Fix densenet-121 throughput issue (#1113)

Author: lvliang-intel

neural_compressor/adaptor/tf_utils/graph_rewriter/generic/pre_optimize.py

@@ -108,6 +108,11 @@ def get_optimized_model(self, itex_mode=False):
 
         self._tmp_graph_def = SplitSharedInputOptimizer(self._tmp_graph_def).do_transformation()
 
+        # Put FuseDecomposedBNOptimizer before GraphFoldConstantOptimizer
+        # The 'Sub' op in the small decomposed ops of BN will be converted to const by GraphFoldConstantOptimizer.
+        # Then the FuseDecomposedBNOptimizer can't fuse the small decomposed ops to BN.
+        self._tmp_graph_def = FuseDecomposedBNOptimizer(self._tmp_graph_def).do_transformation()
+
         # disable fold constant for itex qdq mode
         if not itex_mode:
             self._tmp_graph_def = GraphFoldConstantOptimizer(self._tmp_graph_def).do_transformation()
@@ -121,8 +126,6 @@ def get_optimized_model(self, itex_mode=False):
 
         self._tmp_graph_def = GraphCseOptimizer(self._tmp_graph_def).do_transformation()
 
-        self._tmp_graph_def = FuseDecomposedBNOptimizer(self._tmp_graph_def).do_transformation()
-
         self._tmp_graph_def = FoldBatchNormNodesOptimizer(
             self._tmp_graph_def).do_transformation()