add dsp_mode parameter

python/tools/tf_converter.py

@@ -32,7 +32,7 @@ def main(unused_args):
 
   if FLAGS.runtime == 'dsp':
     output_graph_def = tf_dsp_converter_lib.convert_to_mace_pb(
-      input_graph_def, FLAGS.input_node, FLAGS.output_node)
+      input_graph_def, FLAGS.input_node, FLAGS.output_node, FLAGS.dsp_mode)
   else:
     output_graph_def = tf_converter_lib.convert_to_mace_pb(
       input_graph_def, FLAGS.input_node, FLAGS.output_node, FLAGS.data_type, FLAGS.runtime, FLAGS.winograd)
@@ -119,6 +119,11 @@ def parse_args():
     const=False,
     default=False,
     help="open winograd convolution or not")
+  parser.add_argument(
+    "--dsp_mode",
+    type=int,
+    default=0,
+    help="dsp run mode, defalut=0")
   return parser.parse_known_args()
 
 

python/tools/tf_dsp_converter_lib.py

@@ -359,7 +359,7 @@ def fuse_quantize(net_def, input_node, output_node):
   new_net_def.op.extend(new_ops)
   return new_net_def
 
-def convert_to_mace_pb(input_graph_def, input_node, output_node):
+def convert_to_mace_pb(input_graph_def, input_node, output_node, dsp_mode):
   """
     nnlib does not have batch norm, so use tensorflow optimizer to fold
      batch norm with convolution. The fold optimization reorders ops, so
@@ -394,5 +394,9 @@ def convert_to_mace_pb(input_graph_def, input_node, output_node):
       dtype = mace_pb2.DT_FLOAT
       final_net_def = add_input_output_info(net_def_with_node_id, input_node, output_node, graph, dtype)
 
+      arg = final_net_def.arg.add()
+      arg.name = 'dsp_mode'
+      arg.i = dsp_mode
+
   return final_net_def