diff --git a/api/common/env.py b/api/common/env.py
new file mode 100644
index 0000000000..523273302b
--- /dev/null
+++ b/api/common/env.py
@@ -0,0 +1,24 @@
+#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+import os
+
+
+def tf_use_xla():
+    return os.environ.get("TF_USE_XLA", False)
+
+
+def benchmark_use_feed_fetch():
+    return os.environ.get("BENCHMARK_USE_FEED_FETCH", False)
diff --git a/api/common/main.py b/api/common/main.py
index 3fab250920..24036a3cf8 100644
--- a/api/common/main.py
+++ b/api/common/main.py
@@ -22,6 +22,7 @@
 import collections
 import numpy as np
 
+from common import env
 from common import utils
 from common import system
 from common import api_param
@@ -228,7 +229,8 @@ def test_main_without_json(pd_obj=None,
 
     _adaptive_repeat(config, args)
     config.backward = args.backward
-    use_feed_fetch = True if args.task == "accuracy" else False
+    use_feed_fetch = True if args.task == "accuracy" else env.benchmark_use_feed_fetch(
+    )
 
     feeder_adapter = None
     if _is_tensorflow_enabled(args, config):
diff --git a/api/common/tensorflow_api_benchmark.py b/api/common/tensorflow_api_benchmark.py
index 9e3a65f01b..c734a69779 100644
--- a/api/common/tensorflow_api_benchmark.py
+++ b/api/common/tensorflow_api_benchmark.py
@@ -22,6 +22,7 @@
 import numpy as np
 from common import special_op_list
 
+from . import env
 from . import utils
 from . import api_param
 from . import feeder
@@ -384,6 +385,8 @@ def _init_session(self, use_gpu):
             config = tf.compat.v1.ConfigProto()
             if use_gpu:
                 config.gpu_options.allow_growth = self.allow_growth
+                if env.tf_use_xla():
+                    config.graph_options.optimizer_options.global_jit_level = tf.compat.v1.OptimizerOptions.ON_1
             else:
                 # In default, TF use full cpu cores, but Paddle use one cpu core.
                 # To make the same experiment, set TF use one cpu core as well.
diff --git a/api/tests/batch_norm.py b/api/tests/batch_norm.py
index 123688f8cb..3aaf63f1fd 100644
--- a/api/tests/batch_norm.py
+++ b/api/tests/batch_norm.py
@@ -22,9 +22,21 @@ def __init__(self):
     def init_from_json(self, filename, config_id=0, unknown_dim=16):
         super(BatchNormConfig, self).init_from_json(filename, config_id,
                                                     unknown_dim)
-        # TFBatchNorm does not have data_layout param, it only support NHWC format.
-        if self.data_layout == "NCHW":
-            self.run_tf = False
+        # tf's batch_norm does not have data_format param, it only support NHWC format.
+        if self.data_layout == "NCHW" and len(self.input_shape) == 4:
+            self.feed_spec = [
+                {
+                    "range": [-1, 1],
+                    "permute": [0, 2, 3, 1]
+                },  # input
+                {
+                    "range": [-1, 1],
+                },  # scale
+                {
+                    "range": [-1, 1],
+                }  # bias
+            ]
+
         if len(self.input_shape) == 4:
             if self.data_layout == "NCHW":
                 self.num_channels = self.input_shape[1]
@@ -35,6 +47,17 @@ def init_from_json(self, filename, config_id=0, unknown_dim=16):
 
     def to_tensorflow(self):
         tf_config = super(BatchNormConfig, self).to_tensorflow()
+        if self.data_layout == "NCHW" and len(self.input_shape) == 4:
+            print(
+                "Warning:\n"
+                "  1. tf's batch_norm does not have data_format param, it only support NHWC format. The benchmark test is actually running with NHWC format.\n"
+            )
+            tf_config.data_layout = "NHWC"
+            tf_config.input_shape = [
+                self.input_shape[0], self.input_shape[2], self.input_shape[3],
+                self.input_shape[1]
+            ]
+
         if len(tf_config.input_shape) == 4:
             tf_config.axes = [0, 1, 2]
         else:
@@ -72,20 +95,20 @@ class TFBatchNorm(TensorflowAPIBenchmarkBase):
     def build_graph(self, config):
         input = self.variable(
             name='input', shape=config.input_shape, dtype=config.input_dtype)
+        bias = self.variable(
+            name='bias', shape=[config.num_channels], dtype=config.input_dtype)
         scale = self.variable(
             name='scale',
             shape=[config.num_channels],
             dtype=config.input_dtype)
-        bias = self.variable(
-            name='bias', shape=[config.num_channels], dtype=config.input_dtype)
         mean, var = tf.nn.moments(
             x=input, axes=config.axes, shift=None, keepdims=False)
         result = tf.nn.batch_normalization(
             x=input,
             mean=mean,
             variance=var,
-            offset=bias,
-            scale=scale,
+            offset=bias,  # beta
+            scale=scale,  # gamma
             variance_epsilon=config.epsilon)
 
         self.feed_list = [input, scale, bias]
diff --git a/api/tests/configs/batch_norm.json b/api/tests/configs/batch_norm.json
index 47313e9a10..ad981830dd 100644
--- a/api/tests/configs/batch_norm.json
+++ b/api/tests/configs/batch_norm.json
@@ -38,7 +38,8 @@
             "type": "bool",
             "value": "False"
         }
-    }
+    },
+    "atol": 1E-05
 }, {
     "op": "batch_norm",
     "param_info": {