* Started support for unaligned vectorization.

* Introduce a new highly optimized matrix-matrix product for large
  matrices. The code is still highly experimental and it is activated
  only if you define EIGEN_WIP_PRODUCT at compile time.
  Currently the third dimension of the product must be a factor of
  the packet size (x4 for floats) and the right handed side matrix
  must be column major.
  Moreover, currently c = a*b; actually computes c += a*b !!
  Therefore, the code is provided for experimentation purpose only !
  These limitations will be fixed soon or later to become the default
  product implementation.

14 files changed, 654 insertions, 74 deletions

diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h
index 3862ae831..384059185 100644
--- a/Eigen/src/Core/Assign.h
+++ b/Eigen/src/Core/Assign.h
@@ -78,7 +78,7 @@ struct ei_matrix_assignment_packet_unroller
   {
     ei_matrix_assignment_packet_unroller<Derived1, Derived2,
       Index-ei_packet_traits<typename Derived1::Scalar>::size>::run(dst, src);
-    dst.writePacketCoeff(row, col, src.packetCoeff(row, col));
+    dst.template writePacketCoeff<Aligned>(row, col, src.template packetCoeff<Aligned>(row, col));
   }
 };
 
@@ -87,7 +87,7 @@ struct ei_matrix_assignment_packet_unroller<Derived1, Derived2, 0 >
 {
   static void run(Derived1 &dst, const Derived2 &src)
   {
-    dst.writePacketCoeff(0, 0, src.packetCoeff(0, 0));
+    dst.template writePacketCoeff<Aligned>(0, 0, src.template packetCoeff<Aligned>(0, 0));
   }
 };
 
@@ -211,7 +211,7 @@ struct ei_assignment_impl<Derived, OtherDerived, true, false>
             // FIXME the following is not really efficient
             int i = index/dst.rows();
             int j = index%dst.rows();
-            dst.writePacketCoeff(i, j, src.packetCoeff(i, j));
+            dst.template writePacketCoeff<Aligned>(i, j, src.template packetCoeff<Aligned>(i, j));
           }
           for(int i = alignedSize/dst.rows(); i < dst.rows(); i++)
             for(int j = alignedSize%dst.rows(); j < dst.cols(); j++)
@@ -222,7 +222,7 @@ struct ei_assignment_impl<Derived, OtherDerived, true, false>
 //           std::cout << "vectorized normal row major\n";
           for(int i = 0; i < dst.rows(); i++)
             for(int j = 0; j < dst.cols(); j+=ei_packet_traits<typename Derived::Scalar>::size)
-              dst.writePacketCoeff(i, j, src.packetCoeff(i, j));
+              dst.template writePacketCoeff<Aligned>(i, j, src.template packetCoeff<Aligned>(i, j));
         }
       }
       else
@@ -240,7 +240,7 @@ struct ei_assignment_impl<Derived, OtherDerived, true, false>
             // FIXME the following is not really efficient
             int i = index%dst.rows();
             int j = index/dst.rows();
-            dst.writePacketCoeff(i, j, src.packetCoeff(i, j));
+            dst.template writePacketCoeff<Aligned>(i, j, src.template packetCoeff<Aligned>(i, j));
           }
           for(int j = alignedSize/dst.rows(); j < dst.cols(); j++)
             for(int i = alignedSize%dst.rows(); i < dst.rows(); i++)
@@ -251,7 +251,7 @@ struct ei_assignment_impl<Derived, OtherDerived, true, false>
 //           std::cout << "vectorized normal col major\n";
           for(int j = 0; j < dst.cols(); j++)
             for(int i = 0; i < dst.rows(); i+=ei_packet_traits<typename Derived::Scalar>::size)
-              dst.writePacketCoeff(i, j, src.packetCoeff(i, j));
+              dst.template writePacketCoeff<Aligned>(i, j, src.template packetCoeff<Aligned>(i, j));
         }
       }
     }
diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h
index 6c00e0ba8..247a46b40 100644
--- a/Eigen/src/Core/Block.h
+++ b/Eigen/src/Core/Block.h
@@ -143,6 +143,18 @@ template<typename MatrixType, int BlockRows, int BlockCols> class Block
       return m_matrix.coeff(row + m_startRow.value(), col + m_startCol.value());
     }
 
+    template<int LoadMode>
+    PacketScalar _packetCoeff(int row, int col) const
+    {
+      return m_matrix.packetCoeff<UnAligned>(row + m_startRow.value(), col + m_startCol.value());
+    }
+
+    template<int LoadMode>
+    void _writePacketCoeff(int row, int col, const PacketScalar& x)
+    {
+      m_matrix.const_cast_derived().writePacketCoeff<UnAligned>(row + m_startRow.value(), col + m_startCol.value(), x);
+    }
+
   protected:
 
     const typename MatrixType::Nested m_matrix;
diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h
index 3de3e2dd5..31afbe0f1 100644
--- a/Eigen/src/Core/CwiseBinaryOp.h
+++ b/Eigen/src/Core/CwiseBinaryOp.h
@@ -103,9 +103,10 @@ class CwiseBinaryOp : ei_no_assignment_operator,
       return m_functor(m_lhs.coeff(row, col), m_rhs.coeff(row, col));
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int row, int col) const
     {
-      return m_functor.packetOp(m_lhs.packetCoeff(row, col), m_rhs.packetCoeff(row, col));
+      return m_functor.packetOp(m_lhs.template packetCoeff<LoadMode>(row, col), m_rhs.template packetCoeff<LoadMode>(row, col));
     }
 
   protected:
diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h
index e603280bf..bcc09d0c8 100644
--- a/Eigen/src/Core/CwiseNullaryOp.h
+++ b/Eigen/src/Core/CwiseNullaryOp.h
@@ -82,6 +82,7 @@ class CwiseNullaryOp : ei_no_assignment_operator,
       return m_functor(rows, cols);
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int, int) const
     {
       return m_functor.packetOp();
diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h
index 76c6ed818..438417829 100644
--- a/Eigen/src/Core/CwiseUnaryOp.h
+++ b/Eigen/src/Core/CwiseUnaryOp.h
@@ -82,9 +82,10 @@ class CwiseUnaryOp : ei_no_assignment_operator,
       return m_functor(m_matrix.coeff(row, col));
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int row, int col) const
     {
-      return m_functor.packetOp(m_matrix.packetCoeff(row, col));
+      return m_functor.packetOp(m_matrix.template packetCoeff<LoadMode>(row, col));
     }
 
   protected:
diff --git a/Eigen/src/Core/Lazy.h b/Eigen/src/Core/Lazy.h
index ffe777158..d2f763680 100644
--- a/Eigen/src/Core/Lazy.h
+++ b/Eigen/src/Core/Lazy.h
@@ -72,9 +72,10 @@ template<typename ExpressionType> class Lazy
       return m_expression.coeff(row, col);
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int row, int col) const
     {
-      return m_expression.packetCoeff(row, col);
+      return m_expression.template packetCoeff<LoadMode>(row, col);
     }
 
