finally directly calling the low-level products is faster

Eigen/src/Core/util/Macros.h

@@ -94,6 +94,13 @@
 #define EIGEN_TUNE_FOR_CPU_CACHE_SIZE (sizeof(float)*256*256)
 #endif
 
+/** Defines the maximal width of the blocks used in the triangular solver
+  * for vectors (level 2 blas xTRSV). The default is 8.
+  */
+#ifndef EIGEN_TUNE_TRSV_PANEL_WIDTH
+#define EIGEN_TUNE_TRSV_PANEL_WIDTH 8
+#endif
+
 /** Allows to disable some optimizations which might affect the accuracy of the result.
   * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them.
   * They currently include:

Eigen/src/Core/util/Meta.h

@@ -198,4 +198,16 @@ template<typename T> struct ei_is_diagonal<DiagonalWrapper<T> >
 template<typename T, int S> struct ei_is_diagonal<DiagonalMatrix<T,S> >
 { enum { ret = true }; };
 
+template<bool Conjugate> struct ei_conj_if;
+
+template<> struct ei_conj_if<true> {
+  template<typename T>
+  inline T operator()(const T& x) { return ei_conj(x); }
+};
+
+template<> struct ei_conj_if<false> {
+  template<typename T>
+  inline const T& operator()(const T& x) { return x; }
+};
+
 #endif // EIGEN_META_H