowsAtCompileTime==Dynamic)
+  {
+    m2.rowwise() /= m2.colwise().sum();
+    VERIFY_IS_APPROX(m2, m1.rowwise() / m1.colwise().sum());
+  }
 }
 
 template<typename MatrixType> void vectorwiseop_matrix(const MatrixType& m)