diff --git a/paddle/fluid/imperative/prepared_operator.cc b/paddle/fluid/imperative/prepared_operator.cc
index 82b91d2e77292dbefae54d0f7ecb7a2aff00f979..0336325bef6bd71208a839478741490ec4e10b66 100644
--- a/paddle/fluid/imperative/prepared_operator.cc
+++ b/paddle/fluid/imperative/prepared_operator.cc
@@ -42,23 +42,17 @@ static void PrepareData(const platform::Place& place,
     for (const auto& var_base : name_pair.second) {
       const auto* tensor = GetTensorFromVar(var_base->Var());
       if (tensor && tensor->IsInitialized()) {
-        auto tmp_place = tensor->place();
-
-        // TODO(jiabin): Support transform data layout when we Verify it on more
-        // tests
-        if (!(tmp_place == place)) {
-          auto kernel_type_for_var = op.GetKernelTypeForVar(
-              name_pair.first, *tensor, expected_kernel_key);
-          if (!NeedTransform(kernel_type_for_var, expected_kernel_key)) {
-            continue;
-          } else {
-            VLOG(3) << "Transform Variable " << var_base->Name() << " from "
-                    << kernel_type_for_var << " to " << expected_kernel_key;
-            framework::Tensor out;
-            TransformData(expected_kernel_key, kernel_type_for_var, *tensor,
-                          &out);
-            SetTensorToVariable(var_base->Var(), out, var_base->MutableVar());
-          }
+        auto kernel_type_for_var = op.GetKernelTypeForVar(
+            name_pair.first, *tensor, expected_kernel_key);
+        if (!NeedTransform(kernel_type_for_var, expected_kernel_key)) {
+          continue;
+        } else {
+          VLOG(3) << "Transform Variable " << var_base->Name() << " from "
+                  << kernel_type_for_var << " to " << expected_kernel_key;
+          framework::Tensor out;
+          TransformData(expected_kernel_key, kernel_type_for_var, *tensor,
+                        &out);
+          SetTensorToVariable(var_base->Var(), out, var_base->MutableVar());
         }
       }
     }
@@ -93,6 +87,13 @@ PreparedOp PrepareOpImpl(const NameVarMap<VarType>& ins,
   auto& kernels = kernels_iter->second;
 
   framework::RuntimeContext ctx({}, {});
+#ifdef PADDLE_WITH_MKLDNN
+  // MKLDNN variant of code reads attributes in some of GetKernelTypeForVar and
+  // GetKernelType functions, so we need to copy the attributes there.
+  // Const qualifier of Attrs had to be discarded to overwrite it.
+  auto& mutable_op_attrs = const_cast<framework::AttributeMap&>(op.Attrs());
+  mutable_op_attrs = attrs;
+#endif
   auto expected_kernel_key =
       op.GetExpectedKernelType(DygraphExecutionContext<VarType>(
           op, framework::Scope(), *dev_ctx, ctx, ins, outs, attrs));
diff --git a/paddle/fluid/imperative/tests/test_prepare_op.cc b/paddle/fluid/imperative/tests/test_prepare_op.cc
index c2e30b45a7f6c06ee6eb8945922a4317e9060491..f226c63f0c432e3878c7df6a5a04433ce047ff26 100644
--- a/paddle/fluid/imperative/tests/test_prepare_op.cc
+++ b/paddle/fluid/imperative/tests/test_prepare_op.cc
@@ -176,7 +176,7 @@ TEST(test_prepare_op, test_prepare_data) {
 }
 #endif
 
-TEST(test_prepare_op, test_prepare_data_same_place) {
+void TestPrepareDataSamePlace(framework::AttributeMap attr_map) {
   std::shared_ptr<imperative::VarBase> vin(
       new imperative::VarBase(false, "vin"));
   std::shared_ptr<imperative::VarBase> vout(
@@ -198,7 +198,6 @@ TEST(test_prepare_op, test_prepare_data_same_place) {
   var_pair out_pair = var_pair("Out", vb_vector(1, vout));
   imperative::NameVarBaseMap ins = {x_pair};
   imperative::NameVarBaseMap outs = {out_pair};
-  framework::AttributeMap attr_map;
   const std::string op_type = "relu";
   const auto& info = framework::OpInfoMap::Instance().Get(op_type);
   if (info.Checker()) info.Checker()->Check(&attr_map);
@@ -222,8 +221,21 @@ TEST(test_prepare_op, test_prepare_data_same_place) {
     }
   }
 }
+
+TEST(test_prepare_op, test_prepare_data_same_place) {
+  TestPrepareDataSamePlace({});
+}
+
+#ifdef PADDLE_WITH_MKLDNN
+TEST(test_prepare_op, test_prepare_data_cpu_mkldnn) {
+  TestPrepareDataSamePlace({{"use_mkldnn", true}});
+}
+#endif
 }  // namespace imperative
 }  // namespace paddle
 
 USE_OP(split);
 USE_OP(relu);
+#ifdef PADDLE_WITH_MKLDNN
+USE_OP_DEVICE_KERNEL(relu, MKLDNN);
+#endif
diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc
index a8c4107add1beeb9a7a5aedad9be982b6d8b6aac..9ed169fe3502e0c34b9f37d6520edc1a3fbfa91c 100644
--- a/paddle/fluid/operators/conv_op.cc
+++ b/paddle/fluid/operators/conv_op.cc
@@ -196,7 +196,7 @@ framework::OpKernelType ConvOp::GetKernelTypeForVar(
     auto ar = paddle::framework::AttrReader(attrs);
     const std::string data_format = ar.Get<std::string>("data_format");
     auto dl = framework::StringToDataLayout(data_format);
-    // Some models may have intentionally set "AnyLayout" for pool
+    // Some models may have intentionally set "AnyLayout" for conv
     // op. Treat this as NCHW (default data_format value)
     if (dl != framework::DataLayout::kAnyLayout) {
       return framework::OpKernelType(expected_kernel_type.data_type_,
diff --git a/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
index 55c6bad9af689196f1eda7acf916518ab2c130da..d904bdbfa96ae1df83a0cacde0822611ac55757e 100644
--- a/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
+++ b/python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
@@ -112,13 +112,10 @@ class TestMKLDNNSwishDim2(TestSwish):
     def setUp(self):
         super(TestMKLDNNSwishDim2, self).setUp()
 
-        x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype)
-        beta = 2.3
-        out = x * expit(beta * x)
+        self.attrs["use_mkldnn"] = True
 
-        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
-        self.outputs = {'Out': out}
-        self.attrs = {"use_mkldnn": True, "beta": beta}
+    def init_dtype(self):
+        self.dtype = np.float32
 
     def init_dtype(self):
         self.dtype = np.float32