fix print

src/main.py

@@ -35,9 +35,9 @@ def generate_main_unroll_step():
         unroll += "\tcudaDeviceSynchronize();\n\n"
         if i != len(list_gpusched_step_object) - 1:
             unroll += "\tstep_end = high_resolution_clock::now();\n"
-            unroll += "\tprintf(\"After step {no} at time %d\",step_end.count()),\n"
             unroll += "\telapsed_time = step_end - base;\n"
             unroll += "\telapsed_time_ms = duration_cast<ms>(elapsed_time);\n"
+            unroll += "\tprintf(\"After step {no} at time %d ms\",elapsed_time_ms.count());\n".format(no=currStep.step_no)
             unroll += "\tidle_time = milliseconds({next_step_min_time}) - elapsed_time_ms;\n".format(next_step_min_time=list_gpusched_step_object[i+1].min_time_start)
             unroll += "\tif (idle_time.count() > 0)\n\t\tstd::this_thread::sleep_for(idle_time);\n"