diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc
index 54d177ea77781b0a40bf0c21fbaa2e42df686ad8..75b93088e55028be833ad4fd99df3dff0ee3cb23 100644
--- a/paddle/fluid/memory/detail/system_allocator.cc
+++ b/paddle/fluid/memory/detail/system_allocator.cc
@@ -48,7 +48,7 @@ void* AlignedMalloc(size_t size) {
   void* p = nullptr;
   size_t alignment = 32ul;
 #ifdef PADDLE_WITH_MKLDNN
-  // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
+  // refer to https://github.com/01org/mkl-dnn/blob/master/include/dnnl.hpp
   // memory alignment
   alignment = 4096ul;
 #endif
diff --git a/paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc b/paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
index e1506e37083666366ee47773cdb850fa983bedaa..6ef49e2cf3db7318f2eb8f0f55ffccd0e3bbad15 100644
--- a/paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
+++ b/paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
@@ -30,11 +30,10 @@ class GRUMKLDNNHandler : public RNNMKLDNNHandler<T, dnnl::gru_forward, T_out> {
  public:
   GRUMKLDNNHandler(const paddle::framework::ExecutionContext& ctx,
                    const platform::MKLDNNDeviceContext& dev_ctx,
-                   const mkldnn::engine mkldnn_engine,
-                   platform::Place cpu_place, const LoDTensor* input,
-                   const Tensor* weight_h, const Tensor* h0,
-                   const bool is_reverse, const int64_t N, const int64_t Ti,
-                   const int64_t IC, const int64_t OC,
+                   const dnnl::engine mkldnn_engine, platform::Place cpu_place,
+                   const LoDTensor* input, const Tensor* weight_h,
+                   const Tensor* h0, const bool is_reverse, const int64_t N,
+                   const int64_t Ti, const int64_t IC, const int64_t OC,
                    const std::string& unique_name)
       : RNNMKLDNNHandler<T, dnnl::gru_forward, T_out>(
             ctx, dev_ctx, mkldnn_engine, ctx.GetPlace(), input, weight_h, h0,
diff --git a/paddle/fluid/operators/matmul_op.cc b/paddle/fluid/operators/matmul_op.cc
index 051f97ad4ec8de8a56407e13c7221e6f0e4d1046..29e2cd08ce9fedf8e730dc6f713a87cf6efe57a3 100644
--- a/paddle/fluid/operators/matmul_op.cc
+++ b/paddle/fluid/operators/matmul_op.cc
@@ -730,7 +730,7 @@ class MatMulOp : public framework::OperatorWithKernel {
         OperatorWithKernel::IndicateOrPromoteVarDataTypes(ctx, "X", "Y");
 
 #ifdef PADDLE_WITH_MKLDNN
-    using mkldnn::memory;
+    using dnnl::memory;
     if (this->CanMKLDNNBeUsed(ctx, input_data_type)) {
       return framework::OpKernelType(input_data_type, ctx.GetPlace(),
                                      framework::DataLayout::kMKLDNN,
diff --git a/paddle/fluid/operators/mkldnn/cast_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/cast_mkldnn_op.cc
index 6f79c2dccf56b2683cdb961cb287afb095454509..552e91d6ff82604a03bea9a1a60e6905b03de4e2 100644
--- a/paddle/fluid/operators/mkldnn/cast_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/cast_mkldnn_op.cc
@@ -36,9 +36,8 @@ class CastMKLDNNKernel : public framework::OpKernel<T> {
     auto x_paddle_type = framework::proto::VarType::Type(in_dtype);
     auto out_paddle_type = framework::proto::VarType::Type(out_dtype);
 
