added proper mult_add to gmm++

Former-commit-id: 03a4f13a47

resources/3rdparty/gmm-5.0/include/gmm/gmm_blas.h

@@ -1858,6 +1858,25 @@ namespace gmm {
 		linalg_traits<L1>::sub_orientation>::potype());
     }
   }
+    
+  /** Multiply-accumulate. l4 = l1*l2 + l3; */
+  template <typename L1, typename L2, typename L3, typename L4> inline
+  void mult_add(const L1& l1, const L2& l2, const L3& l3, L4& l4) {
+      size_type m = mat_nrows(l1), n = mat_ncols(l1);
+      if (!m || !n) return;
+      GMM_ASSERT2(n==vect_size(l2) && m==vect_size(l3) && vect_size(l3) == vect_size(l4), "dimensions mismatch");
+      if (!same_origin(l2, l3)) {
+          mult_add_spec(l1, l2, l3, l4, typename principal_orientation_type<typename
+                        linalg_traits<L1>::sub_orientation>::potype());
+      }
+      else {
+          GMM_WARNING2("Warning, A temporary is used for mult\n");
+          typename temporary_vector<L3>::vector_type temp(vect_size(l2));
+          copy(l2, temp);
+          mult_add_spec(l1, temp, l3, l4, typename principal_orientation_type<typename
+                        linalg_traits<L1>::sub_orientation>::potype());
+      }
+  }
   ///@cond DOXY_SHOW_ALL_FUNCTIONS
 
   template <typename L1, typename L2, typename L3> inline
@@ -1893,6 +1912,16 @@ namespace gmm {
       *it += vect_sp(linalg_traits<L1>::row(itr), l2);
   }
 
+  template <typename L1, typename L2, typename L3, typename L4>
+  void mult_add_by_row(const L1& l1, const L2& l2, const L3& l3, L4& l4, abstract_dense) {
+      typename linalg_traits<L3>::const_iterator add_it=vect_begin(l3), add_ite=vect_end(l3);
+      typename linalg_traits<L4>::iterator target_it=vect_begin(l4), target_ite=vect_end(l4);
+      typename linalg_traits<L1>::const_row_iterator
+      itr = mat_row_const_begin(l1);
+      for (; add_it != add_ite; ++add_it, ++target_it, ++itr)
+          *target_it = vect_sp(linalg_traits<L1>::row(itr), l2) + *add_it;
+  }
+
   template <typename L1, typename L2, typename L3>
   void mult_add_by_col(const L1& l1, const L2& l2, L3& l3, abstract_dense) {
     size_type nc = mat_ncols(l1);
@@ -1922,6 +1951,10 @@ namespace gmm {
   void mult_add_spec(const L1& l1, const L2& l2, L3& l3, row_major)
   { mult_add_by_row(l1, l2, l3, typename linalg_traits<L3>::storage_type()); }
 
+  template <typename L1, typename L2, typename L3, typename L4> inline
+  void mult_add_spec(const L1& l1, const L2& l2, const L3& l3, L4& l4, row_major)
+  { mult_add_by_row(l1, l2, l3, l4, typename linalg_traits<L3>::storage_type()); }
+    
   template <typename L1, typename L2, typename L3> inline
   void mult_add_spec(const L1& l1, const L2& l2, L3& l3, col_major)
   { mult_add_by_col(l1, l2, l3, typename linalg_traits<L2>::storage_type()); }