fix stream generation and add comment to signal start of stream generation

src/main.py

@@ -40,12 +40,13 @@ def generate_main_unroll_step():
 def generate_cudaStreams():
     stream_number  = GPUSchedResultStep.TRACK_MAX_CONCURRENT_JOBS
     stream_name = "stream_{no}"
-    dump_stream = "\tcudaStream_t "
+    dump_stream = "\t// Generate cuda streams\n"
+    dump_stream += "\tcudaStream_t "
     for i in range(0, stream_number-1):
         dump_stream += stream_name.format(no=i) + ", "
     dump_stream += stream_name.format(no=stream_number-1) + ";\n"
 
-    for i in range(0, stream_number-1):
+    for i in range(0, stream_number):
         dump_stream += "\tcudaStreamCreate(&{name});\n".format(name=stream_name.format(no=i))
     return dump_stream