   protected:
diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h
index dd1235aa3..922c3ddae 100644
--- a/Eigen/src/Core/Matrix.h
+++ b/Eigen/src/Core/Matrix.h
@@ -116,21 +116,36 @@ class Matrix : public MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Flags, _MaxRows,
         return m_storage.data()[row + col * m_storage.rows()];
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int row, int col) const
     {
       ei_internal_assert(Flags & VectorizableBit);
       if(Flags & RowMajorBit)
-        return ei_pload(&m_storage.data()[col + row * m_storage.cols()]);
+        if (LoadMode==Aligned)
+          return ei_pload(&m_storage.data()[col + row * m_storage.cols()]);
+        else
+          return ei_ploadu(&m_storage.data()[col + row * m_storage.cols()]);
       else
-        return ei_pload(&m_storage.data()[row + col * m_storage.rows()]);
+        if (LoadMode==Aligned)
+          return ei_pload(&m_storage.data()[row + col * m_storage.rows()]);
+        else
+          return ei_ploadu(&m_storage.data()[row + col * m_storage.rows()]);
     }
+
+    template<int StoreMode>
     void _writePacketCoeff(int row, int col, const PacketScalar& x)
     {
       ei_internal_assert(Flags & VectorizableBit);
       if(Flags & RowMajorBit)
-        ei_pstore(&m_storage.data()[col + row * m_storage.cols()], x);
+        if (StoreMode==Aligned)
+          ei_pstore(&m_storage.data()[col + row * m_storage.cols()], x);
+        else
+          ei_pstoreu(&m_storage.data()[col + row * m_storage.cols()], x);
       else
-        ei_pstore(&m_storage.data()[row + col * m_storage.rows()], x);
+        if (StoreMode==Aligned)
+          ei_pstore(&m_storage.data()[row + col * m_storage.rows()], x);
+        else
+          ei_pstoreu(&m_storage.data()[row + col * m_storage.rows()], x);
     }
 
   public:
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index a60ead6ba..97e8cab28 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -207,8 +207,10 @@ template<typename Derived> class MatrixBase
     Scalar& coeffRef(int index);
     Scalar& operator[](int index);
 
-    PacketScalar packetCoeff(int row, int col) const { return derived()._packetCoeff(row,col); }
-    void writePacketCoeff(int row, int col, const PacketScalar& x) { return derived()._writePacketCoeff(row,col,x); }
+    template<int LoadMode>
+    PacketScalar packetCoeff(int row, int col) const { return derived().template _packetCoeff<LoadMode>(row,col); }
+    template<int StoreMode>
+    void writePacketCoeff(int row, int col, const PacketScalar& x) { return derived().template _writePacketCoeff<StoreMode>(row,col,x); }
 
     const Scalar x() const;
     const Scalar y() const;
@@ -555,7 +557,9 @@ template<typename Derived> class MatrixBase
 
   private:
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int , int) const { ei_internal_assert(false && "_packetCoeff not defined"); }
+    template<int StoreMode>
     void _writePacketCoeff(int , int, const PacketScalar&) { ei_internal_assert(false && "_packetCoeff not defined"); }
 
 };
diff --git a/Eigen/src/Core/PacketMath.h b/Eigen/src/Core/PacketMath.h
index 140dcdde5..cfa19eb6a 100644
--- a/Eigen/src/Core/PacketMath.h
+++ b/Eigen/src/Core/PacketMath.h
@@ -54,29 +54,33 @@ template <typename Scalar> inline Scalar ei_pset1(const Scalar& a) { return a; }
 template <typename Scalar> inline void ei_pstore(Scalar* to, const Scalar& from) { (*to) = from; }
 /** \internal \returns the first element of a packet */
 template <typename Scalar> inline Scalar ei_pfirst(const Scalar& a) { return a; }
+/** \internal \returns a packet where the element i contains the sum of the packet of \a vec[i] */
+template <typename Scalar> inline Scalar ei_predux(const Scalar vecs[1]) { return vecs[0]; }
+/** \internal \returns the sum of the elements of \a a*/
+template <typename Scalar> inline Scalar ei_predux(const Scalar& a) { return a; }
 
 #ifdef EIGEN_VECTORIZE_SSE
 
 #ifdef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
 #undef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
-#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 		16
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 16
 #endif
 
 template<> struct ei_packet_traits<float>  { typedef __m128  type; enum {size=4}; };
 template<> struct ei_packet_traits<double> { typedef __m128d type; enum {size=2}; };
 template<> struct ei_packet_traits<int>    { typedef __m128i type; enum {size=4}; };
 
-inline __m128  ei_padd(const __m128&  a, const __m128&  b) { return _mm_add_ps(a,b); }
-inline __m128d ei_padd(const __m128d& a, const __m128d& b) { return _mm_add_pd(a,b); }
-inline __m128i ei_padd(const __m128i& a, const __m128i& b) { return _mm_add_epi32(a,b); }
+template<> inline __m128  ei_padd(const __m128&  a, const __m128&  b) { return _mm_add_ps(a,b); }
+template<> inline __m128d ei_padd(const __m128d& a, const __m128d& b) { return _mm_add_pd(a,b); }
+template<> inline __m128i ei_padd(const __m128i& a, const __m128i& b) { return _mm_add_epi32(a,b); }
 
-inline __m128  ei_psub(const __m128&  a, const __m128&  b) { return _mm_sub_ps(a,b); }
-inline __m128d ei_psub(const __m128d& a, const __m128d& b) { return _mm_sub_pd(a,b); }
-inline __m128i ei_psub(const __m128i& a, const __m128i& b) { return _mm_sub_epi32(a,b); }
+template<> inline __m128  ei_psub(const __m128&  a, const __m128&  b) { return _mm_sub_ps(a,b); }
+template<> inline __m128d ei_psub(const __m128d& a, const __m128d& b) { return _mm_sub_pd(a,b); }
+template<> inline __m128i ei_psub(const __m128i& a, const __m128i& b) { return _mm_sub_epi32(a,b); }
 
-inline __m128  ei_pmul(const __m128&  a, const __m128&  b) { return _mm_mul_ps(a,b); }
-inline __m128d ei_pmul(const __m128d& a, const __m128d& b) { return _mm_mul_pd(a,b); }
-inline __m128i ei_pmul(const __m128i& a, const __m128i& b)
+template<> inline __m128  ei_pmul(const __m128&  a, const __m128&  b) { return _mm_mul_ps(a,b); }
+template<> inline __m128d ei_pmul(const __m128d& a, const __m128d& b) { return _mm_mul_pd(a,b); }
+template<> inline __m128i ei_pmul(const __m128i& a, const __m128i& b)
 {
   return _mm_or_si128(
     _mm_and_si128(
@@ -89,21 +93,21 @@ inline __m128i ei_pmul(const __m128i& a, const __m128i& b)
 }
 
