MKLDNN implementation of batch normalization (#9904)

* Initial implementation of forward pass for MKLDNN batch norm

* Added attributes for MKLDNN batch norm

* MKLDNN batch norm forward pass passes unittest. Started working on backward

* Backward pass for MKLDNN batch norm added

* MKLDNN batch norm: scoring added to forward pass

* MKLDNN batch norm: bias as input added; handling AnyLayout when kernel is looked up

* MKLDNN batch norm: python unit tests added; mkldnn tests removed

* MKLDNN batch norm: changes required by cpplint

* MKLDNN batch norm: refactoring the operator

* MKLDNN batch norm: saved variance inversed in backward pass for correct execution of MKLDNN unit tests

* MKLDNN batch norm: refctoring, function for static/const cast to void* added

* MKLDNN batch norm: remove AnyLayout from batch norm

*  MKLDNN batch norm: only NCHW format is supported. Unittests refactored

* MKDNN batch norm: use_mkldnn added to attributes

* MKLDNN batch norm: AnyLayout removed from unittest

* MKLDNN batch norm: added CUDNN defines to batch norm

* MKLDNN batch norm: undefined data_format variable corrected

* MKLDNN batch norm: use_cudnn added, use of setUp method for configuring attributes

* MKLDNN batch norm: added use_cudnn attribute to batch norm operator

* MKLDNN batch norm: correcting batch norm unit tests for MKLDNN

* MKLDNN batch norm: MKLDNN tests moved to another file; reverting changes for saved variance not being inverted

* Change default layout to NCHW

* MKLDNN batch norm: init_kernel_type method added to unit tests

* MKLDNN batch norm: style changes

* MKLDNN batch norm: unit tests refactored

* MKLDNN batch norm: added use_mkldnn attribute to batch norm python interface

paddle/fluid/operators/batch_norm_mkldnn_op.cc

+/* Copyright (c) 2018 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. */
+
+#include "mkldnn.hpp"
+#include "paddle/fluid/operators/batch_norm_op.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using paddle::platform::MKLDNNDeviceContext;
+using paddle::platform::MKLDNNMemDesc;
+using mkldnn::memory;
+
+template <typename T>
+using EigenArrayMap =
+    Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using ConstEigenArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using EigenVectorArrayMap = Eigen::Map<Eigen::Array<T, Eigen::Dynamic, 1>>;
+template <typename T>
+using ConstEigenVectorArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, 1>>;
+
+namespace {
+template <typename T>
+struct bn_type_traits {
+  using op_type = T;
+  using op_desc = typename op_type::desc;
+  using op_prim = typename op_type::primitive_desc;
+};
+
+template <typename T, typename Container>
+void copy_to_weights(T scale_begin, T scale_end, T shift_begin, T shift_end,
+                     Container *c) {
+  auto it = std::begin(*c);
+
+  std::copy(scale_begin, scale_end, std::inserter(*c, it));
+  std::copy(
+      shift_begin, shift_end,
+      std::inserter(*c, std::next(it, std::distance(scale_begin, scale_end))));
+}
+
+template <typename Op, typename... Args>
+void run_batch_norm_op(Args &&... args) {
+  Op batch_norm_op{args...};
+
+  std::vector<mkldnn::primitive> pipeline;
+  pipeline.push_back(batch_norm_op);
+  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+}
+
+template <typename T>
+inline void *cast_const_to_void(const T *t) {
+  return static_cast<void *>(const_cast<T *>(t));
+}
+}  // namespace
+
+template <typename T>
+class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto data_layout_str = ctx.Attr<std::string>("data_layout");
+    auto data_layout = framework::StringToDataLayout(data_layout_str);
+    PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW,
+                   "MKLDNN batch normalization handles only NCHW data layout");
+
+    const float epsilon = ctx.Attr<float>("epsilon");
+    const float momentum = ctx.Attr<float>("momentum");
+    const bool is_test = ctx.Attr<bool>("is_test");
+
+    const auto *x = ctx.Input<Tensor>("X");
+    const auto *mean = ctx.Input<Tensor>("Mean");
+    const auto *variance = ctx.Input<Tensor>("Variance");
+
+    auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    auto mkldnn_engine = dev_ctx.GetEngine();
+
+    auto *y = ctx.Output<Tensor>("Y");
+    auto *mean_out = ctx.Output<Tensor>("MeanOut");
+    auto *variance_out = ctx.Output<Tensor>("VarianceOut");
+    auto *batch_mean = ctx.Output<Tensor>("SavedMean");
+    auto *batch_variance = ctx.Output<Tensor>("SavedVariance");
+
+    const auto *scale = ctx.Input<Tensor>("Scale");
+    const auto *shift = ctx.Input<Tensor>("Bias");
+
+    y->mutable_data<T>(ctx.GetPlace());
+    mean_out->mutable_data<T>(ctx.GetPlace());
+    variance_out->mutable_data<T>(ctx.GetPlace());
+
+    if (!is_test) {
+      batch_mean->mutable_data<T>(ctx.GetPlace());
+      batch_variance->mutable_data<T>(ctx.GetPlace());
+    }
+
+    auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring
+                                       : mkldnn::prop_kind::forward_training;
+
+    auto dims = paddle::framework::vectorize2int(x->dims());
+
+    auto src_md =
+        MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
+    auto dst_md =
+        MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
+
+    auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine};
+    auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine};
+
+    auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data<T>())};
+    auto dst = mkldnn::memory{dst_pd, y->data<T>()};
+
+    unsigned flags = mkldnn::use_scale_shift;
+    if (is_test) flags |= mkldnn::use_global_stats;
+
+    using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
+    auto batch_norm_fwd_desc =
+        bn_fwd_types::op_desc{propagation, src_md, epsilon, flags};
+    auto batch_norm_fwd_pd =
+        bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine};
+
+    const unsigned int ic = dims[1];
+
+    // MKLDNN requires a single piece of memory for scale and shift/bias data
+    const size_t scaleshift_size = 2 * ic;
+    std::vector<T> scaleshift_data;
+    scaleshift_data.reserve(scaleshift_size);
+
+    copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(),
+                    shift->data<T>() + ic, &scaleshift_data);
+
+    auto scaleshift_memory = mkldnn::memory{
+        batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()};
+
+    if (is_test) {
+      auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(),
+                                        cast_const_to_void(mean->data<T>())};
+
+      auto variance_memory =
+          mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(),
+                         cast_const_to_void(variance->data<T>())};
+
+      run_batch_norm_op<typename bn_fwd_types::op_type>(
+          batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory,
+          (const mkldnn::primitive::at &)variance_memory, scaleshift_memory,
+          dst);
+    } else {
+      auto mean_memory =
+          mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(),
+                         cast_const_to_void(batch_mean->data<T>())};
+
+      auto variance_memory =
+          mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(),
+                         cast_const_to_void(batch_variance->data<T>())};
+
+      run_batch_norm_op<bn_fwd_types::op_type>(batch_norm_fwd_pd, src,
+                                               scaleshift_memory, dst,
+                                               mean_memory, variance_memory);
+    }
+
+    if (!is_test) {
