diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt
index 91101356436c26171eaca2fe01dfd4d937e71717..b0276f4080b3989047f9d181e9d37a18dcfad5fa 100644
--- a/paddle/fluid/operators/math/CMakeLists.txt
+++ b/paddle/fluid/operators/math/CMakeLists.txt
@@ -3,8 +3,8 @@ add_subdirectory(detail)
 endif(NOT WIN32)
 
 function(math_library TARGET)
-    # math_library is a function to create math library. 
-    # The interface is the same as cc_library. 
+    # math_library is a function to create math library.
+    # The interface is the same as cc_library.
     # But it handle split GPU/CPU code and link some common library.
     set(cc_srcs)
     set(cu_srcs)
@@ -53,7 +53,7 @@ cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context)
 math_library(math_function DEPS blas)
 math_library(maxouting)
 math_library(pooling)
-math_library(selected_rows_functor DEPS selected_rows math_function)
+math_library(selected_rows_functor DEPS selected_rows math_function blas)
 math_library(sequence2batch)
 math_library(sequence_padding)
 math_library(sequence_pooling DEPS math_function)
diff --git a/paddle/fluid/operators/math/selected_rows_functor.cc b/paddle/fluid/operators/math/selected_rows_functor.cc
index 8e8baf49b2330e95ff1a868b0b0a03bc10d84484..6810c24227ed074d4ba747399348aa12f55467a0 100644
--- a/paddle/fluid/operators/math/selected_rows_functor.cc
+++ b/paddle/fluid/operators/math/selected_rows_functor.cc
@@ -15,7 +15,6 @@ limitations under the License. */
 #include <set>
 #include <vector>
 
-#include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
 
 namespace paddle {
@@ -150,6 +149,45 @@ template struct SelectedRowsAddTo<platform::CPUDeviceContext, double>;
 template struct SelectedRowsAddTo<platform::CPUDeviceContext, int>;
 template struct SelectedRowsAddTo<platform::CPUDeviceContext, int64_t>;
 
+template <typename T>
+struct SelectedRowsSumTo<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
+                  const std::vector<framework::SelectedRows*>& input1,
+                  const std::vector<int64_t>& input2_offsets,
+                  framework::SelectedRows* input2) {
+    // Ensure all selected rows have the same height
+    size_t size = 0u;
+    for (auto iter = input1.begin(); iter != input1.end(); ++iter) {
+      auto& in_rows = (*iter)->rows();
+      size += in_rows.end() - in_rows.begin();
+      auto in1_height = (*iter)->height();
+      PADDLE_ENFORCE_EQ(in1_height, input2->height());
+    }
+    // concat rows
+    std::vector<int64_t> in2_rows;
+    in2_rows.reserve(in2_rows.size() + size);
+    for (auto iter = input1.begin(); iter != input1.end(); ++iter) {
+      const framework::Vector<int64_t>& in_rows = (*iter)->rows();
+      in2_rows.insert(in2_rows.end(), in_rows.begin(), in_rows.end());
+    }
+    input2->set_rows(in2_rows);
+
+    auto* in2_value = input2->mutable_value();
+    auto* in2_data = in2_value->data<T>();
+    auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
+    size_t offset = 0u;
+    for (size_t i = 0u; i != input1.size(); ++i) {
+      auto& in_value = input1[i]->value();
+      const auto* in_data = in_value.data<T>();
+      offset += input2_offsets[i];
+      blas.VCOPY(in_value.numel(), in_data, in2_data + offset);
+    }
+  }
+};
+
+template struct SelectedRowsSumTo<platform::CPUDeviceContext, float>;
+template struct SelectedRowsSumTo<platform::CPUDeviceContext, double>;
+
 template <typename T>
 struct SelectedRowsAddToTensor<platform::CPUDeviceContext, T> {
   void operator()(const platform::CPUDeviceContext& context,
@@ -208,8 +246,18 @@ struct MergeAdd<platform::CPUDeviceContext, T> {
                   framework::SelectedRows* output) {
     framework::SelectedRows& out = *output;
     auto input_rows = input.rows();
-    std::set<int64_t> row_set(input_rows.begin(), input_rows.end());
-    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
+    std::vector<int64_t> merge_rows;
+    merge_rows.reserve(input_rows.size());
+    std::unordered_map<int64_t, size_t> rows_pos_map;
+    rows_pos_map.reserve(input_rows.size());
+    size_t idx = 0u;
+    for (std::vector<int64_t>::iterator iter = input_rows.begin();
+         iter != input_rows.end(); ++iter) {
+      if (rows_pos_map.find(*iter) == rows_pos_map.end()) {
+        rows_pos_map[*iter] = idx++;
+        merge_rows.emplace_back(*iter);
+      }
+    }
 