-    mkldnn::memory::data_type x_type =
-        framework::ToMKLDNNDataType(x_paddle_type);
-    mkldnn::memory::data_type out_type =
+    dnnl::memory::data_type x_type = framework::ToMKLDNNDataType(x_paddle_type);
+    dnnl::memory::data_type out_type =
         framework::ToMKLDNNDataType(out_paddle_type);
 
     auto x_tz = framework::vectorize(x->dims());
diff --git a/paddle/fluid/operators/mkldnn/expand_v2_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/expand_v2_mkldnn_op.cc
index d537c3dbf9fddfc3f64d17ad2d39d5534a928108..b1be0f0f8fb441cead7c9138268fb6d058fc7496 100644
--- a/paddle/fluid/operators/mkldnn/expand_v2_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/expand_v2_mkldnn_op.cc
@@ -76,9 +76,9 @@ class ExpandMKLDNNKernel : public paddle::framework::OpKernel<T> {
  private:
   dnnl::memory::format_tag GetExtendedFormatTag(
       std::vector<int64_t>& dims, int new_size,
-      mkldnn::memory::format_tag format_tag) const {
-    mkldnn::memory::desc md(dims, paddle::platform::MKLDNNGetDataType<T>(),
-                            format_tag);
+      dnnl::memory::format_tag format_tag) const {
+    dnnl::memory::desc md(dims, paddle::platform::MKLDNNGetDataType<T>(),
+                          format_tag);
     std::vector<int64_t> new_dims(new_size, 1);
     std::copy(dims.begin(), dims.end(),
               new_dims.begin() + new_size - dims.size());
@@ -112,7 +112,7 @@ class ExpandGradMKLDNNKernel : public paddle::framework::OpKernel<T> {
 
     auto& astream = MKLDNNDeviceContext::tls().get_stream();
     if (dout_vec_dims == dx_vec_dims) {
-      mkldnn::memory::data_type dout_type =
+      dnnl::memory::data_type dout_type =
           paddle::framework::ToMKLDNNDataType(dout->type());
       paddle::platform::ReorderMKLDNNHandler reorder_handler(
           dout_vec_dims, dout->type(), dout_type, onednn_engine);
diff --git a/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc
index 51fa5ad021a2b284cd75f297d83326b2102c1e41..ab92a165c76d124e27c2635863846e52815c3d61 100644
--- a/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc
@@ -43,7 +43,7 @@ class GaussianMKLDNNKernel : public paddle::framework::OpKernel<T> {
     }
 
     tensor->set_layout(DataLayout::kMKLDNN);
-    tensor->set_format(mkldnn::memory::format_tag::oihw);
+    tensor->set_format(dnnl::memory::format_tag::oihw);
   }
 };
 }  // namespace operators
diff --git a/paddle/fluid/operators/mkldnn/layer_norm_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/layer_norm_mkldnn_op.cc
index 8ab4612ff04b504b31fddaf245ce4127e5edc8ae..1abd237acbc91d5ec3e725dcf7197f19db6595d0 100644
--- a/paddle/fluid/operators/mkldnn/layer_norm_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/layer_norm_mkldnn_op.cc
@@ -25,7 +25,7 @@ class LayerNormMKLDNNHandler : public platform::MKLDNNHandlerNoCachingT<
   LayerNormMKLDNNHandler(const std::vector<int64_t>& dims, const float& epsilon,
                          const dnnl::normalization_flags& flags,
                          const bool& is_test, const MKLDNNMemoryFormat fmt,
-                         const mkldnn::engine engine, platform::Place cpu_place)
+                         const dnnl::engine engine, platform::Place cpu_place)
       : platform::MKLDNNHandlerNoCachingT<T, dnnl::layer_normalization_forward>(
             engine, cpu_place) {
     auto md = dnnl::memory::desc(dims, platform::MKLDNNGetDataType<T>(), fmt);
@@ -131,7 +131,7 @@ class LayerNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     }
 
     if (with_scaleshift) {
-      std::shared_ptr<mkldnn::memory> scaleshift_memory =
+      std::shared_ptr<dnnl::memory> scaleshift_memory =
           handler.AcquireScaleShiftMemory(scale, bias);
       args.insert({DNNL_ARG_SCALE_SHIFT, *scaleshift_memory});
     }
diff --git a/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.h b/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.h
index 69ae78fcca08a09607fa8231bd0a9fb9c7f02cc8..c82119d06a01ea40760d1c345f36e3d50c38b74e 100644
--- a/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.h
+++ b/paddle/fluid/operators/mkldnn/matmul_mkldnn_op.h
@@ -32,7 +32,7 @@ class MatMulGradMKLDNNKernel : public framework::OpKernel<T> {
  private:
   void ExecuteMatMulGrad(const ExecutionContext& ctx,
                          const MKLDNNDeviceContext& dev_ctx,
-                         const mkldnn::engine& engine, Tensor* x, bool trans_x,
+                         const dnnl::engine& engine, Tensor* x, bool trans_x,
                          bool is_fold_init_dims_x, Tensor* y, bool trans_y,
                          bool is_fold_init_dims_y, Tensor* out) const;
   void RunKernel(const ExecutionContext& ctx) const;
diff --git a/paddle/fluid/operators/mkldnn/matmul_v2_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/matmul_v2_mkldnn_op.cc
index aa0a16944bcfabafa3a8184e7bc44c2c5bb9af20..0457aeed616fa33a8ac05d696ff7327f63138ce9 100644
--- a/paddle/fluid/operators/mkldnn/matmul_v2_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/matmul_v2_mkldnn_op.cc
@@ -33,7 +33,7 @@ template <typename T>
 class MatMulV2MKLDNNHandler
     : public paddle::platform::MKLDNNHandlerNoCachingT<T, dnnl::matmul> {
  public:
-  MatMulV2MKLDNNHandler(const mkldnn::engine engine,
+  MatMulV2MKLDNNHandler(const dnnl::engine engine,
                         paddle::platform::Place cpu_place,
                         const std::vector<int64_t>& x_org_dims, bool trans_x,
                         const std::vector<int64_t>& y_org_dims, bool trans_y,
@@ -132,7 +132,7 @@ class MatMulV2MKLDNNKernel
  protected:
   void ExecuteMatMul(const ExecutionContext& ctx,
                      const MKLDNNDeviceContext& dev_ctx,
-                     const mkldnn::engine onednn_engine,
+                     const dnnl::engine onednn_engine,
                      paddle::platform::Place cpu_place, const Tensor* x,
                      std::vector<int64_t>& x_dims, bool trans_x,
                      const Tensor* y, std::vector<int64_t>& y_dims,
@@ -272,7 +272,7 @@ class MatMulV2GradMKLDNNKernel : public MatMulV2MKLDNNKernel<T> {
 
   void ReduceSumForMatmulGradOutput(const ExecutionContext& ctx,
                                     const MKLDNNDeviceContext& dev_ctx,
-                                    const mkldnn::engine onednn_engine,
+                                    const dnnl::engine onednn_engine,
                                     const Tensor* dx_tmp, Tensor* dx,
                                     std::vector<int64_t> dx_dims) const {
     paddle::platform::ReductionMKLDNNHandler<T> handler(
diff --git a/paddle/fluid/operators/mkldnn/prelu_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/prelu_mkldnn_op.cc
index e2a4482666a1ace818777e9e7e3abaa1e6ff2f22..8296b4739d9541db6e7c0da995c8743fecbec22e 100644
--- a/paddle/fluid/operators/mkldnn/prelu_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/prelu_mkldnn_op.cc
@@ -31,7 +31,7 @@ class PReluMKLDNNHandler
                                       dnnl::prelu_backward> {
  public:
   PReluMKLDNNHandler(const MKLDNNDeviceContext& dev_ctx,
-                     const mkldnn::engine engine, platform::Place cpu_place,
+                     const dnnl::engine engine, platform::Place cpu_place,
                      const Tensor* x, const Tensor* weights,
                      const std::string& uniq_name, const std::string& mode,
                      bool is_test = false)
diff --git a/paddle/fluid/operators/mkldnn/reshape_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/reshape_mkldnn_op.cc
index 6c3f4ec06201a115d50074a2d9c5fd9aa63743fa..06142e95532c523edc68ea229de792f3fb4dbada 100644
--- a/paddle/fluid/operators/mkldnn/reshape_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/reshape_mkldnn_op.cc
@@ -74,7 +74,7 @@ class ReshapeMKLDNNKernel : public framework::OpKernel<T> {
 
     auto x_vec_dims = framework::vectorize(x_dims);
 
-    mkldnn::memory::data_type x_type = framework::ToMKLDNNDataType(x->type());
+    dnnl::memory::data_type x_type = framework::ToMKLDNNDataType(x->type());
     platform::ReorderMKLDNNHandler reorder_handler(x_vec_dims, x->type(),
                                                    x_type, onednn_engine);
 
@@ -197,7 +197,7 @@ class ReshapeMKLDNNKernel : public framework::OpKernel<T> {
   }
 
  protected:
-  static mkldnn::memory::format_tag getPlainFormatTag(const Tensor* tensor) {
+  static dnnl::memory::format_tag getPlainFormatTag(const Tensor* tensor) {
     auto tensor_dims_size = tensor->dims().size();
     PADDLE_ENFORCE_EQ(
         tensor_dims_size <= 6 && tensor_dims_size >= 1, true,
@@ -206,17 +206,17 @@ class ReshapeMKLDNNKernel : public framework::OpKernel<T> {
 
     switch (tensor_dims_size) {
       case 1:
-        return mkldnn::memory::format_tag::a;
+        return dnnl::memory::format_tag::a;
       case 2:
-        return mkldnn::memory::format_tag::ab;
+        return dnnl::memory::format_tag::ab;
       case 3:
-        return mkldnn::memory::format_tag::abc;
+        return dnnl::memory::format_tag::abc;
       case 4:
-        return mkldnn::memory::format_tag::abcd;
+        return dnnl::memory::format_tag::abcd;
       case 5:
-        return mkldnn::memory::format_tag::abcde;
+        return dnnl::memory::format_tag::abcde;
       default:
-        return mkldnn::memory::format_tag::abcdef;
+        return dnnl::memory::format_tag::abcdef;
     }
   }
 
@@ -324,7 +324,7 @@ class ReshapeGradMKLDNNKernel : public ReshapeMKLDNNKernel<T, op_name> {
 
     auto dout_vec_dims = framework::vectorize(dout->dims());
 
-    mkldnn::memory::data_type dout_type =
+    dnnl::memory::data_type dout_type =
         framework::ToMKLDNNDataType(dout->type());
     platform::ReorderMKLDNNHandler reorder_handler(dout_vec_dims, dout->type(),
                                                    dout_type, onednn_engine);
diff --git a/paddle/fluid/operators/mkldnn/slice_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/slice_mkldnn_op.cc
index 6bc3413604e221b3a9af16d131e4858fc0746510..d9bd843a9d0cf07ea23074c7605849cc147734ef 100644
--- a/paddle/fluid/operators/mkldnn/slice_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/slice_mkldnn_op.cc
@@ -15,24 +15,24 @@ limitations under the License. */
 #include "paddle/fluid/operators/utils.h"
 #include "paddle/fluid/platform/mkldnn_reuse.h"
 
-static mkldnn::memory::format_tag get_plain_format_tag(
+static dnnl::memory::format_tag get_plain_format_tag(
     const paddle::framework::Tensor* tensor) {
   auto tensor_dims_size = tensor->dims().size();
 
   switch (tensor_dims_size) {
     case 1:
-      return mkldnn::memory::format_tag::a;
+      return dnnl::memory::format_tag::a;
     case 2:
-      return mkldnn::memory::format_tag::ab;
+      return dnnl::memory::format_tag::ab;
     case 3:
-      return mkldnn::memory::format_tag::abc;
+      return dnnl::memory::format_tag::abc;
     case 4:
-      return mkldnn::memory::format_tag::abcd;
+      return dnnl::memory::format_tag::abcd;
     case 5:
-      return mkldnn::memory::format_tag::abcde;
+      return dnnl::memory::format_tag::abcde;
   }
 
-  return mkldnn::memory::format_tag::abcdef;
+  return dnnl::memory::format_tag::abcdef;
 }
 
 namespace paddle {
@@ -97,7 +97,7 @@ class SliceMKLDNNKernel : public framework::OpKernel<T> {
 
     out->Resize(framework::make_ddim(slice_dims));
 
