fuse Assign and fix optimizeAssign

mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_helper.cc

@@ -583,7 +583,7 @@ std::vector<PrimitivePtr> GetFusibleOpList() {
     prim::kPrimExpandDims, prim::kPrimMul,        prim::kPrimMinimum, prim::kPrimMaximum, prim::kPrimLog,
     prim::kPrimPow,        prim::kPrimSub,        prim::kPrimRsqrt,   prim::kPrimSqrt,    prim::kPrimAddN,
     prim::kPrimEqual,      prim::kPrimReciprocal, prim::kPrimTanh,    prim::kPrimReshape, prim::kPrimTranspose,
-    prim::kPrimCast,       prim::kPrimRealDiv,    prim::kPrimMatMul};
+    prim::kPrimCast,       prim::kPrimRealDiv,    prim::kPrimMatMul,  prim::kPrimAssign};
 #elif ENABLE_GPU
   std::vector<PrimitivePtr> fusible_basic_ops = {
     prim::kPrimAbs,     prim::kPrimRound,      prim::kPrimNeg,       prim::kPrimExp,     prim::kPrimAdd,

mindspore/ccsrc/backend/optimizer/graph_kernel/optimize_assign.cc

@@ -146,7 +146,7 @@ void UpdateUsersOfGraphKernel(const FuncGraphPtr &func_graph, const AnfNodePtr &
       //   input, without Depend node.
       // 2. If the `cnode` has another path to the getitem_user, it's unnecessary to add update_state and load node to
       // keep exec_order.
-      if (!AnfAlgo::IsRealKernel(getitem_user) || HasPathToParamUser(cnode, getitem_user, getitem)) {
+      if (HasPathToParamUser(cnode, getitem_user, getitem)) {
         mng->Replace(getitem, assign_to);
         continue;
       }