     auto input_width = input.value().dims()[1];
     out.set_rows(merge_rows);
@@ -226,7 +274,7 @@ struct MergeAdd<platform::CPUDeviceContext, T> {
     auto* input_data = input.value().data<T>();
 
     for (size_t i = 0; i < input_rows.size(); i++) {
-      size_t out_i = FindPos(merge_rows, input_rows[i]);
+      size_t out_i = rows_pos_map[input_rows[i]];
       for (int64_t j = 0; j < input_width; j++) {
         out_data[out_i * input_width + j] += input_data[i * input_width + j];
       }
@@ -234,8 +282,6 @@ struct MergeAdd<platform::CPUDeviceContext, T> {
   }
 };
 
-template struct MergeAdd<platform::CPUDeviceContext, float>;
-template struct MergeAdd<platform::CPUDeviceContext, double>;
 template struct MergeAdd<platform::CPUDeviceContext, int>;
 template struct MergeAdd<platform::CPUDeviceContext, int64_t>;
 
diff --git a/paddle/fluid/operators/math/selected_rows_functor.h b/paddle/fluid/operators/math/selected_rows_functor.h
index aa419f74fcd2a53cdd734ec270bc154b78c9f2ff..3d99c9b3f2303c941a150f77516703b8103c6c25 100644
--- a/paddle/fluid/operators/math/selected_rows_functor.h
+++ b/paddle/fluid/operators/math/selected_rows_functor.h
@@ -12,8 +12,13 @@ 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. */
 #pragma once
+
+#include <vector>
+
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/device_context.h"
 
 #define INLINE_FOR2(sizei, sizej)     \
@@ -49,6 +54,15 @@ struct SelectedRowsAddTo {
                   const int64_t input2_offset, framework::SelectedRows* input2);
 };
 
+// input2 = [all input in input1] + input2
+template <typename DeviceContext, typename T>
+struct SelectedRowsSumTo {
+  void operator()(const DeviceContext& context,
+                  const std::vector<framework::SelectedRows*>& input1,
+                  const std::vector<int64_t>& input2_offsets,
+                  framework::SelectedRows* input2);
+};
+
 // input2 = input1 + input2
 template <typename DeviceContext, typename T>
 struct SelectedRowsAddToTensor {
@@ -70,6 +84,108 @@ struct MergeAdd {
                   framework::SelectedRows* output);
 };
 
+template <>
+struct MergeAdd<platform::CPUDeviceContext, float> {
+  framework::SelectedRows operator()(const platform::CPUDeviceContext& context,
+                                     const framework::SelectedRows& input) {
+    framework::SelectedRows out;
+    (*this)(context, input, &out);
+    return out;
+  }
+
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::SelectedRows& input,
+                  framework::SelectedRows* output) {
+    framework::SelectedRows& out = *output;
+    auto input_rows = input.rows();
+    std::vector<int64_t> merge_rows;
+    merge_rows.reserve(input_rows.size());
+    std::unordered_map<int64_t, size_t> rows_pos_map;
+    rows_pos_map.reserve(input_rows.size());
+    size_t idx = 0u;
+    for (std::vector<int64_t>::iterator iter = input_rows.begin();
+         iter != input_rows.end(); ++iter) {
+      if (rows_pos_map.find(*iter) == rows_pos_map.end()) {
+        rows_pos_map[*iter] = idx++;
+        merge_rows.emplace_back(*iter);
+      }
+    }
+
+    auto input_width = input.value().dims()[1];
+    out.set_rows(merge_rows);
+    out.set_height(input.height());
+    out.mutable_value()->mutable_data<float>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), input_width}),
+        context.GetPlace());
+
+    math::SetConstant<platform::CPUDeviceContext, float> constant_functor;
+    constant_functor(context, out.mutable_value(), 0.0);
+
+    auto* out_data = out.mutable_value()->data<float>();
+    auto* input_data = input.value().data<float>();
+
+    auto blas = GetBlas<platform::CPUDeviceContext, float>(context);
+    for (size_t i = 0; i < input_rows.size(); i++) {
+      size_t out_i = rows_pos_map[input_rows[i]];
+      float* y = out_data + out_i * input_width;
+      const float* x = input_data + i * input_width;
+      blas.AXPY(input_width, 1., x, y);
+    }
+  }
+};
+
+template <>
+struct MergeAdd<platform::CPUDeviceContext, double> {
+  framework::SelectedRows operator()(const platform::CPUDeviceContext& context,
+                                     const framework::SelectedRows& input) {
+    framework::SelectedRows out;
+    (*this)(context, input, &out);
+    return out;
+  }
+
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::SelectedRows& input,
+                  framework::SelectedRows* output) {
+    framework::SelectedRows& out = *output;
+    auto input_rows = input.rows();
+    std::vector<int64_t> merge_rows;
+    merge_rows.reserve(input_rows.size());
+    std::unordered_map<int64_t, size_t> rows_pos_map;
+    rows_pos_map.reserve(input_rows.size());
+    size_t idx = 0u;
+    for (std::vector<int64_t>::iterator iter = input_rows.begin();
+         iter != input_rows.end(); ++iter) {
+      if (rows_pos_map.find(*iter) == rows_pos_map.end()) {
+        rows_pos_map[*iter] = idx++;
+        merge_rows.emplace_back(*iter);
+      }
+    }
+
+    auto input_width = input.value().dims()[1];
+    out.set_rows(merge_rows);
+    out.set_height(input.height());
+    out.mutable_value()->mutable_data<double>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), input_width}),
+        context.GetPlace());
+
+    math::SetConstant<platform::CPUDeviceContext, double> constant_functor;
+    constant_functor(context, out.mutable_value(), 0.0);
+
+    auto* out_data = out.mutable_value()->data<double>();
+    auto* input_data = input.value().data<double>();
+
+    auto blas = GetBlas<platform::CPUDeviceContext, double>(context);
+    for (size_t i = 0; i < input_rows.size(); i++) {
+      size_t out_i = rows_pos_map[input_rows[i]];
+      double* y = out_data + out_i * input_width;
+      const double* x = input_data + i * input_width;
+      blas.AXPY(input_width, 1., x, y);
+    }
+  }
+};
+
 template <typename DeviceContext, typename T>
 struct Add {
   framework::SelectedRows operator()(const DeviceContext& context,
diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cc b/paddle/fluid/operators/math/selected_rows_functor_test.cc
index 70bed820ee58885861fa8c5535c931f258625572..835589356042b44c9fa5988aed726434fd66910a 100644
--- a/paddle/fluid/operators/math/selected_rows_functor_test.cc
+++ b/paddle/fluid/operators/math/selected_rows_functor_test.cc
@@ -219,3 +219,174 @@ TEST(selected_rows_functor, cpu_add_to) {
   // row9: 2.0 + 3.0
   EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0);
 }
+
+TEST(selected_rows_functor, cpu_merge_add_float) {
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
+                                       float>
+      functor;
+  int64_t height = 10;
+  int64_t row_numel = 10;
+
+  std::vector<int64_t> rows{0, 4, 4, 7};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows{
+      new paddle::framework::SelectedRows(rows, height)};
+  auto* in_value = selected_rows->mutable_value();
+  in_value->mutable_data<float>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in_value, 1.0);
+
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
+
+  paddle::operators::math::scatter::MergeAdd<paddle::platform::CPUDeviceContext,
+                                             float>
+      merge_add_functor;
+  merge_add_functor(ctx, *selected_rows, output.get());
+
+  auto out_height = output->height();