-    mkldnn::memory::data_type x_type = framework::ToMKLDNNDataType(x->type());
+    dnnl::memory::data_type x_type = framework::ToMKLDNNDataType(x->type());
 
     platform::ReorderMKLDNNHandler reorder_handler(x_vec_dims, x->type(),
                                                    x_type, onednn_engine);
@@ -192,11 +192,11 @@ class SliceGradMKLDNNKernel : public framework::OpKernel<T> {
       slice_dims[axes[i]] = ends[i] - starts[i];
     }
 
-    mkldnn::memory::data_type dout_type =
+    dnnl::memory::data_type dout_type =
         framework::ToMKLDNNDataType(dout->type());
-    mkldnn::memory::desc md(dout_vec_dims, platform::MKLDNNGetDataType<T>(),
-                            dout->format());
-    mkldnn::memory::format_tag reorder_format_tag =
+    dnnl::memory::desc md(dout_vec_dims, platform::MKLDNNGetDataType<T>(),
+                          dout->format());
+    dnnl::memory::format_tag reorder_format_tag =
         platform::GetMKLDNNFormat(md.reshape(slice_dims));
 
     platform::ReorderMKLDNNHandler reorder_handler(slice_dims, dout->type(),
diff --git a/paddle/fluid/operators/mkldnn/softplus_mkldnn_op.h b/paddle/fluid/operators/mkldnn/softplus_mkldnn_op.h
index c8c539a956505892b5b000681c58ebf298e05ce0..053a1d0f7da4985edc102620c57b4885673e3abf 100644
--- a/paddle/fluid/operators/mkldnn/softplus_mkldnn_op.h
+++ b/paddle/fluid/operators/mkldnn/softplus_mkldnn_op.h
@@ -25,7 +25,7 @@ class SoftplusMKLDNNHandler
     : public platform::MKLDNNHandlerNoCachingT<T, dnnl::binary> {
  public:
   SoftplusMKLDNNHandler(const framework::ExecutionContext& ctx, const Tensor* x,
-                        const float beta, const mkldnn::engine engine)
+                        const float beta, const dnnl::engine engine)
       : platform::MKLDNNHandlerNoCachingT<T, dnnl::binary>(engine,
                                                            ctx.GetPlace()) {
     auto x_tz = framework::vectorize(x->dims());
@@ -53,7 +53,7 @@ class SoftplusMKLDNNHandler
                                             x_md, beta_md, x_md);
   }
 
-  std::shared_ptr<mkldnn::memory> AcquireBetaMemory(const float* beta) {
+  std::shared_ptr<dnnl::memory> AcquireBetaMemory(const float* beta) {
     return this->AcquireMemoryFromPrimitive(
         this->fwd_pd_->src1_desc(), platform::to_void_cast<float>(beta));
   }
diff --git a/paddle/fluid/operators/mkldnn/split_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/split_mkldnn_op.cc
index 411f33276c361c24bbf89dae7365bd9953ceed9d..a8e1e6e8a0e4c3a084290cf7fb47fe2e742bc13c 100644
--- a/paddle/fluid/operators/mkldnn/split_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/split_mkldnn_op.cc
@@ -90,7 +90,7 @@ class SplitMKLDNNKernel : public framework::OpKernel<T> {
 
     auto x_vec_dims = framework::vectorize(x_dims);
 
-    mkldnn::memory::data_type x_type = framework::ToMKLDNNDataType(x->type());
+    dnnl::memory::data_type x_type = framework::ToMKLDNNDataType(x->type());
 
     auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
 
diff --git a/paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc
index a92e8e6cb047f9e9b533bdbd045f8ee29c9ff329..f8698cdbb64075651d702764fe66d5f81410d25e 100644
--- a/paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc
@@ -29,14 +29,14 @@ class TransposeMKLDNNHandler {
  public:
   TransposeMKLDNNHandler(std::vector<int64_t>& dims,  // NOLINT
                          std::vector<int>& axis,      // NOLINT
-                         mkldnn::engine engine)
+                         dnnl::engine engine)
       : dims_(dims),
         axis_(axis),
         logical_axis_(dims.size(), 0),
         engine_(engine) {}
 
-  std::shared_ptr<mkldnn::memory> AcquireSrcMemory(
-      const MKLDNNMemoryFormat& fmt, void* ptr) {
+  std::shared_ptr<dnnl::memory> AcquireSrcMemory(const MKLDNNMemoryFormat& fmt,
+                                                 void* ptr) {
     // Make memory descriptor using input format, unless it
     // cannot be trusted (nchw) then make up memory fmt manually
     for (size_t i = 0; i < this->logical_axis_.size(); ++i) {
@@ -47,26 +47,26 @@ class TransposeMKLDNNHandler {
                       ? platform::MKLDNNMemDesc(
                             dims_, platform::MKLDNNGetDataType<T>(), fmt)
                       : Axis2MemoryDesc(dims_, logical_axis_);
-    return std::make_shared<mkldnn::memory>(src_md, engine_, ptr);
+    return std::make_shared<dnnl::memory>(src_md, engine_, ptr);
   }
 
-  std::shared_ptr<mkldnn::memory> AcquireDstMemory(framework::Tensor* output,
-                                                   platform::Place place) {
+  std::shared_ptr<dnnl::memory> AcquireDstMemory(framework::Tensor* output,
+                                                 platform::Place place) {
     auto dst_md = Axis2MemoryDesc(dims_, axis_);
     auto dst_data = output->mutable_data<T>(place, dst_md.get_size());
-    return std::make_shared<mkldnn::memory>(dst_md, engine_, dst_data);
+    return std::make_shared<dnnl::memory>(dst_md, engine_, dst_data);
   }
 
-  std::shared_ptr<mkldnn::reorder> AcquireTranspose(
-      std::shared_ptr<mkldnn::memory> dst_memory_p,
-      std::shared_ptr<mkldnn::memory> src_memory_p) {
-    return std::make_shared<mkldnn::reorder>(*(src_memory_p), *(dst_memory_p));
+  std::shared_ptr<dnnl::reorder> AcquireTranspose(
+      std::shared_ptr<dnnl::memory> dst_memory_p,
+      std::shared_ptr<dnnl::memory> src_memory_p) {
+    return std::make_shared<dnnl::reorder>(*(src_memory_p), *(dst_memory_p));
   }
 
  protected:
-  mkldnn::memory::desc Axis2MemoryDesc(std::vector<int64_t>& nchw_tz,  // NOLINT
-                                       std::vector<int>& axis          // NOLINT
-                                       ) {
+  dnnl::memory::desc Axis2MemoryDesc(std::vector<int64_t>& nchw_tz,  // NOLINT
+                                     std::vector<int>& axis          // NOLINT
+                                     ) {
     size_t ndims = axis.size();
 
     std::vector<int64_t> strides(ndims);
@@ -75,8 +75,8 @@ class TransposeMKLDNNHandler {
       strides[axis[i]] = total_stride;
       total_stride *= nchw_tz[axis[i]];
     }
-    mkldnn::memory::desc mem_d(nchw_tz, platform::MKLDNNGetDataType<T>(),
-                               strides);
+    dnnl::memory::desc mem_d(nchw_tz, platform::MKLDNNGetDataType<T>(),
+                             strides);
 
     return mem_d;
   }
@@ -85,7 +85,7 @@ class TransposeMKLDNNHandler {
   std::vector<int64_t> dims_;
   std::vector<int> axis_;
   std::vector<int> logical_axis_;
-  mkldnn::engine engine_;
+  dnnl::engine engine_;
 };
 
 template <typename T>
diff --git a/paddle/fluid/operators/optimizers/sgd_op.cc b/paddle/fluid/operators/optimizers/sgd_op.cc
index b2e258f815d7256501de462d15b2ac2668fd6dc3..e7c09430e9157632ab643bd791102af035392540 100644
--- a/paddle/fluid/operators/optimizers/sgd_op.cc
+++ b/paddle/fluid/operators/optimizers/sgd_op.cc
@@ -72,7 +72,7 @@ class SGDOp : public framework::OperatorWithKernel {
     auto data_type = OperatorWithKernel::IndicateVarDataType(ctx, "Param");
 
 #ifdef PADDLE_WITH_MKLDNN
-    using mkldnn::memory;
+    using dnnl::memory;
     if (this->CanMKLDNNBeUsed(ctx, data_type)) {
       const auto *param_var = ctx.InputVar("Param");
       const auto *grad_var = ctx.InputVar("Grad");
diff --git a/paddle/fluid/operators/reduce_ops/mkldnn/reduce_mkldnn_op.h b/paddle/fluid/operators/reduce_ops/mkldnn/reduce_mkldnn_op.h
index 17801454da2fa2cbbea9b9ddce1eeb1e6263cf6c..ad7defd22d05631f23e19a5621191824a96ee583 100644
--- a/paddle/fluid/operators/reduce_ops/mkldnn/reduce_mkldnn_op.h
+++ b/paddle/fluid/operators/reduce_ops/mkldnn/reduce_mkldnn_op.h
@@ -69,7 +69,7 @@ class ReduceMKLDNNKernel : public framework::OpKernel<T> {
     // In that case reorder must be executed to maintain compatibility with
     // PaddlePaddle reduce op
     if (input_dims == output_dims) {
-      mkldnn::memory::data_type input_type =
+      dnnl::memory::data_type input_type =
           framework::ToMKLDNNDataType(input->type());
       platform::ReorderMKLDNNHandler reorder_handler(input_dims, input->type(),
                                                      input_type, onednn_engine);
@@ -132,7 +132,7 @@ class ReduceGradMKLDNNKernel : public framework::OpKernel<T> {
     auto* input_dy = ctx.Input<Tensor>(framework::GradVarName("Out"));
     auto* output_dx = ctx.Output<Tensor>(framework::GradVarName("X"));
 
-    mkldnn::memory::format_tag x_format_tag;
+    dnnl::memory::format_tag x_format_tag;
     auto input_dims =
         CalculateReducedDims(output_dx, input_dy, dims, reduce_all, keep_dim);
     auto output_dims = framework::vectorize(output_dx->dims());
@@ -175,7 +175,7 @@ class ReduceGradMKLDNNKernel : public framework::OpKernel<T> {
   }
 
  protected:
-  mkldnn::memory::format_tag getPlainFormatTag(const Tensor* tensor) const {
+  dnnl::memory::format_tag getPlainFormatTag(const Tensor* tensor) const {
     auto tensor_dims_size = tensor->dims().size();
     PADDLE_ENFORCE_EQ(
         tensor_dims_size <= 5 && tensor_dims_size >= 1, true,
@@ -184,16 +184,16 @@ class ReduceGradMKLDNNKernel : public framework::OpKernel<T> {
 
     switch (tensor_dims_size) {
       case 1:
-        return mkldnn::memory::format_tag::a;
+        return dnnl::memory::format_tag::a;
       case 2:
-        return mkldnn::memory::format_tag::ab;
+        return dnnl::memory::format_tag::ab;
       case 3:
-        return mkldnn::memory::format_tag::abc;
+        return dnnl::memory::format_tag::abc;
       case 4:
-        return mkldnn::memory::format_tag::abcd;
+        return dnnl::memory::format_tag::abcd;
     }
 
-    return mkldnn::memory::format_tag::abcde;
+    return dnnl::memory::format_tag::abcde;
   }
 };
 
diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h
index e15b62413a16fecfe307d3a2ed16c382db682bf0..73232994516b61c28dc7d00be541a64d19ec55e8 100644
--- a/paddle/fluid/platform/device_context.h
+++ b/paddle/fluid/platform/device_context.h
@@ -48,7 +48,6 @@ limitations under the License. */
 #ifdef PADDLE_WITH_MKLDNN
 #include "dnnl.hpp"
 #include "paddle/fluid/framework/data_layout.h"
-namespace mkldnn = dnnl;
 #endif
 
 #include <map>
@@ -65,10 +64,6 @@ namespace mkldnn = dnnl;
 #endif
 #include "unsupported/Eigen/CXX11/Tensor"
 
-// This aias is required for now so that namespace name changes can be made to
-// less than 20 files at a time. After all the names are changed it will be
-// removed.
-
 namespace Eigen {
 struct DefaultDevice;
 struct GpuDevice;