 // for some weird raisons, it has to be overloaded for packet integer
-inline __m128i ei_pmadd(const __m128i& a, const __m128i& b, const __m128i& c) { return ei_padd(ei_pmul(a,b), c); }
+template<> inline __m128i ei_pmadd(const __m128i& a, const __m128i& b, const __m128i& c) { return ei_padd(ei_pmul(a,b), c); }
 
-inline __m128  ei_pmin(const __m128&  a, const __m128&  b) { return _mm_min_ps(a,b); }
-inline __m128d ei_pmin(const __m128d& a, const __m128d& b) { return _mm_min_pd(a,b); }
+template<> inline __m128  ei_pmin(const __m128&  a, const __m128&  b) { return _mm_min_ps(a,b); }
+template<> inline __m128d ei_pmin(const __m128d& a, const __m128d& b) { return _mm_min_pd(a,b); }
 // FIXME this vectorized min operator is likely to be slower than the standard one
-inline __m128i ei_pmin(const __m128i& a, const __m128i& b)
+template<> inline __m128i ei_pmin(const __m128i& a, const __m128i& b)
 {
   __m128i mask = _mm_cmplt_epi32(a,b);
   return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
 }
 
-inline __m128  ei_pmax(const __m128&  a, const __m128&  b) { return _mm_max_ps(a,b); }
-inline __m128d ei_pmax(const __m128d& a, const __m128d& b) { return _mm_max_pd(a,b); }
+template<> inline __m128  ei_pmax(const __m128&  a, const __m128&  b) { return _mm_max_ps(a,b); }
+template<> inline __m128d ei_pmax(const __m128d& a, const __m128d& b) { return _mm_max_pd(a,b); }
 // FIXME this vectorized max operator is likely to be slower than the standard one
-inline __m128i ei_pmax(const __m128i& a, const __m128i& b)
+template<> inline __m128i ei_pmax(const __m128i& a, const __m128i& b)
 {
   __m128i mask = _mm_cmpgt_epi32(a,b);
   return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
@@ -113,6 +117,10 @@ inline __m128  ei_pload(const float*   from) { return _mm_load_ps(from); }
 inline __m128d ei_pload(const double*  from) { return _mm_load_pd(from); }
 inline __m128i ei_pload(const int* from) { return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); }
 
+inline __m128  ei_ploadu(const float*   from) { return _mm_loadu_ps(from); }
+inline __m128d ei_ploadu(const double*  from) { return _mm_loadu_pd(from); }
+inline __m128i ei_ploadu(const int* from) { return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from)); }
+
 inline __m128  ei_pset1(const float&  from) { return _mm_set1_ps(from); }
 inline __m128d ei_pset1(const double& from) { return _mm_set1_pd(from); }
 inline __m128i ei_pset1(const int&    from) { return _mm_set1_epi32(from); }
@@ -121,15 +129,39 @@ inline void ei_pstore(float*  to, const __m128&  from) { _mm_store_ps(to, from);
 inline void ei_pstore(double* to, const __m128d& from) { _mm_store_pd(to, from); }
 inline void ei_pstore(int*    to, const __m128i& from) { _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
 
+inline void ei_pstoreu(float*  to, const __m128&  from) { _mm_storeu_ps(to, from); }
+inline void ei_pstoreu(double* to, const __m128d& from) { _mm_storeu_pd(to, from); }
+inline void ei_pstoreu(int*    to, const __m128i& from) { _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
+
 inline float  ei_pfirst(const __m128&  a) { return _mm_cvtss_f32(a); }
 inline double ei_pfirst(const __m128d& a) { return _mm_cvtsd_f64(a); }
 inline int    ei_pfirst(const __m128i& a) { return _mm_cvtsi128_si32(a); }
 
+#ifdef __SSE3__
+// TODO implement SSE2 versions as well as integer versions
+inline __m128 ei_predux(const __m128* vecs)
+{
+  return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
+}
+inline __m128d ei_predux(const __m128d* vecs)
+{
+  return _mm_hadd_pd(vecs[0], vecs[1]);
+}
+
+inline float ei_predux(const __m128& a)
+{
+  __m128 tmp0 = _mm_hadd_ps(a,a);
+  return ei_pfirst(_mm_hadd_ps(tmp0, tmp0));
+}
+
+inline double ei_predux(const __m128d& a) { return ei_pfirst(_mm_hadd_pd(a, a)); }
+#endif
+
 #elif defined(EIGEN_VECTORIZE_ALTIVEC)
 
 #ifdef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
 #undef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
-#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 		4
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
 #endif
 
 static const vector int   v0i   = vec_splat_u32(0);
diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h
index 160d437fa..c1b2f5457 100644
--- a/Eigen/src/Core/Product.h
+++ b/Eigen/src/Core/Product.h
@@ -69,7 +69,7 @@ struct ei_packet_product_unroller<true, Index, Size, Lhs, Rhs, PacketScalar>
   static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
   {
     ei_packet_product_unroller<true, Index-1, Size, Lhs, Rhs, PacketScalar>::run(row, col, lhs, rhs, res);
-    res =  ei_pmadd(ei_pset1(lhs.coeff(row, Index)), rhs.packetCoeff(Index, col), res);
+    res =  ei_pmadd(ei_pset1(lhs.coeff(row, Index)), rhs.template packetCoeff<Aligned>(Index, col), res);
   }
 };
 
@@ -79,7 +79,7 @@ struct ei_packet_product_unroller<false, Index, Size, Lhs, Rhs, PacketScalar>
   static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
   {
     ei_packet_product_unroller<false, Index-1, Size, Lhs, Rhs, PacketScalar>::run(row, col, lhs, rhs, res);
-    res =  ei_pmadd(lhs.packetCoeff(row, Index), ei_pset1(rhs.coeff(Index, col)), res);
+    res =  ei_pmadd(lhs.template packetCoeff<Aligned>(row, Index), ei_pset1(rhs.coeff(Index, col)), res);
   }
 };
 
@@ -88,7 +88,7 @@ struct ei_packet_product_unroller<true, 0, Size, Lhs, Rhs, PacketScalar>
 {
   static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
   {
-    res = ei_pmul(ei_pset1(lhs.coeff(row, 0)),rhs.packetCoeff(0, col));
+    res = ei_pmul(ei_pset1(lhs.coeff(row, 0)),rhs.template packetCoeff<Aligned>(0, col));
   }
 };
 
@@ -97,7 +97,7 @@ struct ei_packet_product_unroller<false, 0, Size, Lhs, Rhs, PacketScalar>
 {
   static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
   {
-    res = ei_pmul(lhs.packetCoeff(row, 0), ei_pset1(rhs.coeff(0, col)));
+    res = ei_pmul(lhs.template packetCoeff<Aligned>(row, 0), ei_pset1(rhs.coeff(0, col)));
   }
 };
 
@@ -196,10 +196,10 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
     }
 
     /** \internal */
-    template<typename DestDerived>
+    template<typename DestDerived, int AlignedMode>
     void _cacheOptimalEval(DestDerived& res, ei_meta_false) const;
     #ifdef EIGEN_VECTORIZE
-    template<typename DestDerived>
+    template<typename DestDerived, int AlignedMode>
     void _cacheOptimalEval(DestDerived& res, ei_meta_true) const;
     #endif
 
@@ -228,6 +228,7 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
       return res;
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int row, int col) const
     {
       if(Lhs::ColsAtCompileTime <= EIGEN_UNROLLING_LIMIT)
@@ -247,21 +248,30 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
     PacketScalar _packetCoeffRowMajor(int row, int col) const
     {
       PacketScalar res;
-      res = ei_pmul(ei_pset1(m_lhs.coeff(row, 0)),m_rhs.packetCoeff(0, col));
+      res = ei_pmul(ei_pset1(m_lhs.coeff(row, 0)),m_rhs.template packetCoeff<Aligned>(0, col));
       for(int i = 1; i < m_lhs.cols(); i++)
-        res =  ei_pmadd(ei_pset1(m_lhs.coeff(row, i)), m_rhs.packetCoeff(i, col), res);
+        res =  ei_pmadd(ei_pset1(m_lhs.coeff(row, i)), m_rhs.template packetCoeff<Aligned>(i, col), res);
       return res;
     }
 
     PacketScalar _packetCoeffColumnMajor(int row, int col) const
     {
       PacketScalar res;
-      res = ei_pmul(m_lhs.packetCoeff(row, 0), ei_pset1(m_rhs.coeff(0, col)));
+      res = ei_pmul(m_lhs.template packetCoeff<Aligned>(row, 0), ei_pset1(m_rhs.coeff(0, col)));
       for(int i = 1; i < m_lhs.cols(); i++)
-        res =  ei_pmadd(m_lhs.packetCoeff(row, i), ei_pset1(m_rhs.coeff(i, col)), res);
+        res =  ei_pmadd(m_lhs.template packetCoeff<Aligned>(row, i), ei_pset1(m_rhs.coeff(i, col)), res);
       return res;
+//       const PacketScalar tmp[4];
+//       ei_punpack(m_rhs.packetCoeff(0,col), tmp);
+//
+//       return
+//         ei_pmadd(m_lhs.packetCoeff(row, 0), tmp[0],
+//           ei_pmadd(m_lhs.packetCoeff(row, 1), tmp[1],
+//             ei_pmadd(m_lhs.packetCoeff(row, 2), tmp[2]
+//               ei_pmul(m_lhs.packetCoeff(row, 3), tmp[3]))));
     }
 
+
   protected:
     const LhsNested m_lhs;
     const RhsNested m_rhs;
@@ -298,7 +308,7 @@ template<typename Derived>
 template<typename Lhs, typename Rhs>
 Derived& MatrixBase<Derived>::lazyAssign(const Product<Lhs,Rhs,CacheOptimalProduct>& product)
 {
-  product._cacheOptimalEval(*this,
+  product.template _cacheOptimalEval<Derived, Aligned>(derived(),
     #ifdef EIGEN_VECTORIZE
     typename ei_meta_if<Flags & VectorizableBit, ei_meta_true, ei_meta_false>::ret()
     #else
@@ -309,7 +319,7 @@ Derived& MatrixBase<Derived>::lazyAssign(const Product<Lhs,Rhs,CacheOptimalProdu
 }
 
 template<typename Lhs, typename Rhs, int EvalMode>
-template<typename DestDerived>
+template<typename DestDerived, int AlignedMode>
 void Product<Lhs,Rhs,EvalMode>::_cacheOptimalEval(DestDerived& res, ei_meta_false) const
 {
   res.setZero();
@@ -372,14 +382,14 @@ void Product<Lhs,Rhs,EvalMode>::_cacheOptimalEval(DestDerived& res, ei_meta_fals
 
 #ifdef EIGEN_VECTORIZE
 template<typename Lhs, typename Rhs, int EvalMode>
-template<typename DestDerived>
+template<typename DestDerived, int AlignedMode>
 void Product<Lhs,Rhs,EvalMode>::_cacheOptimalEval(DestDerived& res, ei_meta_true) const
 {
 
   if (((Lhs::Flags&RowMajorBit) && (_cols() % ei_packet_traits<Scalar>::size != 0))
     || (_rows() % ei_packet_traits<Scalar>::size != 0))
   {
-    return _cacheOptimalEval(res, ei_meta_false());
+    return _cacheOptimalEval<DestDerived, AlignedMode>(res, ei_meta_false());
   }
 
   res.setZero();
@@ -398,12 +408,12 @@ void Product<Lhs,Rhs,EvalMode>::_cacheOptimalEval(DestDerived& res, ei_meta_true
         const typename ei_packet_traits<Scalar>::type tmp3 = ei_pset1(m_lhs.coeff(k,j+3));
         for (int i=0; i<this->cols(); i+=ei_packet_traits<Scalar>::size)
         {
-          res.writePacketCoeff(k,i,
-            ei_pmadd(tmp0, m_rhs.packetCoeff(j+0,i),
-              ei_pmadd(tmp1, m_rhs.packetCoeff(j+1,i),
-                ei_pmadd(tmp2, m_rhs.packetCoeff(j+2,i),
-                  ei_pmadd(tmp3, m_rhs.packetCoeff(j+3,i),
-                    res.packetCoeff(k,i)))))
+          res.template writePacketCoeff<AlignedMode>(k,i,
+            ei_pmadd(tmp0, m_rhs.template packetCoeff<AlignedMode>(j+0,i),
+              ei_pmadd(tmp1, m_rhs.template packetCoeff<AlignedMode>(j+1,i),
+                ei_pmadd(tmp2, m_rhs.template packetCoeff<AlignedMode>(j+2,i),
+                  ei_pmadd(tmp3, m_rhs.template packetCoeff<AlignedMode>(j+3,i),
+                    res.template packetCoeff<AlignedMode>(k,i)))))
           );
         }
       }
@@ -414,41 +424,44 @@ void Product<Lhs,Rhs,EvalMode>::_cacheOptimalEval(DestDerived& res, ei_meta_true
       {
         const typename ei_packet_traits<Scalar>::type tmp = ei_pset1(m_lhs.coeff(k,j));
         for (int i=0; i<this->cols(); i+=ei_packet_traits<Scalar>::size)
-          res.writePacketCoeff(k,i, ei_pmadd(tmp, m_rhs.packetCoeff(j,i), res.packetCoeff(k,i)));
+          res.template writePacketCoeff<AlignedMode>(k,i,
+            ei_pmadd(tmp, m_rhs.template packetCoeff<AlignedMode>(j,i), res.template packetCoeff<AlignedMode>(k,i)));
       }
     }
   }
   else
   {
 //     std::cout << "packet lhs\n";
-    int j=0;
-    for(; j<cols4; j+=4)
+    int k=0;
+    for(; k<cols4; k+=4)
     {
-      for(int k=0; k<this->cols(); k++)
+      for(int j=0; j<this->cols(); j+=1)
       {
-        const typename ei_packet_traits<Scalar>::type tmp0 = ei_pset1(m_rhs.coeff(j+0,k));
-        const typename ei_packet_traits<Scalar>::type tmp1 = ei_pset1(m_rhs.coeff(j+1,k));
-        const typename ei_packet_traits<Scalar>::type tmp2 = ei_pset1(m_rhs.coeff(j+2,k));
-        const typename ei_packet_traits<Scalar>::type tmp3 = ei_pset1(m_rhs.coeff(j+3,k));
+        const typename ei_packet_traits<Scalar>::type tmp0 = ei_pset1(m_rhs.coeff(k+0,j));
+        const typename ei_packet_traits<Scalar>::type tmp1 = ei_pset1(m_rhs.coeff(k+1,j));
+        const typename ei_packet_traits<Scalar>::type tmp2 = ei_pset1(m_rhs.coeff(k+2,j));
+        const typename ei_packet_traits<Scalar>::type tmp3 = ei_pset1(m_rhs.coeff(k+3,j));
+
         for (int i=0; i<this->rows(); i+=ei_packet_traits<Scalar>::size)
         {
-          res.writePacketCoeff(i,k,
-            ei_pmadd(tmp0, m_lhs.packetCoeff(i,j),
-              ei_pmadd(tmp1, m_lhs.packetCoeff(i,j+1),
-                ei_pmadd(tmp2, m_lhs.packetCoeff(i,j+2),
-                  ei_pmadd(tmp3, m_lhs.packetCoeff(i,j+3),
-                    res.packetCoeff(i,k)))))
+          res.template writePacketCoeff<AlignedMode>(i,j,
+            ei_pmadd(tmp0, m_lhs.template packetCoeff<AlignedMode>(i,k),
+              ei_pmadd(tmp1, m_lhs.template packetCoeff<AlignedMode>(i,k+1),
+                ei_pmadd(tmp2, m_lhs.template packetCoeff<AlignedMode>(i,k+2),
+                  ei_pmadd(tmp3, m_lhs.template packetCoeff<AlignedMode>(i,k+3),
+                    res.template packetCoeff<AlignedMode>(i,j)))))
           );
         }
       }
     }
-    for(; j<m_lhs.cols(); ++j)
+    for(; k<m_lhs.cols(); ++k)
     {
-      for(int k=0; k<this->cols(); k++)
+      for(int j=0; j<this->cols(); j++)
       {
-        const typename ei_packet_traits<Scalar>::type tmp = ei_pset1(m_rhs.coeff(j,k));
+        const typename ei_packet_traits<Scalar>::type tmp = ei_pset1(m_rhs.coeff(k,j));
         for (int i=0; i<this->rows(); i+=ei_packet_traits<Scalar>::size)
-          res.writePacketCoeff(i,k, ei_pmadd(tmp, m_lhs.packetCoeff(i,j), res.packetCoeff(i,k)));
+          res.template writePacketCoeff<AlignedMode>(k,j,
+            ei_pmadd(tmp, m_lhs.template packetCoeff<AlignedMode>(i,k), res.template packetCoeff<AlignedMode>(i,j)));
       }
     }
   }
diff --git a/Eigen/src/Core/ProductWIP.h b/Eigen/src/Core/ProductWIP.h
new file mode 100644
index 000000000..a1c10d5d8
--- /dev/null
+++ b/Eigen/src/Core/ProductWIP.h
@@ -0,0 +1,496 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra. Eigen itself is part of the KDE project.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob@math.jussieu.fr>
+// Copyright (C) 2008 Gael Guennebaud <g.gael@free.fr>
+//
+// Eigen is free software; you can redistribute it and/or
+// modify it under the terms of the GNU Lesser General Public
+// License as published by the Free Software Foundation; either
+// version 3 of the License, or (at your option) any later version.
+//
+// Alternatively, you can redistribute it and/or
+// modify it under the terms of the GNU General Public License as
+// published by the Free Software Foundation; either version 2 of
+// the License, or (at your option) any later version.
+//
+// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY
+// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public
+// License and a copy of the GNU General Public License along with
+// Eigen. If not, see <http://www.gnu.org/licenses/>.
+
+#ifndef EIGEN_PRODUCT_H
+#define EIGEN_PRODUCT_H
+
+#ifndef EIGEN_VECTORIZE
+#error you must enable vectorization to try this experimental product implementation
+#endif
+
+template<int Index, int Size, typename Lhs, typename Rhs>
+struct ei_product_unroller
+{
+  static void run(int row, int col, const Lhs& lhs, const Rhs& rhs,
+                  typename Lhs::Scalar &res)
+  {
+    ei_product_unroller<Index-1, Size, Lhs, Rhs>::run(row, col, lhs, rhs, res);
+    res += lhs.coeff(row, Index) * rhs.coeff(Index, col);
+  }
+};
+
+template<int Size, typename Lhs, typename Rhs>
+struct ei_product_unroller<0, Size, Lhs, Rhs>
+{
+  static void run(int row, int col, const Lhs& lhs, const Rhs& rhs,
+                  typename Lhs::Scalar &res)
+  {
+    res = lhs.coeff(row, 0) * rhs.coeff(0, col);
+  }
+};
+
+template<int Index, typename Lhs, typename Rhs>
+struct ei_product_unroller<Index, Dynamic, Lhs, Rhs>
+{
+  static void run(int, int, const Lhs&, const Rhs&, typename Lhs::Scalar&) {}
+};
+
+// prevent buggy user code from causing an infinite recursion
+template<int Index, typename Lhs, typename Rhs>
+struct ei_product_unroller<Index, 0, Lhs, Rhs>
+{
+  static void run(int, int, const Lhs&, const Rhs&, typename Lhs::Scalar&) {}
+};
+
+template<bool RowMajor, int Index, int Size, typename Lhs, typename Rhs, typename PacketScalar>
+struct ei_packet_product_unroller;
+
+template<int Index, int Size, typename Lhs, typename Rhs, typename PacketScalar>
+struct ei_packet_product_unroller<true, Index, Size, Lhs, Rhs, PacketScalar>
+{
+  static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
+  {
+    ei_packet_product_unroller<true, Index-1, Size, Lhs, Rhs, PacketScalar>::run(row, col, lhs, rhs, res);
+    res =  ei_pmadd(ei_pset1(lhs.coeff(row, Index)), rhs.packetCoeff(Index, col), res);
+  }
+};
+
+template<int Index, int Size, typename Lhs, typename Rhs, typename PacketScalar>
+struct ei_packet_product_unroller<false, Index, Size, Lhs, Rhs, PacketScalar>
+{
+  static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
+  {
+    ei_packet_product_unroller<false, Index-1, Size, Lhs, Rhs, PacketScalar>::run(row, col, lhs, rhs, res);
+    res =  ei_pmadd(lhs.packetCoeff(row, Index), ei_pset1(rhs.coeff(Index, col)), res);
+  }
+};
+
+template<int Size, typename Lhs, typename Rhs, typename PacketScalar>
+struct ei_packet_product_unroller<true, 0, Size, Lhs, Rhs, PacketScalar>
+{
+  static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
+  {
+    res = ei_pmul(ei_pset1(lhs.coeff(row, 0)),rhs.packetCoeff(0, col));
+  }
+};
+
+template<int Size, typename Lhs, typename Rhs, typename PacketScalar>
+struct ei_packet_product_unroller<false, 0, Size, Lhs, Rhs, PacketScalar>
+{
+  static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, PacketScalar &res)
+  {
+    res = ei_pmul(lhs.packetCoeff(row, 0), ei_pset1(rhs.coeff(0, col)));
+  }
+};
+
+template<bool RowMajor, int Index, typename Lhs, typename Rhs, typename PacketScalar>
+struct ei_packet_product_unroller<RowMajor, Index, Dynamic, Lhs, Rhs, PacketScalar>
+{
+  static void run(int, int, const Lhs&, const Rhs&, PacketScalar&) {}
+};
+
+template<typename Product, bool RowMajor = true> struct ProductPacketCoeffImpl {
+  inline static typename Product::PacketScalar execute(const Product& product, int row, int col)
+  { return product._packetCoeffRowMajor(row,col); }
+};
+
+template<typename Product> struct ProductPacketCoeffImpl<Product, false> {
+  inline static typename Product::PacketScalar execute(const Product& product, int row, int col)
+  { return product._packetCoeffColumnMajor(row,col); }
+};
+
+/** \class Product
+  *
+  * \brief Expression of the product of two matrices
+  *
+  * \param Lhs the type of the left-hand side
+  * \param Rhs the type of the right-hand side
+  * \param EvalMode internal use only
+  *
+  * This class represents an expression of the product of two matrices.
+  * It is the return type of the operator* between matrices, and most of the time
+  * this is the only way it is used.
+  *
+  * \sa class Sum, class Difference
+  */
+template<typename Lhs, typename Rhs> struct ei_product_eval_mode
+{
+  enum{ value =  Lhs::MaxRowsAtCompileTime >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+              && Rhs::MaxColsAtCompileTime >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+              && (!( (Lhs::Flags&RowMajorBit) && ((Rhs::Flags&RowMajorBit) ^ RowMajorBit)))
+              ? CacheOptimalProduct : NormalProduct };
+};
+
+template<typename Lhs, typename Rhs, int EvalMode>
+struct ei_traits<Product<Lhs, Rhs, EvalMode> >
+{
+  typedef typename Lhs::Scalar Scalar;
+  typedef typename ei_nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
+  typedef typename ei_nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
+  typedef typename ei_unref<LhsNested>::type _LhsNested;
+  typedef typename ei_unref<RhsNested>::type _RhsNested;
+  enum {
+    LhsCoeffReadCost = _LhsNested::CoeffReadCost,
+    RhsCoeffReadCost = _RhsNested::CoeffReadCost,
+    LhsFlags = _LhsNested::Flags,
+    RhsFlags = _RhsNested::Flags,
+    RowsAtCompileTime = Lhs::RowsAtCompileTime,
+    ColsAtCompileTime = Rhs::ColsAtCompileTime,
+    MaxRowsAtCompileTime = Lhs::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = Rhs::MaxColsAtCompileTime,
+    _RhsVectorizable = (RhsFlags & RowMajorBit) && (RhsFlags & VectorizableBit) && (ColsAtCompileTime % ei_packet_traits<Scalar>::size == 0),
+    _LhsVectorizable = (!(LhsFlags & RowMajorBit)) && (LhsFlags & VectorizableBit) && (RowsAtCompileTime % ei_packet_traits<Scalar>::size == 0),
+    _Vectorizable = (_LhsVectorizable || _RhsVectorizable) ? 1 : 0,
+    _RowMajor = (RhsFlags & RowMajorBit)
+              && (EvalMode==(int)CacheOptimalProduct ? (int)LhsFlags & RowMajorBit : (!_LhsVectorizable)),
+    _LostBits = DefaultLostFlagMask & ~(
+                (_RowMajor ? 0 : RowMajorBit)
+              | ((RowsAtCompileTime == Dynamic || ColsAtCompileTime == Dynamic) ? 0 : LargeBit)),
+    Flags = ((unsigned int)(LhsFlags | RhsFlags) & _LostBits)
+//          | EvalBeforeAssigningBit //FIXME
+          | EvalBeforeNestingBit
+          | (_Vectorizable ? VectorizableBit : 0),
+    CoeffReadCost
+      = Lhs::ColsAtCompileTime == Dynamic
+      ? Dynamic
+      : Lhs::ColsAtCompileTime
+        * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost)
+        + (Lhs::ColsAtCompileTime - 1) * NumTraits<Scalar>::AddCost
+  };
+};
+
+template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignment_operator,
+  public MatrixBase<Product<Lhs, Rhs, EvalMode> >
+{
+  public:
+
+    EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
+    friend class ProductPacketCoeffImpl<Product,Flags&RowMajorBit>;
+    typedef typename ei_traits<Product>::LhsNested LhsNested;
+    typedef typename ei_traits<Product>::RhsNested RhsNested;
+    typedef typename ei_traits<Product>::_LhsNested _LhsNested;
+    typedef typename ei_traits<Product>::_RhsNested _RhsNested;
+
+    Product(const Lhs& lhs, const Rhs& rhs)
+      : m_lhs(lhs), m_rhs(rhs)
+    {
+      ei_assert(lhs.cols() == rhs.rows());
+    }
+
+    /** \internal */
+    template<typename DestDerived> void _cacheFriendlyEval(DestDerived& res) const;
+
+  private:
+
+    int _rows() const { return m_lhs.rows(); }
+    int _cols() const { return m_rhs.cols(); }
+
+    const Scalar _coeff(int row, int col) const
+    {
+      Scalar res;
+      const bool unroll = CoeffReadCost <= EIGEN_UNROLLING_LIMIT;
+      if(unroll)
+      {
+        ei_product_unroller<Lhs::ColsAtCompileTime-1,
+                            unroll ? Lhs::ColsAtCompileTime : Dynamic,
+                            _LhsNested, _RhsNested>
+          ::run(row, col, m_lhs, m_rhs, res);
+      }
+      else
+      {
+        res = m_lhs.coeff(row, 0) * m_rhs.coeff(0, col);
+        for(int i = 1; i < m_lhs.cols(); i++)
+          res += m_lhs.coeff(row, i) * m_rhs.coeff(i, col);
+      }
+      return res;
+    }
+
+    template<int LoadMode>
+    PacketScalar _packetCoeff(int row, int col) const
+    {
+      if(Lhs::ColsAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+      {
+        PacketScalar res;
+        ei_packet_product_unroller<Flags&RowMajorBit, Lhs::ColsAtCompileTime-1,
+                            Lhs::ColsAtCompileTime <= EIGEN_UNROLLING_LIMIT
+                              ? Lhs::ColsAtCompileTime : Dynamic,
+                            _LhsNested, _RhsNested, PacketScalar>
+          ::run(row, col, m_lhs, m_rhs, res);
+        return res;
+      }
+      else
+        return ProductPacketCoeffImpl<Product,Flags&RowMajorBit>::execute(*this, row, col);
+    }
+
+    PacketScalar _packetCoeffRowMajor(int row, int col) const
+    {
+      PacketScalar res;
+      res = ei_pmul(ei_pset1(m_lhs.coeff(row, 0)),m_rhs.packetCoeff(0, col));
+      for(int i = 1; i < m_lhs.cols(); i++)
+        res =  ei_pmadd(ei_pset1(m_lhs.coeff(row, i)), m_rhs.packetCoeff(i, col), res);
+      return res;
+    }
+
+    PacketScalar _packetCoeffColumnMajor(int row, int col) const
+    {
+      PacketScalar res;
+      res = ei_pmul(m_lhs.packetCoeff(row, 0), ei_pset1(m_rhs.coeff(0, col)));
+      for(int i = 1; i < m_lhs.cols(); i++)
+        res =  ei_pmadd(m_lhs.packetCoeff(row, i), ei_pset1(m_rhs.coeff(i, col)), res);
+      return res;
+    }
+
+
+  protected:
+    const LhsNested m_lhs;
+    const RhsNested m_rhs;
+};
+
+/** \returns the matrix product of \c *this and \a other.
+  *
+  * \note This function causes an immediate evaluation. If you want to perform a matrix product
+  * without immediate evaluation, call .lazy() on one of the matrices before taking the product.
+  *
+  * \sa lazy(), operator*=(const MatrixBase&)
+  */
+template<typename Derived>
+template<typename OtherDerived>
+const Product<Derived,OtherDerived>
+MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
+{
+  return Product<Derived,OtherDerived>(derived(), other.derived());
+}
+
+/** replaces \c *this by \c *this * \a other.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+Derived &
+MatrixBase<Derived>::operator*=(const MatrixBase<OtherDerived> &other)
+{
+  return *this = *this * other;
+}
+
+template<typename Derived>
+template<typename Lhs, typename Rhs>
+Derived& MatrixBase<Derived>::lazyAssign(const Product<Lhs,Rhs,CacheOptimalProduct>& product)
+{
+  product._cacheFriendlyEval(*this);
+  return derived();
+}
+
+template<typename Lhs, typename Rhs, int EvalMode>
+template<typename DestDerived>
+void Product<Lhs,Rhs,EvalMode>::_cacheFriendlyEval(DestDerived& res) const
+{
+  // allow direct access to data for benchmark purpose
+  const Scalar* __restrict__ a = m_lhs.derived().data();
+  const Scalar* __restrict__ b = m_rhs.derived().data();
+  Scalar* __restrict__ c = res.derived().data();
+
+  // FIXME find a way to optimize: (an_xpr) + (a * b)
+  // then we don't need to clear res and avoid and additional mat-mat sum
+//   res.setZero();
+
+  const int ps = ei_packet_traits<Scalar>::size;  // size of a packet
+  #if (defined __i386__)
+  // i386 architectures provides only 8 xmmm register,
+  // so let's reduce the max number of rows processed at once
+  const int bw = 4;                               // number of rows treated at once
+  #else
+  const int bw = 8;                               // number of rows treated at once
+  #endif
+  const int bs = ps * bw;                         // total number of elements treated at once
+  const int rows = _rows();
+  const int cols = _cols();
+  const int size = m_lhs.cols();                  // third dimension of the product
+  const int l2blocksize = 256 > _cols() ? _cols() : 256;
+  const bool rhsIsAligned = ((size%ps) == 0);
+  const bool resIsAligned = ((cols%ps) == 0);
+  Scalar* __restrict__ block = new Scalar[l2blocksize*size];
+
+  // loops on each L2 cache friendly blocks of the result
+  for(int l2i=0; l2i<_rows(); l2i+=l2blocksize)
+  {
+    const int l2blockRowEnd = std::min(l2i+l2blocksize, rows);
+    const int l2blockRowEndBW = l2blockRowEnd & 0xFFFFF8;             // end of the rows aligned to bw
+    const int l2blockRowRemaining = l2blockRowEnd - l2blockRowEndBW;  // number of remaining rows
+
+    // build a cache friendly block
+    int count = 0;
+
+    // copy l2blocksize rows of m_lhs to blocks of ps x bw
+    for(int l2k=0; l2k<size; l2k+=l2blocksize)
+    {
+      const int l2blockSizeEnd = std::min(l2k+l2blocksize, size);
+
+      for (int i = l2i; i<l2blockRowEndBW; i+=bw)
+      {
+        for (int k=l2k; k<l2blockSizeEnd; k+=ps)
+        {
+          // TODO write these two loops using meta unrolling
+          // negligible for large matrices but useful for small ones
+          for (int w=0; w<bw; ++w)
+            for (int s=0; s<ps; ++s)
+              block[count++] = m_lhs.coeff(i+w,k+s);
+        }
+      }
+      if (l2blockRowRemaining>0)
+      {
+        for (int k=l2k; k<l2blockSizeEnd; k+=ps)
+        {
+          for (int w=0; w<l2blockRowRemaining; ++w)
+            for (int s=0; s<ps; ++s)
+              block[count++] = m_lhs.coeff(l2blockRowEndBW+w,k+s);
+        }
+      }
+    }
+
+    for(int l2j=0; l2j<cols; l2j+=l2blocksize)
+    {
+      int l2blockColEnd = std::min(l2j+l2blocksize, cols);
+
+      for(int l2k=0; l2k<size; l2k+=l2blocksize)
+      {
+        // acumulate a full row of current a block time 4 cols of current a block
+        // to a 1x4 c block
+        int l2blockSizeEnd = std::min(l2k+l2blocksize, size);
+
+        // for each 4x1 result's block sub blocks...
+        for(int l1i=l2i; l1i<l2blockRowEndBW; l1i+=bw)
+        {
+          for(int l1j=l2j; l1j<l2blockColEnd; l1j+=1)
+          {
+            int offsetblock = l2k * (l2blockRowEnd-l2i) + (l1i-l2i)*(l2blockSizeEnd-l2k) - l2k*bw/*bs*/;
+            const Scalar* localB = &block[offsetblock];
+
+            int l1jsize = l1j * size; //TODO find a better way to optimize address computation ?
+
+            PacketScalar dst[bw];
+            dst[0] = ei_pset1(Scalar(0.));
+            dst[1] = dst[0];
+            dst[2] = dst[0];
+            dst[3] = dst[0];
+            if (bw==8)
+            {
+              dst[4] = dst[0];
+              dst[5] = dst[0];
+              dst[6] = dst[0];
+              dst[7] = dst[0];
+            }
+            PacketScalar b0, b1, tmp;
+            // TODO in unaligned mode, preload the next element
+//             PacketScalar tmp1 = _mm_load_ps(&m_rhs.derived().data()[l1jsize+l2k]);
+            asm("#eigen begincore");
+            for(int k=l2k; k<l2blockSizeEnd; k+=ps)
+            {
+              //PacketScalar tmp = m_rhs.packetCoeff(k, l1j);
+              if (rhsIsAligned)
+                tmp = ei_pload(&m_rhs.derived().data()[l1jsize + k]);
+              else
+                tmp = ei_ploadu(&m_rhs.derived().data()[l1jsize + k]);
+
+              b0 = ei_pload(&(localB[k*bw]));
+              b1 = ei_pload(&(localB[k*bw+ps]));
+              dst[0] = ei_pmadd(tmp, b0, dst[0]);
+              b0 = ei_pload(&(localB[k*bw+2*ps]));
+              dst[1] = ei_pmadd(tmp, b1, dst[1]);
+              b1 = ei_pload(&(localB[k*bw+3*ps]));
+              dst[2] = ei_pmadd(tmp, b0, dst[2]);
+              if (bw==8)
+                b0 = ei_pload(&(localB[k*bw+4*ps]));
+              dst[3] = ei_pmadd(tmp, b1, dst[3]);
+              if (bw==8)
+              {
+                b1 = ei_pload(&(localB[k*bw+5*ps]));
+                dst[4] = ei_pmadd(tmp, b0, dst[4]);
+                b0 = ei_pload(&(localB[k*bw+6*ps]));
+                dst[5] = ei_pmadd(tmp, b1, dst[5]);
+                b1 = ei_pload(&(localB[k*bw+7*ps]));
+                dst[6] = ei_pmadd(tmp, b0, dst[6]);
+                dst[7] = ei_pmadd(tmp, b1, dst[7]);
+              }
+            }
+
+            res.template writePacketCoeff<Aligned>(l1i, l1j, ei_padd(res.template packetCoeff<Aligned>(l1i, l1j), ei_predux(dst)));
+            if (ps==2)
+              res.template writePacketCoeff<Aligned>(l1i+2,l1j, ei_padd(res.template packetCoeff<Aligned>(l1i+2,l1j), ei_predux(&(dst[2]))));
+            if (bw==8)
+            {
+              res.template writePacketCoeff<Aligned>(l1i+4,l1j, ei_padd(res.template packetCoeff<Aligned>(l1i+4,l1j), ei_predux(&(dst[4]))));
+              if (ps==2)
+                res.template writePacketCoeff<Aligned>(l1i+6,l1j, ei_padd(res.template packetCoeff<Aligned>(l1i+6,l1j), ei_predux(&(dst[6]))));
+            }
+
+            asm("#eigen endcore");
+          }
+        }
+        if (l2blockRowRemaining>0)
+        {
+          // TODO optimize this part using a generic templated function that processes N rows
+          // here we process the remaining l2blockRowRemaining rows
+          for(int l1j=l2j; l1j<l2blockColEnd; l1j+=1)
+          {
+            int offsetblock = l2k * (l2blockRowEnd-l2i) + (l2blockRowEndBW-l2i)*(l2blockSizeEnd-l2k) - l2k*l2blockRowRemaining;
+            const Scalar* localB = &block[offsetblock];
+
+            int l1jsize = l1j * size;
+
+            PacketScalar dst[bw];
+            dst[0] = ei_pset1(Scalar(0.));
+            for (int w = 1; w<l2blockRowRemaining; ++w)
+              dst[w] = dst[0];
+            PacketScalar b0, b1, tmp;
+            asm("#eigen begincore dynamic");
+            for(int k=l2k; k<l2blockSizeEnd; k+=ps)
+            {
+              //PacketScalar tmp = m_rhs.packetCoeff(k, l1j);
+              if (rhsIsAligned)
+                tmp = ei_pload(&m_rhs.derived().data()[l1jsize + k]);
+              else
+                tmp = ei_ploadu(&m_rhs.derived().data()[l1jsize + k]);
+
+              // TODO optimize this loop
+              for (int w = 0; w<l2blockRowRemaining; ++w)
+                dst[w] = ei_pmadd(tmp, ei_pload(&(localB[k*l2blockRowRemaining+w*ps])), dst[w]);
+            }
+
+            // TODO optimize this loop
+            for (int w = 0; w<l2blockRowRemaining; ++w)
+              res.coeffRef(l2blockRowEndBW+w, l1j) += ei_predux(dst[w]);
+
+            asm("#eigen endcore dynamic");
+          }
+        }
+      }
+    }
+  }
+
+  delete[] block;
+}
+
+#endif // EIGEN_PRODUCT_H
diff --git a/Eigen/src/Core/Temporary.h b/Eigen/src/Core/Temporary.h
index 9157b10e4..854bb4698 100644
--- a/Eigen/src/Core/Temporary.h
+++ b/Eigen/src/Core/Temporary.h
@@ -71,9 +71,10 @@ template<typename ExpressionType> class Temporary
       return m_expression.coeff(row, col);
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int row, int col) const
     {
-      return m_expression.packetCoeff(row, col);
+      return m_expression.template packetCoeff<LoadMode>(row, col);
     }
 
   protected:
diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h
index c2ca4a63d..e4af78c7c 100644
--- a/Eigen/src/Core/Transpose.h
+++ b/Eigen/src/Core/Transpose.h
@@ -79,14 +79,16 @@ template<typename MatrixType> class Transpose
       return m_matrix.coeff(col, row);
     }
 
+    template<int LoadMode>
     PacketScalar _packetCoeff(int row, int col) const
     {
-      return m_matrix.packetCoeff(col, row);
+      return m_matrix.template packetCoeff<LoadMode>(col, row);
     }
 
+    template<int LoadMode>
     void _writePacketCoeff(int row, int col, const PacketScalar& x)
     {
-      m_matrix.const_cast_derived().writePacketCoeff(col, row, x);
+      m_matrix.const_cast_derived().template writePacketCoeff<LoadMode>(col, row, x);
     }
 
   protected:
diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h
index ae0451156..c74d8f87b 100644
--- a/Eigen/src/Core/util/Constants.h
+++ b/Eigen/src/Core/util/Constants.h
@@ -45,6 +45,7 @@ const unsigned int NullLowerBit = 0x200;    ///< means the strictly triangular l
 const unsigned int NullUpperBit = 0x400;    ///< means the strictly triangular upper part is 0
 
 enum { Upper=NullLowerBit, Lower=NullUpperBit };
+enum { Aligned=0, UnAligned=1 };
 
 // list of flags that are lost by default
 const unsigned int DefaultLostFlagMask = ~(VectorizableBit | Like1DArrayBit | NullDiagBit | UnitDiagBit | NullLowerBit | NullUpperBit);