+  EXPECT_EQ(out_height, height);
+
+  auto& out_rows = output->rows();
+  EXPECT_EQ(out_rows[0], 0);
+  EXPECT_EQ(out_rows[1], 4);
+  EXPECT_EQ(out_rows[2], 7);
+
+  auto* out_data = output->value().data<float>();
+
+  EXPECT_EQ(out_data[0 * row_numel], 1.0);
+  EXPECT_EQ(out_data[1 * row_numel], 2.0);
+  EXPECT_EQ(out_data[2 * row_numel], 1.0);
+}
+
+TEST(selected_rows_functor, cpu_merge_add_int) {
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext, int>
+      functor;
+  int64_t height = 10;
+  int64_t row_numel = 10;
+
+  std::vector<int64_t> rows{0, 4, 4, 7};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows{
+      new paddle::framework::SelectedRows(rows, height)};
+  auto* in_value = selected_rows->mutable_value();
+  in_value->mutable_data<int>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in_value, 1);
+
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
+
+  paddle::operators::math::scatter::MergeAdd<paddle::platform::CPUDeviceContext,
+                                             int>
+      merge_add_functor;
+  merge_add_functor(ctx, *selected_rows, output.get());
+
+  auto out_height = output->height();
+  EXPECT_EQ(out_height, height);
+
+  auto& out_rows = output->rows();
+  EXPECT_EQ(out_rows[0], 0);
+  EXPECT_EQ(out_rows[1], 4);
+  EXPECT_EQ(out_rows[2], 7);
+
+  auto* out_data = output->value().data<int>();
+
+  EXPECT_EQ(out_data[0 * row_numel], 1);
+  EXPECT_EQ(out_data[1 * row_numel], 2);
+  EXPECT_EQ(out_data[2 * row_numel], 1);
+}
+TEST(selected_rows_functor, cpu_sum_to) {
+  paddle::platform::CPUPlace cpu_place;
+  paddle::platform::CPUDeviceContext ctx(cpu_place);
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
+                                       float>
+      functor;
+  int64_t height = 10;
+  int64_t row_numel = 10;
+  std::vector<int64_t> rows1{0, 4, 7};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
+      new paddle::framework::SelectedRows(rows1, height)};
+  auto* in1_value = selected_rows1->mutable_value();
+  in1_value->mutable_data<float>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows1.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in1_value, 1.0);
+  std::vector<int64_t> rows2{0, 5, 7, 9};
+  std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
+      new paddle::framework::SelectedRows(rows2, height)};
+  auto* in2_value = selected_rows2->mutable_value();
+  in2_value->mutable_data<float>(
+      paddle::framework::make_ddim(
+          {static_cast<int64_t>(rows2.size()), row_numel}),
+      cpu_place);
+  functor(ctx, in2_value, 2.0);
+  std::unique_ptr<paddle::framework::SelectedRows> output{
+      new paddle::framework::SelectedRows()};
+  output->set_height(height);
+  auto* out_value = output->mutable_value();
+  // simplely concat two SelectedRows
+  out_value->mutable_data<float>(paddle::framework::make_ddim({7, 10}),
+                                 cpu_place);
+  paddle::operators::math::SelectedRowsSumTo<paddle::platform::CPUDeviceContext,
+                                             float>
+      sum_to_functor;
+  sum_to_functor(ctx, std::vector<paddle::framework::SelectedRows*>(
+                          {selected_rows1.get(), selected_rows2.get()}),
+                 std::vector<int64_t>({0, in1_value->numel()}), output.get());
+  auto out_height = output->height();
+  EXPECT_EQ(out_height, height);
+  auto& out_rows = output->rows();
+  // input1 rows
+  EXPECT_EQ(out_rows[0], 0);
+  EXPECT_EQ(out_rows[1], 4);
+  EXPECT_EQ(out_rows[2], 7);
+  // input2 rows
+  EXPECT_EQ(out_rows[3], 0);
+  EXPECT_EQ(out_rows[4], 5);
+  EXPECT_EQ(out_rows[5], 7);
+  EXPECT_EQ(out_rows[6], 9);
+  auto* out_data = output->value().data<float>();
+  // input1 value
+  EXPECT_EQ(out_data[0 * row_numel + 0], 1.0);
+  EXPECT_EQ(out_data[0 * row_numel + 8], 1.0);
+  EXPECT_EQ(out_data[1 * row_numel + 1], 1.0);
+  EXPECT_EQ(out_data[2 * row_numel + 6], 1.0);
+  // input2 value
+  EXPECT_EQ(out_data[3 * row_numel + 3], 2.0);
+  EXPECT_EQ(out_data[3 * row_numel + 8], 2.0);
+  EXPECT_EQ(out_data[4 * row_numel + 4], 2.0);
+  EXPECT_EQ(out_data[5 * row_numel + 7], 2.0);
+  EXPECT_EQ(out_data[6 * row_numel + 9], 2.0);
+  std::unique_ptr<paddle::framework::Tensor> tensor1{
+      new paddle::framework::Tensor()};
+  tensor1->mutable_data<float>(
+      paddle::framework::make_ddim({height, row_numel}), cpu_place);
+  functor(ctx, tensor1.get(), 3.0);
+  paddle::operators::math::SelectedRowsAddToTensor<
+      paddle::platform::CPUDeviceContext, float>
+      add_to_tensor_functor;
+  add_to_tensor_functor(ctx, *output, tensor1.get());
+  auto* tensor1_data = tensor1->data<float>();
+  // row0: 1.0 + 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[0 * row_numel + 0], 6.0);
+  // row1: 3.0
+  EXPECT_EQ(tensor1_data[1 * row_numel + 1], 3.0);
+  // row4 : 1.0 + 3.0
+  EXPECT_EQ(tensor1_data[4 * row_numel + 6], 4.0);
+  // row5: 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[5 * row_numel + 7], 5.0);
+  // row6: 3.0
+  EXPECT_EQ(tensor1_data[6 * row_numel + 1], 3.0);
+  // row7: 1.0 + 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[7 * row_numel + 3], 6.0);
+  // row9: 2.0 + 3.0
+  EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0);
+}