+      const unsigned int in = dims[0];
+      const unsigned int sample_size = x->numel() / in / ic;
+
+      // saved_xx is use just in this batch of data
+      EigenVectorArrayMap<T> saved_mean_e(
+          batch_mean->mutable_data<T>(ctx.GetPlace()), ic);
+      EigenVectorArrayMap<T> saved_variance_e(
+          batch_variance->mutable_data<T>(ctx.GetPlace()), ic);
+      saved_mean_e.setZero();
+      saved_variance_e.setZero();
+
+      const unsigned int x_arr_size = in * ic;
+      ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, x_arr_size);
+      for (unsigned int nc = 0; nc < x_arr_size; ++nc) {
+        saved_mean_e(nc % ic) += x_arr.col(nc).sum();
+      }
+      saved_mean_e /= in * sample_size;
+      for (unsigned int nc = 0; nc < x_arr_size; ++nc) {
+        saved_variance_e(nc % ic) +=
+            (x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm();
+      }
+      saved_variance_e /= in * sample_size;
+
+      ConstEigenVectorArrayMap<T> mean_arr{mean->data<T>(), ic};
+      ConstEigenVectorArrayMap<T> variance_arr{variance->data<T>(), ic};
+
+      EigenVectorArrayMap<T> running_mean_arr(
+          mean_out->mutable_data<T>(ctx.GetPlace()), ic);
+      EigenVectorArrayMap<T> running_var_arr(
+          variance_out->mutable_data<T>(ctx.GetPlace()), ic);
+
+      auto one_minus_momentum = 1. - momentum;
+      running_mean_arr =
+          mean_arr * momentum + saved_mean_e * one_minus_momentum;
+      running_var_arr =
+          variance_arr * momentum + saved_variance_e * one_minus_momentum;
+    }
+  }
+};
+
+template <typename T>
+class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext &ctx) const override {
+    auto data_layout_str = ctx.Attr<std::string>("data_layout");
+    auto data_layout = framework::StringToDataLayout(data_layout_str);
+    PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW,
+                   "MKLDNN batch normalization handles only NCHW data layout");
+
+    auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    auto mkldnn_engine = dev_ctx.GetEngine();
+
+    const float epsilon = ctx.Attr<float>("epsilon");
+
+    const auto *x = ctx.Input<Tensor>("X");
+    const auto *scale = ctx.Input<Tensor>("Scale");
+    const auto *shift = ctx.Input<Tensor>("Bias");
+    const auto *batch_mean = ctx.Input<Tensor>("SavedMean");
+    const auto *batch_variance = ctx.Input<Tensor>("SavedVariance");
+
+    const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto *diff_x = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *diff_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
+    auto *diff_shift = ctx.Output<Tensor>(framework::GradVarName("Bias"));
+
+    diff_x->mutable_data<T>(ctx.GetPlace());
+    diff_scale->mutable_data<T>(ctx.GetPlace());
+    diff_shift->mutable_data<T>(ctx.GetPlace());
+
+    auto dims = paddle::framework::vectorize2int(x->dims());
+    unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats;
+
+    auto src_md =
+        MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
+    auto dst_md =
+        MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
+    auto diff_src_md =
+        MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
+    auto diff_dst_md =
+        MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
+
+    using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>;
+    using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
+
+    auto batch_norm_fwd_desc = bn_fwd_types::op_desc{
+        mkldnn::prop_kind::forward_training, src_md, epsilon, flags};
+    auto batch_norm_fwd_pd =
+        bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine};
+
+    auto batch_norm_bwd_desc = bn_bwd_types::op_desc{
+        mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags};
+    auto batch_norm_bwd_pd = bn_bwd_types::op_prim{
+        batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd};
+
+    auto src = mkldnn::memory{{src_md, mkldnn_engine},
+                              cast_const_to_void(x->data<T>())};
+
+    auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(),
+                               cast_const_to_void(batch_mean->data<T>())};
+
+    auto variance =
+        mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(),
+                       cast_const_to_void(batch_variance->data<T>())};
+
+    auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine},
+                                   cast_const_to_void(diff_y->data<T>())};
+
+    const unsigned int ic = dims[1];
+
+    const size_t scaleshift_size = 2 * ic;
+
+    std::vector<T> scaleshift_data;
+    scaleshift_data.reserve(scaleshift_size);
+    copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(),
+                    shift->data<T>() + ic, &scaleshift_data);
+
+    auto scaleshift_memory = mkldnn::memory{
+        batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()};
+
+    std::vector<T> diff_scaleshift_data;
+    diff_scaleshift_data.reserve(scaleshift_size);
+    copy_to_weights(diff_scale->data<T>(), diff_scale->data<T>() + ic,
+                    diff_shift->data<T>(), diff_shift->data<T>() + ic,
+                    &diff_scaleshift_data);
+
+    auto diff_scaleshift_memory =
+        mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(),
+                       diff_scaleshift_data.data()};
+
+    auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine},
+                                   static_cast<void *>(diff_x->data<T>())};
+
+    run_batch_norm_op<bn_bwd_types::op_type>(
+        batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory,
+        diff_src, diff_scaleshift_memory);
+
+    auto it = std::begin(diff_scaleshift_data);
+    std::copy(it, std::next(it, ic), diff_scale->data<T>());
+    std::copy(std::next(it, ic), std::end(diff_scaleshift_data),
+              diff_shift->data<T>());
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace,
+                   ops::BatchNormMKLDNNOpKernel<float>);
+REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace,
+                   ops::BatchNormMKLDNNGradOpKernel<float>);

paddle/fluid/operators/batch_norm_op.cc

@@ -15,6 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/batch_norm_op.h"
 #include <string>
 #include "paddle/fluid/framework/data_layout.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 
 namespace paddle {
 namespace operators {
@@ -106,7 +109,18 @@ class BatchNormOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType(
                                          ctx.Input<Tensor>("Variance")->type()),
                       "Variance input should be of float type");
-    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_MKLDNN
+    if (library_ == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kMKLDNN;
+    }
+#endif
+    // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
+    framework::DataLayout layout = framework::DataLayout::kAnyLayout;
+    return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
+                                   library_);
   }
 };
 
@@ -151,6 +165,9 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
               "Variance of the current mini batch, "
               "will apply to output when training")
         .AsIntermediate();
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Batch Normalization.
 
@@ -349,8 +366,19 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
     if (t == nullptr) {
       PADDLE_THROW("can't find Y@GRAD");
     }
-    return framework::OpKernelType(framework::ToDataType(t->type()),
-                                   ctx.GetPlace());
+
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_MKLDNN
+    if (library_ == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kMKLDNN;
+    }
+#endif
+    // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
+    framework::DataLayout layout = framework::DataLayout::kAnyLayout;
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
+        layout, library_);
   }
 };
 
@@ -474,6 +502,7 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker {
     op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
 
     op->SetInput("Scale", Input("Scale"));
+    op->SetInput("Bias", Input("Bias"));
     op->SetInput("SavedMean", Output("SavedMean"));
     op->SetInput("SavedVariance", Output("SavedVariance"));
 

python/paddle/fluid/layers/nn.py

@@ -1496,6 +1496,7 @@ def batch_norm(input,
                bias_attr=None,
                data_layout='NCHW',
                in_place=False,
+               use_mkldnn=False,
                name=None,
                moving_mean_name=None,
                moving_variance_name=None,
@@ -1574,9 +1575,12 @@ def batch_norm(input,
             "SavedMean": saved_mean,
             "SavedVariance": saved_variance
         },
-        attrs={"momentum": momentum,
-               "epsilon": epsilon,
-               "is_test": is_test})
+        attrs={
+            "momentum": momentum,
+            "epsilon": epsilon,
+            "is_test": is_test,
+            "use_mkldnn": use_mkldnn
+        })
 
     return helper.append_activation(batch_norm_out)
 

python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py

+#   Copyright (c) 2018 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.
+
+import unittest
+import numpy as np
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
+import paddle.fluid as fluid
+from op_test import OpTest
+from paddle.fluid.framework import grad_var_name
+from test_batch_norm_op import TestBatchNormOpInference, TestBatchNormOpTraining, _reference_training, _reference_grad
+
+
+class TestMKLDNNBatchNormOpTraining(TestBatchNormOpTraining):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+        self.data_formats = ["NCHW"]
+
+    def ref_forward_backward(self, x, y_grad, scale, bias, mean, variance,
+                             epsilon, momentum, shape, data_layout):
+        # run forward
+        y, saved_mean, saved_variance = _reference_training(
+            x, scale, bias, epsilon, data_layout)
+        mean_out = saved_mean * (1. - momentum) + momentum * mean
+        variance_out = saved_variance * (1. - momentum) + momentum * variance
+        # run backward
+        x_grad, scale_grad, bias_grad = _reference_grad(
+            x, y_grad, scale, saved_mean, saved_variance, epsilon, data_layout)
+
+        return y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad
+
+
+class TestMKLDNNBatchNormOpInference(TestBatchNormOpInference):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+    def test_check_output(self):
+        place = core.CPUPlace()
+        data_format = "NCHW"
+
+        self.check_with_place(place, data_format, self.dtype, [2, 3, 4, 5])
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_batch_norm_op.py

@@ -158,6 +158,8 @@ def set_output_grad(scope, outputs, place, feed_dict=None):
 class TestBatchNormOpInference(unittest.TestCase):
     def setUp(self):
         self.dtype = np.float32
+        self.use_mkldnn = False
+        self.init_kernel_type()
 
     def __assert_close(self, tensor, np_array, msg, atol=1e-4):
         self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg)
@@ -230,6 +232,7 @@ class TestBatchNormOpInference(unittest.TestCase):
             # attrs
             is_test=True,
             data_layout=data_layout,
+            use_mkldnn=self.use_mkldnn,
             epsilon=epsilon)
 
         batch_norm_op.run(scope, place)
@@ -254,10 +257,15 @@ class TestBatchNormOpInference(unittest.TestCase):
                                       [2, 3, 4, 5])
                 self.check_with_place(place, data_format, self.dtype, [2, 3])
 
+    def init_kernel_type(self):
+        pass
+
 
 class TestFP16BatchNormOpInference(TestBatchNormOpInference):
     def setUp(self):
         self.dtype = np.float16
+        self.use_mkldnn = False
+        self.init_kernel_type()
 
     def test_check_output(self):
         places = []
@@ -274,9 +282,28 @@ class TestFP16BatchNormOpInference(TestBatchNormOpInference):
 
 
 class TestBatchNormOpTraining(unittest.TestCase):
+    def setUp(self):
+        self.use_mkldnn = False
+        self.data_formats = ["NCHW", "NHWC"]
+        self.init_kernel_type()
+
     def __assert_close(self, tensor, np_array, msg, atol=1e-4):
         np.allclose(np.array(tensor), np_array, atol=atol)
 
+    def ref_forward_backward(self, x, y_grad, scale, bias, mean, variance,
+                             epsilon, momentum, shape, data_layout):
+        # run forward
+        y, saved_mean, var_ref = _reference_training(x, scale, bias, epsilon,
+                                                     data_layout)
+        mean_out = saved_mean * (1. - momentum) + momentum * mean
+        variance_out = var_ref * (1. - momentum) + momentum * variance
+        saved_variance = 1. / np.sqrt(var_ref + epsilon)
+        # run backward
+        x_grad, scale_grad, bias_grad = _reference_grad(
+            x, y_grad, scale, saved_mean, var_ref, epsilon, data_layout)
+
+        return y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad
+
     def test_forward_backward(self):
         def test_with_place(place, data_layout, shape):
             # attr
@@ -295,16 +322,11 @@ class TestBatchNormOpTraining(unittest.TestCase):
             mean = np.zeros(scale_shape).astype(np.float32)
             variance = np.ones(scale_shape).astype(np.float32)
 
-            # run forward
-            y, saved_mean, var_ref = _reference_training(x, scale, bias,
-                                                         epsilon, data_layout)
-            mean_out = saved_mean * (1. - momentum) + momentum * mean
-            variance_out = var_ref * (1. - momentum) + momentum * variance
-            saved_variance = 1. / np.sqrt(var_ref + epsilon)
-            # run backward
             y_grad = np.random.random_sample(shape).astype(np.float32)
-            x_grad, scale_grad, bias_grad = _reference_grad(
-                x, y_grad, scale, saved_mean, var_ref, epsilon, data_layout)
+
+            y, mean_out, variance_out, saved_mean, saved_variance, x_grad, scale_grad, bias_grad = self.ref_forward_backward(
+                x, y_grad, scale, bias, mean, variance, epsilon, momentum,
+                shape, data_layout)
 
             var_dict = locals()
             var_dict['y@GRAD'] = y_grad
@@ -344,7 +366,8 @@ class TestBatchNormOpTraining(unittest.TestCase):
                         "momentum": momentum,
                         "epsilon": epsilon,
                         "is_test": False,
-                        "data_layout": data_layout
+                        "data_layout": data_layout,
+                        "use_mkldnn": self.use_mkldnn
                     })
                 block.create_var(name='y@GRAD', dtype='float32', shape=y.shape)
 
@@ -387,13 +410,17 @@ class TestBatchNormOpTraining(unittest.TestCase):
             print "op test forward passed: ", str(place), data_layout
 
         places = [core.CPUPlace()]
+
         if core.is_compiled_with_cuda() and core.op_support_gpu("batch_norm"):
             places.append(core.CUDAPlace(0))
 
         for place in places:
-            for data_format in ["NCHW", "NHWC"]:
+            for data_format in self.data_formats:
                 test_with_place(place, data_format, [2, 3, 4, 5])
 
+    def init_kernel_type(self):
+        pass
+
 
 if __name__ == '__main__':
     unittest.main()