check idle time

src/main.py

@@ -33,6 +33,14 @@ def generate_main_unroll_step():
             currTask    = list_gpu_tasks_object[relatedTask]
             unroll +="\t{kernel_name}<<<{block},{thread},{smem},{stream}>>>({params});\n".format(kernel_name=currTask.name, block = currTask.blocks, thread = currTask.threads, smem = currTask.smem,stream = "stream_{no}".format(no=job),params="test")
         unroll += "\tcudaDeviceSynchronize();\n\n"
+        if i != len(list_gpusched_step_object) - 1:
+            unroll += "\tstep_end = high_resolution_clock::now();\n"
+            unroll += "\telapsed_time = step_end - base;\n"
+            unroll += "\telapsed_time_ms = duration_cast<ms>(elasped_time);\n"
+            unroll += "\tidle_time = {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 > 0)\n\t\tusleep(idle_time);\n"
+
+
         print(currStep.list_jobs)
 
     return unroll
@@ -56,6 +64,9 @@ def write_timer_initialization():
     chrono_dump += "\tusing namespace std::chrono\n";
     chrono_dump += "\ttypedef duration<double> dsec\n";
     chrono_dump += "\ttypedef milliseconds ms;\n";
+    chrono_dump += "\tdsec elapsed_time;\n"
+    chrono_dump += "\tms elapsed_time_ms;\n"
+    chrono_dump += "\tint idle_time;\n"
     chrono_dump += "\thigh_resolution_clock::time_point step_end;\n"
     chrono_dump += "\thigh_resolution_clock::time_point base = high_resolution_block::now();\n"
     return chrono_dump
@@ -67,6 +78,7 @@ def write_main(args):
         outFile.write("#include <stdio.h>\n"
                       "#include <chrono>\n"
                       "#include <string>\n"
+                      "#include <unistd.h>\n"
                       "#include \"steps_struct.hpp\"\n"
                       "#include \"gpu_tasks.hpp\"\n"
                       "#include \"{kernel_file}\"\n".format(kernel_file=args.kernel_declaration)