diff --git a/libs/libarchfpga/src/physical_types.h b/libs/libarchfpga/src/physical_types.h
index ec42adfe4d3..387f5004fbd 100644
--- a/libs/libarchfpga/src/physical_types.h
+++ b/libs/libarchfpga/src/physical_types.h
@@ -1225,6 +1225,7 @@ struct t_pin_to_pin_annotation {
  *      parent_pb_graph_node  : parent pb graph node
  *      total_primitive_count : Total number of this primitive type in the cluster. If there are 10 ALMs per cluster
  *                              and 2 FFs per ALM (given the mode of the parent of this primitive) then the total is 20.
+ *      flat_site_index       : index of this primitive within its primitive type in this cluster; in [0,...,total_primitive_count]
  *      illegal_modes         : vector containing illegal modes that result in conflicts during routing
  */
 class t_pb_graph_node {
@@ -1281,6 +1282,7 @@ class t_pb_graph_node {
     int num_output_pin_class;   /* number of output pin classes that this pb_graph_node has */
 
     int total_primitive_count; /* total number of this primitive type in the cluster */
+    int flat_site_index;       /* index of this primitive within sites of its type in this cluster  */
 
     /* Interconnect instances for this pb
      * Only used for power
diff --git a/vpr/src/pack/pb_type_graph.cpp b/vpr/src/pack/pb_type_graph.cpp
index 72467921f21..7c51b00ee93 100644
--- a/vpr/src/pack/pb_type_graph.cpp
+++ b/vpr/src/pack/pb_type_graph.cpp
@@ -48,6 +48,7 @@ static void alloc_and_load_pb_graph(t_pb_graph_node* pb_graph_node,
                                     t_pb_graph_node* parent_pb_graph_node,
                                     t_pb_type* pb_type,
                                     const int index,
+                                    const int flat_index,
                                     bool load_power_structures,
                                     int& pin_count_in_cluster);
 
@@ -151,6 +152,7 @@ void alloc_and_load_all_pb_graphs(bool load_power_structures, bool is_flat) {
                                     nullptr,
                                     type.pb_type,
                                     0,
+                                    0,
                                     load_power_structures,
                                     pin_count_in_cluster);
             type.pb_graph_head->total_pb_pins = pin_count_in_cluster;
@@ -224,6 +226,7 @@ static void alloc_and_load_pb_graph(t_pb_graph_node* pb_graph_node,
                                     t_pb_graph_node* parent_pb_graph_node,
                                     t_pb_type* pb_type,
                                     const int index,
+                                    const int flat_index,
                                     bool load_power_structures,
                                     int& pin_count_in_cluster) {
     int i, j, k, i_input, i_output, i_clockport;
@@ -237,6 +240,7 @@ static void alloc_and_load_pb_graph(t_pb_graph_node* pb_graph_node,
     pb_graph_node->num_clock_ports = 0;
 
     pb_graph_node->total_primitive_count = 0;
+    pb_graph_node->flat_site_index = flat_index;
 
     /* Generate ports for pb graph node */
     for (i = 0; i < pb_type->num_ports; i++) {
@@ -349,10 +353,13 @@ static void alloc_and_load_pb_graph(t_pb_graph_node* pb_graph_node,
                                                                                 sizeof(t_pb_graph_node*));
         for (j = 0; j < pb_type->modes[i].num_pb_type_children; j++) {
             pb_graph_node->child_pb_graph_nodes[i][j] = (t_pb_graph_node*)vtr::calloc(pb_type->modes[i].pb_type_children[j].num_pb, sizeof(t_pb_graph_node));
+            int base = flat_index*(pb_type->modes[i].pb_type_children[j].num_pb);
+
             for (k = 0; k < pb_type->modes[i].pb_type_children[j].num_pb; k++) {
                 alloc_and_load_pb_graph(&pb_graph_node->child_pb_graph_nodes[i][j][k],
                                         pb_graph_node,
                                         &pb_type->modes[i].pb_type_children[j],
+                                        base + k,
                                         k,
                                         load_power_structures,
                                         pin_count_in_cluster);
diff --git a/vpr/src/pack/re_cluster_util.h b/vpr/src/pack/re_cluster_util.h
index 4aca0fa5f95..45026f555ba 100644
--- a/vpr/src/pack/re_cluster_util.h
+++ b/vpr/src/pack/re_cluster_util.h
@@ -89,7 +89,7 @@ bool start_new_cluster_for_mol(t_pack_molecule* molecule,
                                t_lb_router_data** router_data,
                                PartitionRegion& temp_cluster_pr,
                                NocGroupId& temp_cluster_noc_grp_id,
-                               const int detailed_routing_stage = 1);
+                               const int& detailed_routing_stage = 1);
 
 /**
  * @brief A function that packs a molecule into an existing cluster
@@ -109,7 +109,7 @@ bool pack_mol_in_existing_cluster(t_pack_molecule* molecule,
                                   bool during_packing,
                                   t_clustering_data& clustering_data,
                                   t_lb_router_data*& router_data,
-                                  const int detailed_routing_stage = 1);
+                                  const int& detailed_routing_stage = 1);
 
 /**
  * @brief A function that fix the clustered netlist if the move is performed