diff --git a/dory/Hardware_targets/PULP/Common/Templates/main.c.t b/dory/Hardware_targets/PULP/Common/Templates/main.c.t
index 4885c25f..ece05b4d 100644
--- a/dory/Hardware_targets/PULP/Common/Templates/main.c.t
+++ b/dory/Hardware_targets/PULP/Common/Templates/main.c.t
@@ -64,6 +64,7 @@ void application(void * arg) {
 #endif
   size_t l2_input_size = ${int(DORY_HW_graph[0].tiling_dimensions["L2"]["input_activation_memory"])};
   size_t input_size = 1000000;
+  int initial_dir = 1;
   % if l3_supported:
 
   void *ram_input = ram_malloc(input_size);
@@ -79,7 +80,7 @@ void application(void * arg) {
   % if l3_supported:
       ram_read(l2_buffer, ram_input, l2_input_size);
   % endif
-      ${prefix}network_run(l2_buffer, ${l2_buffer_size}, l2_buffer, ${"0" if single_input else "exec"}${f", {prefix}L2_input_h{' + exec * l2_input_size' if not single_input else ''}" if not l3_supported else ""});
+      ${prefix}network_run(l2_buffer, ${l2_buffer_size}, l2_buffer, ${"0" if single_input else "exec"}, initial_dir${f", {prefix}L2_input_h{' + exec * l2_input_size' if not single_input else ''}" if not l3_supported else ""});
 
   % if not single_input:
   }
diff --git a/dory/Hardware_targets/PULP/Common/Templates/network.c.t b/dory/Hardware_targets/PULP/Common/Templates/network.c.t
index 698eb346..a4b70247 100644
--- a/dory/Hardware_targets/PULP/Common/Templates/network.c.t
+++ b/dory/Hardware_targets/PULP/Common/Templates/network.c.t
@@ -104,7 +104,7 @@ void ${prefix}execute_layer_fork(void *args) {
   if (pi_core_id() == 0) pmsis_l1_malloc_free(layer_args->L1_buffer, ${l1_buffer});
 }
 
-struct ${prefix}network_run_token ${prefix}network_run_async(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""})
+struct ${prefix}network_run_token ${prefix}network_run_async(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec, int initial_dir${", void *L2_input_h" if not l3_supported else ""})
 {
   struct pi_device cluster_dev = {0};
   struct pi_cluster_conf conf;
@@ -120,8 +120,9 @@ struct ${prefix}network_run_token ${prefix}network_run_async(void *l2_buffer, si
   args[1] = (unsigned int) l2_buffer_size;
   args[2] = (unsigned int) l2_final_output;
   args[3] = (unsigned int) exec;
+  args[4] = (unsigned int) initial_dir;
   % if not l3_supported:
-  args[4] = (unsigned int) L2_input_h;
+  args[5] = (unsigned int) L2_input_h;
   % endif
   // open cluster...
   pi_cluster_task(&cluster_task, ${prefix}network_run_cluster, args);
@@ -145,9 +146,9 @@ void ${prefix}network_run_wait(struct ${prefix}network_run_token token)
   % endif
 }
 
-void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""})
+void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec, int initial_dir${", void *L2_input_h" if not l3_supported else ""})
 {
-  ${prefix}network_run_wait(network_run_async(l2_buffer, l2_buffer_size, l2_final_output, exec${", L2_input_h" if not l3_supported else ""}));
+  ${prefix}network_run_wait(network_run_async(l2_buffer, l2_buffer_size, l2_final_output, exec, initial_dir${", L2_input_h" if not l3_supported else ""}));
 }
 
 void ${prefix}network_run_cluster(void *args) {
@@ -156,8 +157,9 @@ void ${prefix}network_run_cluster(void *args) {
   size_t l2_buffer_size = (size_t) real_args[1];
   void * l2_final_output = (void *) real_args[2];
   int exec = (int) real_args[3];
+  int dir = (int) real_args[4];
   % if not l3_supported:
-  void * L2_input_h = (void *)real_args[4];
+  void * L2_input_h = (void *)real_args[5];
   % endif
 /*
   - initial buffer allocation L2 and L1
@@ -172,7 +174,6 @@ void ${prefix}network_run_cluster(void *args) {
   void *L3_weights_curr = L3_weights;
   void *bypass_activations = NULL;
 
-  int dir = 1;
   int residual_number = 0;
   int bypass_dimension = 0;
   % if not l3_supported:
diff --git a/dory/Hardware_targets/PULP/Common/Templates/network.h.t b/dory/Hardware_targets/PULP/Common/Templates/network.h.t
index 595aebaf..0ffff345 100644
--- a/dory/Hardware_targets/PULP/Common/Templates/network.h.t
+++ b/dory/Hardware_targets/PULP/Common/Templates/network.h.t
@@ -41,9 +41,9 @@ void ${prefix}network_terminate();
 void ${prefix}network_initialize();
 % endif
 void ${prefix}network_run_cluster(void * args);
-struct ${prefix}network_run_token ${prefix}network_run_async(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""});
+struct ${prefix}network_run_token ${prefix}network_run_async(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec, int initial_dir${", void *L2_input_h" if not l3_supported else ""});
 void network_run_wait(struct ${prefix}network_run_token token);
-void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""});
+void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec, int initial_dir${", void *L2_input_h" if not l3_supported else ""});
 void ${prefix}execute_layer_fork(void *arg);
 
 % if l3_supported and not single_input: