diff --git a/CMakeLists.txt b/CMakeLists.txt
index 487aa200d7fc4df5850d9afd35e35fa119ca0e71..6c2848d0b1969d352bc9f4e20d843610912886df 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -84,6 +84,8 @@ if(WIN32)
         endforeach(flag_var)
     endif()
 
+    # NOTE(Avin0323): Less parallel count result in faster compilation.
+    math(EXPR PROCESS_MAX "${CPU_CORES} * 2 / 3")
     # windows build turn off warnings, use parallel compiling.
     foreach(flag_var
         CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
@@ -91,13 +93,7 @@ if(WIN32)
         CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
         CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO)
         string(REGEX REPLACE "/W[1-4]" " /W0 " ${flag_var} "${${flag_var}}")
-        # NOTE(Avin0323): Less parallel count result in faster compilation with
-        # Unity Build on GPU.
-        if(WITH_UNITY_BUILD AND WITH_GPU)
-            set(${flag_var} "${${flag_var}} /MP8")
-        else()
-            set(${flag_var} "${${flag_var}} /MP")
-        endif()
+        set(${flag_var} "${${flag_var}} /MP${PROCESS_MAX}")
     endforeach(flag_var)
     foreach(flag_var CMAKE_CXX_FLAGS CMAKE_C_FLAGS)
         set(${flag_var} "${${flag_var}} /w")
diff --git a/paddle/fluid/operators/average_accumulates_op.h b/paddle/fluid/operators/average_accumulates_op.h
index 338e46111fca83230aca1c7877578e557cef5a31..6813f566758261f6c6f8af4fa3711b345dad546a 100644
--- a/paddle/fluid/operators/average_accumulates_op.h
+++ b/paddle/fluid/operators/average_accumulates_op.h
@@ -23,10 +23,6 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-
 template <typename DeviceContext>
 void GetAccumulators(const framework::ExecutionContext& ctx,
                      int64_t* num_updates, int64_t* num_accumulates,
@@ -67,18 +63,18 @@ class AverageAccumulatesKernel : public framework::OpKernel<T> {
     auto* in_sum_1 = ctx.Input<Tensor>("in_sum_1");
     auto* in_sum_2 = ctx.Input<Tensor>("in_sum_2");
     auto* in_sum_3 = ctx.Input<Tensor>("in_sum_3");
-    auto param_tensor = EigenVector<T>::Flatten(*param);
-    auto in_sum_1_tensor = EigenVector<T>::Flatten(*in_sum_1);
-    auto in_sum_2_tensor = EigenVector<T>::Flatten(*in_sum_2);
-    auto in_sum_3_tensor = EigenVector<T>::Flatten(*in_sum_3);
+    auto param_tensor = framework::EigenVector<T>::Flatten(*param);
+    auto in_sum_1_tensor = framework::EigenVector<T>::Flatten(*in_sum_1);
+    auto in_sum_2_tensor = framework::EigenVector<T>::Flatten(*in_sum_2);
+    auto in_sum_3_tensor = framework::EigenVector<T>::Flatten(*in_sum_3);
 
     // Get outputs
     auto* out_sum_1 = ctx.Output<Tensor>("out_sum_1");
     auto* out_sum_2 = ctx.Output<Tensor>("out_sum_2");
     auto* out_sum_3 = ctx.Output<Tensor>("out_sum_3");
-    auto out_sum_1_tensor = EigenVector<T>::Flatten(*out_sum_1);
-    auto out_sum_2_tensor = EigenVector<T>::Flatten(*out_sum_2);
-    auto out_sum_3_tensor = EigenVector<T>::Flatten(*out_sum_3);
+    auto out_sum_1_tensor = framework::EigenVector<T>::Flatten(*out_sum_1);
+    auto out_sum_2_tensor = framework::EigenVector<T>::Flatten(*out_sum_2);
+    auto out_sum_3_tensor = framework::EigenVector<T>::Flatten(*out_sum_3);
 
     // Compute
     auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
diff --git a/paddle/fluid/operators/conv_transpose_cudnn_op.cu b/paddle/fluid/operators/conv_transpose_cudnn_op.cu
index a12629b7a4959179d176f24c26a42af7aad3a277..edf00eb2ba9a76b1a288999acff4dee6e50f8817 100644
--- a/paddle/fluid/operators/conv_transpose_cudnn_op.cu
+++ b/paddle/fluid/operators/conv_transpose_cudnn_op.cu
@@ -25,7 +25,6 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
-using DataLayout = platform::DataLayout;
 
 template <typename T, int D>
 static void DataTranspose(const framework::ExecutionContext& ctx,
@@ -67,14 +66,15 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
     int groups = ctx.Attr<int>("groups");
     const T* filter_data = filter->data<T>();
     const std::string data_layout_str = ctx.Attr<std::string>("data_format");
-    const paddle::operators::DataLayout data_layout =
-        (data_layout_str != "NHWC" ? DataLayout::kNCHW : DataLayout::kNHWC);
+    const paddle::platform::DataLayout data_layout =
+        (data_layout_str != "NHWC" ? platform::DataLayout::kNCHW
+                                   : platform::DataLayout::kNHWC);
 
     // if channel_last, transpose to channel_first
     Tensor input_transpose;
     std::vector<int> input_vec = framework::vectorize<int>(input->dims());
     std::vector<int> output_vec = framework::vectorize<int>(output->dims());
-    if (data_layout == DataLayout::kNHWC) {
+    if (data_layout == platform::DataLayout::kNHWC) {
       if (strides.size() == 2U) {
         std::vector<int> axis = {0, 3, 1, 2};
         for (size_t i = 0; i < axis.size(); ++i) {
@@ -195,7 +195,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
     }
     T* transformed_output_data = transformed_output.data<T>();
 
-    DataLayout layout;
+    platform::DataLayout layout;
 
     int iwo_groups = groups;
     int c_groups = 1;
@@ -206,9 +206,9 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
 #endif
 
     if (strides.size() == 2U) {
-      layout = DataLayout::kNCHW;
+      layout = platform::DataLayout::kNCHW;
     } else {
-      layout = DataLayout::kNCDHW;
+      layout = platform::DataLayout::kNCDHW;
     }
 
     size_t workspace_size = 0;
@@ -269,7 +269,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
           ctx, &transformed_output, output, starts, ends, axes);
     }
 
-    if (data_layout == DataLayout::kNHWC) {
+    if (data_layout == platform::DataLayout::kNHWC) {
       Tensor output_transpose;
       Tensor output_nchw;
       output_nchw.ShareDataWith(*output);
@@ -309,8 +309,9 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
     std::string padding_algorithm = ctx.Attr<std::string>("padding_algorithm");
     int user_workspace_size = ctx.Attr<int>("workspace_size_MB");
     const std::string data_layout_str = ctx.Attr<std::string>("data_format");
-    const paddle::operators::DataLayout data_layout =
-        (data_layout_str != "NHWC" ? DataLayout::kNCHW : DataLayout::kNHWC);
+    const paddle::platform::DataLayout data_layout =
+        (data_layout_str != "NHWC" ? platform::DataLayout::kNCHW
+                                   : platform::DataLayout::kNHWC);
 
     // if channel_last, transpose to channel_first
     Tensor input_transpose;
@@ -318,7 +319,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
     std::vector<int> input_vec = framework::vectorize<int>(input->dims());
     std::vector<int> output_vec =
         framework::vectorize<int>(output_grad->dims());
-    if (data_layout == DataLayout::kNHWC) {
+    if (data_layout == platform::DataLayout::kNHWC) {
       if (strides.size() == 2U) {
         std::vector<int> axis = {0, 3, 1, 2};
         for (size_t i = 0; i < axis.size(); ++i) {
@@ -416,12 +417,12 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
     output_vec = framework::vectorize<int>(transformed_output_grad.dims());
 
     // ------------------- cudnn descriptors ---------------------
-    DataLayout layout;
+    platform::DataLayout layout;
 
     if (strides.size() == 2U) {
-      layout = DataLayout::kNCHW;
+      layout = platform::DataLayout::kNCHW;
     } else {
-      layout = DataLayout::kNCDHW;
+      layout = platform::DataLayout::kNCDHW;
     }
 
     int iwo_groups = groups;
@@ -515,7 +516,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
         workspace_handle.RunFunc(cudnn_func, workspace_size);
       }
 
-      if (data_layout == DataLayout::kNHWC) {
+      if (data_layout == platform::DataLayout::kNHWC) {
         Tensor input_grad_transpose;
         Tensor input_grad_nchw;
         input_grad_nchw.ShareDataWith(*input_grad);
@@ -849,7 +850,7 @@ class CUDNNConvTransposeDoubleGradOpKernel : public framework::OpKernel<T> {
     cudnnConvolutionBwdFilterAlgo_t filter_algo =
         static_cast<cudnnConvolutionBwdFilterAlgo_t>(0);
 
-    auto layout = GetCudnnTensorFormat(DataLayout::kNCHW);
+    auto layout = GetCudnnTensorFormat(platform::DataLayout::kNCHW);
 
     // ddo = conv(ddI, W) + conv(I, ddW)
     size_t workspace_size = 0;
@@ -916,12 +917,12 @@ class CUDNNConvTransposeDoubleGradOpKernel : public framework::OpKernel<T> {
     }
 
     int i_n, i_c, i_d, i_h, i_w;
-    GetNCDHW(transformed_X.dims(), DataLayout::kNCHW, &i_n, &i_c, &i_d, &i_h,
-             &i_w);
+    GetNCDHW(transformed_X.dims(), platform::DataLayout::kNCHW, &i_n, &i_c,
+             &i_d, &i_h, &i_w);
 
     int o_n, o_c, o_d, o_h, o_w;
-    GetNCDHW(transformed_dO.dims(), DataLayout::kNCHW, &o_n, &o_c, &o_d, &o_h,
-             &o_w);
+    GetNCDHW(transformed_dO.dims(), platform::DataLayout::kNCHW, &o_n, &o_c,
+             &o_d, &o_h, &o_w);
 
     int group_offset_in =
         transformed_X.numel() / transformed_X.dims()[0] / groups;
diff --git a/paddle/fluid/operators/correlation_op.cu b/paddle/fluid/operators/correlation_op.cu
index 0d177f653ec3d5cacbd4d938e6e5da6689b1bc74..6cf1ff5e7284026d784e9c855ddadd860cdae23b 100644
--- a/paddle/fluid/operators/correlation_op.cu
+++ b/paddle/fluid/operators/correlation_op.cu
@@ -23,7 +23,6 @@ namespace operators {
 #define FULL_MASK 0xffffffff
 
 using framework::Tensor;
-using DataLayout = framework::DataLayout;
 
 template <typename T>
 __forceinline__ __device__ T warpReduceSum(T val) {
diff --git a/paddle/fluid/operators/dot_op.h b/paddle/fluid/operators/dot_op.h
index c78ac87084cafe5668015904b18cc4937f41119f..a197e2149ee027d847b549b584dd314905ab8bdc 100644
--- a/paddle/fluid/operators/dot_op.h
+++ b/paddle/fluid/operators/dot_op.h
@@ -26,10 +26,6 @@ using Tensor = framework::Tensor;
 using complex64 = platform::complex64;
 using complex128 = platform::complex128;
 
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
-
 template <typename T, typename R>
 struct P {
   void operator()(T a, R b);
@@ -85,11 +81,11 @@ struct DotGradFunction<DeviceContext, T, math::EnableComplex<T>> {
         dy.device(dev) = dy * dout.broadcast(size);
       }
     } else {
-      auto dout = EigenMatrix<T>::From(*tensor_dout);
+      auto dout = framework::EigenMatrix<T>::From(*tensor_dout);
 
       if (tensor_dx) {
         tensor_dx->mutable_data<T>(ctx.GetPlace());
-        auto y = EigenMatrix<T>::From(*tensor_y);
+        auto y = framework::EigenMatrix<T>::From(*tensor_y);
         auto& dev_raw = ctx.template device_context<DeviceContext>();
         auto& dev = *dev_raw.eigen_device();
         Eigen::DSizes<int, 2> size(1, tensor_dx->dims()[1]);
@@ -99,14 +95,14 @@ struct DotGradFunction<DeviceContext, T, math::EnableComplex<T>> {
         math::ConjFunctor<T> functor(tensor_y->data<T>(), tensor_y->numel(),
                                      tensor_dx->data<T>());
         for_range(functor);
-        auto dx = EigenMatrix<T>::From(*tensor_dx);
+        auto dx = framework::EigenMatrix<T>::From(*tensor_dx);
 
         dx.device(dev) = dx * dout.broadcast(size);
       }
 
       if (tensor_dy) {
         tensor_dy->mutable_data<T>(ctx.GetPlace());
-        auto x = EigenMatrix<T>::From(*tensor_x);
+        auto x = framework::EigenMatrix<T>::From(*tensor_x);
         auto& dev_raw = ctx.template device_context<DeviceContext>();
         auto& dev = *dev_raw.eigen_device();
         Eigen::DSizes<int, 2> size(1, tensor_dy->dims()[1]);
@@ -117,7 +113,7 @@ struct DotGradFunction<DeviceContext, T, math::EnableComplex<T>> {
                                      tensor_dy->data<T>());
         for_range(functor);
 
-        auto dy = EigenMatrix<T>::From(*tensor_dy);
+        auto dy = framework::EigenMatrix<T>::From(*tensor_dy);
 
         dy.device(dev) = dy * dout.broadcast(size);
       }
@@ -186,12 +182,12 @@ struct DotGradFunction<DeviceContext, T, math::DisableComplex<T>> {
         dy.device(dev) = x * dout.broadcast(size);
       }
     } else {
-      auto dout = EigenMatrix<T>::From(*tensor_dout);
+      auto dout = framework::EigenMatrix<T>::From(*tensor_dout);
 
       if (tensor_dx) {
         tensor_dx->mutable_data<T>(ctx.GetPlace());
-        auto y = EigenMatrix<T>::From(*tensor_y);
-        auto dx = EigenMatrix<T>::From(*tensor_dx);
+        auto y = framework::EigenMatrix<T>::From(*tensor_y);
+        auto dx = framework::EigenMatrix<T>::From(*tensor_dx);
         auto& dev =
             *ctx.template device_context<DeviceContext>().eigen_device();
         Eigen::DSizes<int, 2> size(1, tensor_dx->dims()[1]);
@@ -200,8 +196,8 @@ struct DotGradFunction<DeviceContext, T, math::DisableComplex<T>> {
 
       if (tensor_dy) {
         tensor_dy->mutable_data<T>(ctx.GetPlace());
-        auto x = EigenMatrix<T>::From(*tensor_x);
-        auto dy = EigenMatrix<T>::From(*tensor_dy);
+        auto x = framework::EigenMatrix<T>::From(*tensor_x);
+        auto dy = framework::EigenMatrix<T>::From(*tensor_dy);
         auto& dev =
             *ctx.template device_context<DeviceContext>().eigen_device();
         Eigen::DSizes<int, 2> size(1, tensor_dy->dims()[1]);
@@ -262,9 +258,9 @@ class DotKernel : public framework::OpKernel<T> {
       auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
       out.device(dev) = (x * y).sum();
     } else {
-      auto out = EigenMatrix<T>::From(*tensor_out);
-      auto x = EigenMatrix<T>::From(*tensor_x);
-      auto y = EigenMatrix<T>::From(*tensor_y);
+      auto out = framework::EigenMatrix<T>::From(*tensor_out);
+      auto x = framework::EigenMatrix<T>::From(*tensor_x);
+      auto y = framework::EigenMatrix<T>::From(*tensor_y);
 
       auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
       out.device(dev) = (x * y).sum(Eigen::DSizes<int, 1>(1));
diff --git a/paddle/fluid/operators/meshgrid_op.h b/paddle/fluid/operators/meshgrid_op.h
index d591912bef800afac50a3e9753dab7f402655b30..11cd43b22045c3b04d5495901be50845cccae206 100644
--- a/paddle/fluid/operators/meshgrid_op.h
+++ b/paddle/fluid/operators/meshgrid_op.h
@@ -50,16 +50,6 @@
 namespace paddle {
 namespace operators {
 
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-template <typename T, size_t D, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
-
 template <typename DeviceContext, typename T>
 class MeshgridKernel : public framework::OpKernel<T> {
  public:
@@ -120,9 +110,9 @@ class MeshgridKernel : public framework::OpKernel<T> {
       bcast_dims[i] = 1;
 
       outs[i]->Resize(out_dims);
-      auto x = EigenTensor<T, Rank>::From(reshape_ins_tensor);
+      auto x = framework::EigenTensor<T, Rank>::From(reshape_ins_tensor);
       outs[i]->mutable_data<T>(context.GetPlace());
-      auto y = EigenTensor<T, Rank>::From(*outs[i]);
+      auto y = framework::EigenTensor<T, Rank>::From(*outs[i]);
       auto& place =
           *context.template device_context<DeviceContext>().eigen_device();
       y.device(place) = x.broadcast(bcast_dims);
@@ -159,8 +149,8 @@ class MeshgridGradKernel : public framework::OpKernel<T> {
 
     for (int i = 0; i < n; i++) {
       outs[i]->mutable_data<T>(context.GetPlace());
-      auto out_grad_tmp = EigenVector<T>::Flatten(*out_grad[i]);
-      auto in_grad = EigenVector<T>::Flatten(*outs[i]);
+      auto out_grad_tmp = framework::EigenVector<T>::Flatten(*out_grad[i]);
+      auto in_grad = framework::EigenVector<T>::Flatten(*outs[i]);
 
       std::vector<int> reduce_dims_vec;
       std::vector<int> reshape_dims_vec;
diff --git a/paddle/fluid/operators/rank_loss_op.h b/paddle/fluid/operators/rank_loss_op.h
index 28626c0e2e697cef29db3b969fa55a1da78dd8a4..8609958476f60a0c03b399f8fa2a00b29f3a9011 100644
--- a/paddle/fluid/operators/rank_loss_op.h
+++ b/paddle/fluid/operators/rank_loss_op.h
@@ -37,7 +37,7 @@ class RankLossKernel : public framework::OpKernel<T> {
 
     auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
     out.device(dev) =
-        (1. + (left - right).exp()).log() - label * (left - right);
+        (1.0f + (left - right).exp()).log() - label * (left - right);
   }
 };
 
@@ -65,14 +65,15 @@ class RankLossGradKernel : public framework::OpKernel<T> {
     if (d_left_t) {
       d_left_t->mutable_data<T>(ctx.GetPlace());
       auto d_left = framework::EigenVector<T>::Flatten(*d_left_t);
-      d_left.device(dev) = d_out * (1. / (1. + (right - left).exp()) - label);
+      d_left.device(dev) =
+          d_out * (1.0f / (1.0f + (right - left).exp()) - label);
     }
     // compute d_right
     if (d_right_t) {
       d_right_t->mutable_data<T>(ctx.GetPlace());
       auto d_right = framework::EigenVector<T>::Flatten(*d_right_t);
       d_right.device(dev) =
-          -d_out * (1.0 / (1. + (right - left).exp()) - label);
+          -d_out * (1.0f / (1.0f + (right - left).exp()) - label);
     }
   }
 };
diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.h b/paddle/fluid/operators/softmax_with_cross_entropy_op.h
index 93f2552c3cee90a3eb1a948494fb231a41f6f74d..35663bd9b77c234c72c800fb42d1892f94e79acf 100644
--- a/paddle/fluid/operators/softmax_with_cross_entropy_op.h
+++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.h
@@ -23,9 +23,6 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename T>
 class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
@@ -95,12 +92,12 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
     labels_2d.ShareDataWith(*labels).Resize({n, labels->numel() / n});
     out_grad_2d.ShareDataWith(*out_grad).Resize({n, d / axis_dim});
 
-    auto out_grad_mat = EigenMatrix<T>::From(out_grad_2d);
-    auto logit_grad_mat = EigenMatrix<T>::From(logit_grad_2d);
+    auto out_grad_mat = framework::EigenMatrix<T>::From(out_grad_2d);
+    auto logit_grad_mat = framework::EigenMatrix<T>::From(logit_grad_2d);
     auto& place = *context.template device_context<platform::CPUDeviceContext>()
                        .eigen_device();
     if (soft_label) {
-      auto lbl_mat = EigenMatrix<T>::From(labels_2d);
+      auto lbl_mat = framework::EigenMatrix<T>::From(labels_2d);
       logit_grad_mat.device(place) =
           out_grad_mat.broadcast(Eigen::DSizes<int, 2>(1, axis_dim)) *
           (logit_grad_mat - lbl_mat);
diff --git a/paddle/fluid/operators/squared_l2_distance_op.h b/paddle/fluid/operators/squared_l2_distance_op.h
index b8735a69c4b60e493543d7267cba76401e4ce8f6..5472ecaf99c59947ddc09e7d11f073abfa11074f 100644
--- a/paddle/fluid/operators/squared_l2_distance_op.h
+++ b/paddle/fluid/operators/squared_l2_distance_op.h
@@ -20,12 +20,6 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename DeviceContext, typename T>
 class SquaredL2DistanceKernel : public framework::OpKernel<T> {
@@ -41,15 +35,15 @@ class SquaredL2DistanceKernel : public framework::OpKernel<T> {
 
     int cols = in0->numel() / in0_dims[0];
     // reduce dimensions except the first
-    auto x =
-        EigenMatrix<T>::From(*in0, framework::make_ddim({in0_dims[0], cols}));
-    auto y =
-        EigenMatrix<T>::From(*in1, framework::make_ddim({in1_dims[0], cols}));
+    auto x = framework::EigenMatrix<T>::From(
+        *in0, framework::make_ddim({in0_dims[0], cols}));
+    auto y = framework::EigenMatrix<T>::From(
+        *in1, framework::make_ddim({in1_dims[0], cols}));
 
     out0->mutable_data<T>(context.GetPlace());
     out1->mutable_data<T>(context.GetPlace());
-    auto sub_result = EigenMatrix<T>::From(*out0);
-    auto z = EigenVector<T>::Flatten(*out1);
+    auto sub_result = framework::EigenMatrix<T>::From(*out0);
+    auto z = framework::EigenVector<T>::Flatten(*out1);
 
     auto& place =
         *context.template device_context<DeviceContext>().eigen_device();
@@ -88,8 +82,8 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
                  "in scope for operator 'squared_l2_distance_grad'.",
                  framework::GradVarName("Y")));
 
-    auto sub_result = EigenMatrix<T>::From(*in0);
-    auto out_grad = EigenMatrix<T>::From(*in1);
+    auto sub_result = framework::EigenMatrix<T>::From(*in0);
+    auto out_grad = framework::EigenMatrix<T>::From(*in1);
 
     auto x_dims = x_g->dims();
     auto y_dims = y_g->dims();
@@ -106,8 +100,8 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
 
     x_g->mutable_data<T>(context.GetPlace());
     // eigen matrix
-    auto x_grad =
-        EigenMatrix<T>::From(*x_g, framework::make_ddim({x_dims[0], cols}));
+    auto x_grad = framework::EigenMatrix<T>::From(
+        *x_g, framework::make_ddim({x_dims[0], cols}));
     // dimensions are same with subResult
     x_grad.device(eigen_place) = grad_mat;
 
@@ -121,12 +115,12 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
                           sub_result.dimensions()[0], y_dims[0]));
 
     if (sub_result.dimensions()[0] == y_dims[0]) {
-      auto y_grad =
-          EigenMatrix<T>::From(*y_g, framework::make_ddim({y_dims[0], cols}));
+      auto y_grad = framework::EigenMatrix<T>::From(
+          *y_g, framework::make_ddim({y_dims[0], cols}));
       y_grad.device(eigen_place) = -1 * grad_mat;
     } else {
       auto col_sum_res = -1 * (grad_mat.sum(Eigen::array<int, 1>({{0}})));
-      auto y_grad = EigenVector<T>::Flatten(*y_g);
+      auto y_grad = framework::EigenVector<T>::Flatten(*y_g);
       y_grad.device(eigen_place) = col_sum_res;
     }
   }
diff --git a/paddle/fluid/operators/unity_build_rule.cmake b/paddle/fluid/operators/unity_build_rule.cmake
index c59a239c4b429aafe01f5199a8f61639b45ac7e3..d21f6b2d69d84f4894fd74e260bb77ac47781d72 100644
--- a/paddle/fluid/operators/unity_build_rule.cmake
+++ b/paddle/fluid/operators/unity_build_rule.cmake
@@ -307,32 +307,36 @@ register_unity_group(cc
     spp_op.cu.cc
     squeeze_op.cu.cc
     unbind_op.cu.cc
-    unique_op.cu
     unpool_op.cu.cc
     unsqueeze_op.cu.cc)
 register_unity_group(cu
     addmm_op.cu
     affine_channel_op.cu
     allclose_op.cu
-    argsort_op.cu
     assign_value_op.cu
     bce_loss_op.cu
     bernoulli_op.cu
-    bilateral_slice_op.cu)
+    bilateral_slice_op.cu
+    batch_norm_op.cu)
 register_unity_group(cu
     bilinear_tensor_product_op.cu
     bmm_op.cu
     cast_op.cu
     cholesky_op.cu
     clip_by_norm_op.cu
-    clip_op.cu)
+    clip_op.cu
+    conv_cudnn_op.cu
+    affine_grid_op.cu)
 register_unity_group(cu
     center_loss_op.cu
     conv_op.cu
     conv_transpose_cudnn_op.cu
     conv_transpose_op.cu
     cos_sim_op.cu
-    crop_op.cu)
+    crop_op.cu
+    average_accumulates_op.cu
+    conj_op.cu
+    correlation_op.cu)
 register_unity_group(cu
     cross_entropy_op.cu
     cross_op.cu
@@ -349,7 +353,9 @@ register_unity_group(cu
     diag_op.cu
     diag_v2_op.cu
     edit_distance_op.cu
-    erf_op.cu)
+    erf_op.cu
+    meshgrid_op.cu
+    imag_op.cu)
 register_unity_group(cu
     expand_v2_op.cu
     fake_dequantize_op.cu
@@ -377,10 +383,8 @@ register_unity_group(cu
     inplace_abn_op.cu
     interpolate_v2_op.cu
     isfinite_op.cu
-    kron_op.cu
     l1_norm_op.cu
     label_smooth_op.cu
-    layer_norm_op.cu
     linspace_op.cu
     load_combine_op.cu
     load_op.cu)
@@ -388,20 +392,30 @@ register_unity_group(cu
     lod_reset_op.cu
     log_softmax_op.cu
     lrn_op.cu
-    lstm_unit_op.cu)
+    lstm_unit_op.cu
+    dot_op.cu
+    psroi_pool_op.cu
+    rank_loss_op.cu
+    real_op.cu)
 register_unity_group(cu
     log_loss_op.cu
     lookup_table_v2_op.cu
     margin_rank_loss_op.cu
     masked_select_op.cu
-    merge_selected_rows_op.cu)
+    merge_selected_rows_op.cu
+    lstmp_op.cu
+    shuffle_channel_op.cu
+    softmax_cudnn_op.cu
+    squared_l2_distance_op.cu)
 register_unity_group(cu
     conv_shift_op.cu
     dequantize_log_op.cu
     dropout_op.cu
     fake_quantize_op.cu
     gelu_op.cu
-    lookup_table_op.cu)
+    lookup_table_op.cu
+    sigmoid_cross_entropy_with_logits_op.cu
+    softmax_with_cross_entropy_op.cu)
 register_unity_group(cu
     mean_iou_op.cu
     mean_op.cu
@@ -430,7 +444,10 @@ register_unity_group(cu
     random_crop_op.cu
     randperm_op.cu
     range_op.cu
-    reverse_op.cu)
+    reverse_op.cu
+    partial_concat_op.cu
+    kldiv_loss_op.cu
+    instance_norm_op.cu)
 register_unity_group(cu
     roi_align_op.cu
     roll_op.cu
@@ -457,40 +474,42 @@ register_unity_group(cu
     split_op.cu
     split_selected_rows_op.cu
     squared_l2_norm_op.cu
-    stack_op.cu
-    strided_slice_op.cu
     sum_op.cu
-    temporal_shift_op.cu)
+    temporal_shift_op.cu
+    arg_max_op.cu)
 register_unity_group(cu
     row_conv_op.cu
-    tile_op.cu
-    trace_op.cu
-    transpose_op.cu
     tree_conv_op.cu
     tril_triu_op.cu
     truncated_gaussian_random_op.cu
-    unfold_op.cu)
+    unfold_op.cu
+    arg_min_op.cu
+    crop_tensor_op.cu)
 register_unity_group(cu
     smooth_l1_loss_op.cu
     uniform_random_op.cu
-    unique_op.cu
     unstack_op.cu
     where_index_op.cu
-    where_op.cu)
+    where_op.cu
+    layer_norm_op.cu)
+register_unity_group(cu
+    expand_as_op.cu
+    stack_op.cu)
 # The following groups are to make better use of `/MP` which MSVC's parallel
 # compilation instruction when compiling in Unity Build.
 register_unity_group(cu activation_op.cu)
-register_unity_group(cu arg_max_op.cu)
-register_unity_group(cu arg_min_op.cu)
-register_unity_group(cu batch_norm_op.cu)
-register_unity_group(cu crop_tensor_op.cu)
 register_unity_group(cu dist_op.cu)
-register_unity_group(cu expand_as_op.cu)
 register_unity_group(cu expand_as_v2_op.cu)
 register_unity_group(cu gru_unit_op.cu)
-register_unity_group(cu instance_norm_op.cu)
-register_unity_group(cu kldiv_loss_op.cu)
-register_unity_group(cu partial_concat_op.cu)
-register_unity_group(cu softmax_with_cross_entropy_op.cu)
-register_unity_group(cu squared_l2_distance_op.cu)
 register_unity_group(cu top_k_op.cu)
+register_unity_group(cu argsort_op.cu)
+register_unity_group(cu kron_op.cu)
+register_unity_group(cu unique_op.cu)
+register_unity_group(cu tile_op.cu)
+register_unity_group(cu trace_op.cu)
+register_unity_group(cu transpose_op.cu)
+register_unity_group(cu strided_slice_op.cu)
+register_unity_group(cu expand_op.cu)
+register_unity_group(cu matmul_v2_op.cu)
+register_unity_group(cu top_k_v2_op.cu)
+register_unity_group(cu set_value_op.cu)
diff --git a/paddle/scripts/paddle_build.bat b/paddle/scripts/paddle_build.bat
index a6c6a065d2f7bd37bb54b3978201468003508b2d..eb356b5869326d44c07914af4841412075531975 100644
--- a/paddle/scripts/paddle_build.bat
+++ b/paddle/scripts/paddle_build.bat
@@ -262,7 +262,7 @@ echo    ========================================
 echo    Step 2. Buile Paddle ...
 echo    ========================================
 
-for /F %%# in ('wmic cpu get NumberOfLogicalProcessors^|findstr [0-9]') do set /a PARALLEL_PROJECT_COUNT=%%#*9/10
+for /F %%# in ('wmic cpu get NumberOfLogicalProcessors^|findstr [0-9]') do set /a PARALLEL_PROJECT_COUNT=%%#*2/3
 set build_times=1
 :build_tp
 echo Build third_party the %build_times% time: