diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc
index 6ff363d766..ab7c612271 100644
--- a/paddle/fluid/operators/activation_mkldnn_op.cc
+++ b/paddle/fluid/operators/activation_mkldnn_op.cc
@@ -13,8 +13,8 @@
    limitations under the License. */
 
 #include "mkldnn.hpp"
-#include "mkldnn_activation_op.h"
 #include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/operators/mkldnn_activation_op.h"
 
 namespace paddle {
 namespace operators {
@@ -40,18 +40,24 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
   const T *dst_data = dst->template mutable_data<T>(ctx.GetPlace());
 
   // get memory dim
-  PADDLE_ENFORCE(src->dims().size() == 4,
-                 "Input dim must be with 4, i.e. NCHW");
+  PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4,
+                 "Input dim must be with 2 or 4");
   std::vector<int> src_tz = framework::vectorize2int(src->dims());
 
   // create memory description
-  // TODO(kbinias-intel): support more formats
-  auto data_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                         mkldnn::memory::format::nchw);
+  auto data_md = src_tz.size() == 2
+                     ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                               mkldnn::memory::format::nc)
+                     : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                               mkldnn::memory::format::nchw);
 
   // create memory primitives
-  auto src_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)src_data);
-  auto dst_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)dst_data);
+  auto src_memory =
+      mkldnn::memory({data_md, mkldnn_engine},
+                     static_cast<void *>(const_cast<float *>(src_data)));
+  auto dst_memory =
+      mkldnn::memory({data_md, mkldnn_engine},
+                     static_cast<void *>(const_cast<float *>(dst_data)));
 
   auto forward_desc = mkldnn::eltwise_forward::desc(
       mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta);
@@ -91,15 +97,21 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm,
   std::vector<int> src_tz = framework::vectorize2int(x->dims());
 
   // create memory description
-  auto data_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                         mkldnn::memory::format::nchw);
+  auto data_md = src_tz.size() == 2
+                     ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                               mkldnn::memory::format::nc)
+                     : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                               mkldnn::memory::format::nchw);
 
   // create memory primitives
-  auto src_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)src);
+  auto src_memory = mkldnn::memory(
+      {data_md, mkldnn_engine}, static_cast<void *>(const_cast<float *>(src)));
   auto diff_src_memory =
-      mkldnn::memory({data_md, mkldnn_engine}, (void *)diff_src);
+      mkldnn::memory({data_md, mkldnn_engine},
+                     static_cast<void *>(const_cast<float *>(diff_src)));
   auto diff_dst_memory =
-      mkldnn::memory({data_md, mkldnn_engine}, (void *)diff_dst);
+      mkldnn::memory({data_md, mkldnn_engine},
+                     static_cast<void *>(const_cast<float *>(diff_dst)));
 
   auto backward_desc =
       mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, alpha, beta);
diff --git a/python/paddle/fluid/tests/unittests/test_activation_op.py b/python/paddle/fluid/tests/unittests/test_activation_op.py
index fb162f8b73..c5b53902bc 100644
--- a/python/paddle/fluid/tests/unittests/test_activation_op.py
+++ b/python/paddle/fluid/tests/unittests/test_activation_op.py
@@ -535,9 +535,37 @@ class TestSwish(OpTest):
 
 
 #--------------------test MKLDNN--------------------
-class TestMKLDNNRelu(TestRelu):
+class TestMKLDNNReluDim2(TestRelu):
     def setUp(self):
-        super(TestMKLDNNRelu, self).setUp()
+        super(TestMKLDNNReluDim2, self).setUp()
+
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNTanhDim2(TestTanh):
+    def setUp(self):
+        super(TestMKLDNNTanhDim2, self).setUp()
+
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNSqrtDim2(TestSqrt):
+    def setUp(self):
+        super(TestMKLDNNSqrtDim2, self).setUp()
+
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNAbsDim2(TestAbs):
+    def setUp(self):
+        super(TestMKLDNNAbsDim2, self).setUp()
+
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNReluDim4(TestRelu):
+    def setUp(self):
+        super(TestMKLDNNReluDim4, self).setUp()
 
         x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
         # The same reason with TestAbs
@@ -549,9 +577,9 @@ class TestMKLDNNRelu(TestRelu):
         self.attrs = {"use_mkldnn": True}
 
 
-class TestMKLDNNTanh(TestTanh):
+class TestMKLDNNTanhDim4(TestTanh):
     def setUp(self):
-        super(TestMKLDNNTanh, self).setUp()
+        super(TestMKLDNNTanhDim4, self).setUp()
 
         self.inputs = {
             'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
@@ -560,9 +588,9 @@ class TestMKLDNNTanh(TestTanh):
         self.attrs = {"use_mkldnn": True}
 
 
-class TestMKLDNNSqrt(TestSqrt):
+class TestMKLDNNSqrtDim4(TestSqrt):
     def setUp(self):
-        super(TestMKLDNNSqrt, self).setUp()
+        super(TestMKLDNNSqrtDim4, self).setUp()
 
         self.inputs = {
             'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
@@ -571,9 +599,9 @@ class TestMKLDNNSqrt(TestSqrt):
         self.attrs = {"use_mkldnn": True}
 
 
-class TestMKLDNNAbs(TestAbs):
+class TestMKLDNNAbsDim4(TestAbs):
     def setUp(self):
-        super(TestMKLDNNAbs, self).setUp()
+        super(TestMKLDNNAbsDim4, self).setUp()
 
         x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
         # The same reason with TestAbs