diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index 373611cc0ee952de813f01d32d1516e1a8384750..e5ff3b2f7ec3740955a3b50d60f54230caac8c16 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -44,6 +44,8 @@ endfunction()
 add_subdirectory(math)
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
 
+cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
+
 cc_library(net_op SRCS net_op.cc DEPS op_registry)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
 
diff --git a/paddle/operators/gather.h b/paddle/operators/gather.h
index 0c73717d38aca9f3430e66cafc3ecccdd2eec776..d6e6990394e46ba06c4bacfe33ca522f3ff1413a 100644
--- a/paddle/operators/gather.h
+++ b/paddle/operators/gather.h
@@ -29,7 +29,7 @@ void CPUGather(const T* params, const int* indices, const int slice_size,
                const int index_size, T* output) {
   const size_t slice_bytes = slice_size * sizeof(T);
 
-  for (size_t i = 0; i < index_size; ++i) {
+  for (int i = 0; i < index_size; ++i) {
     int index_ = indices[i];
     memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes);
   }
@@ -60,7 +60,7 @@ void Gather(const platform::Place& place, const paddle::framework::Tensor* src,
 
   // slice size
   int slice_size = 1;
-  for (size_t i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
+  for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
 
   // Gathering
   if (platform::is_cpu_place(place)) {
diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc
index 5de748ec461e4b1a34b75b57c9cd7d5bc9326059..d24d83f299fdb071e60fa3cc7b223c0228cb29af 100644
--- a/paddle/operators/gather_test.cc
+++ b/paddle/operators/gather_test.cc
@@ -35,7 +35,7 @@ TEST(Gather, GatherData) {
   p_src = src->mutable_data<int>(make_ddim({3, 4}), CPUPlace());
   p_index = index->mutable_data<int>(make_ddim({2}), CPUPlace());
 
-  for (size_t i = 0; i < 12; ++i) p_src[i] = i;
+  for (int i = 0; i < 12; ++i) p_src[i] = i;
   p_index[0] = 1;
   p_index[1] = 0;
 
@@ -43,6 +43,6 @@ TEST(Gather, GatherData) {
 
   Gather<int>(CPUPlace(), src, index, output);
 
-  for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
-  for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
+  for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
+  for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
 }
diff --git a/paddle/operators/scatter.h b/paddle/operators/scatter.h
new file mode 100644
index 0000000000000000000000000000000000000000..6b542675c291607b35f180123cf42fee6a783a85
--- /dev/null
+++ b/paddle/operators/scatter.h
@@ -0,0 +1,92 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+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 <cstring>
+
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
+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>;
+
+// Implementation of CPU copy
+template <typename T>
+void CPUScatterUpdate(const paddle::framework::Tensor* src, const int* index,
+                      const size_t index_size,
+                      paddle::framework::Tensor* output) {
+  paddle::framework::DDim output_dims = output->dims();
+
+  for (size_t i = 0; i < index_size; ++i) {
+    int index_ = index[i];
+
+    paddle::framework::Tensor src_ = *src;
+    paddle::framework::Tensor output_ = *output;
+    if (index_size > 1) src_ = src->Slice<T>(i, i + 1);
+    if (output_dims[0] > 1) output_ = output->Slice<T>(index_, index_ + 1);
+
+    auto X = EigenVector<T>::Flatten(src_);
+    auto Y = EigenVector<T>::Flatten(output_);
+
+    Y = X + Y;
+  }
+}
+
+// Implementation of GPU scatter:
+template <typename T>
+void GPUScatterUpdate(const T* src, const int* index, const int slice_size,
+                      const int index_size, T* output);
+
+/**
+ * Return a updated tensor from source tensor, scattered according to index:
+ * dst[i] += src[index[i]]
+ * input[src]: type-T source Tensor
+ * input[index]: type-int index Tensor (1-D)
+ * return: output tensor
+ */
+template <typename T>
+void ScatterUpdate(const platform::Place& place,
+                   const paddle::framework::Tensor* src,
+                   const paddle::framework::Tensor* index,
+                   paddle::framework::Tensor* output) {
+  // check index of shape 1-D
+  PADDLE_ENFORCE(index->dims().size() == 1);
+  int index_size = index->dims()[0];
+
+  auto src_dims = src->dims();
+  auto dst_dims = output->dims();
+
+  // check src shape and dst shape should match
+  for (int i = 1; i < src_dims.size(); i++)
+    PADDLE_ENFORCE(src_dims[i] == dst_dims[i]);
+
+  // slice size
+  size_t slice_size = 1;
+  for (int i = 0; i < src_dims.size(); ++i) slice_size *= src_dims[i];
+
+  if (platform::is_cpu_place(place)) {
+    CPUScatterUpdate<T>(src, index->data<int>(), index_size, output);
+  } else {
+  }
+}
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/scatter_test.cc b/paddle/operators/scatter_test.cc
new file mode 100644
index 0000000000000000000000000000000000000000..4449ce6564396f1971506efb7458c00c834db19f
--- /dev/null
+++ b/paddle/operators/scatter_test.cc
@@ -0,0 +1,52 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+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. */
+
+#include "paddle/operators/scatter.h"
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
+#include <gtest/gtest.h>
+#include <iostream>
+#include <string>
+
+TEST(scatter, ScatterUpdate) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  using namespace paddle::operators;
+
+  Tensor* src = new Tensor();
+  Tensor* index = new Tensor();
+  Tensor* output = new Tensor();
+
+  float* p_src = nullptr;
+  int* p_index = nullptr;
+  p_src = src->mutable_data<float>(make_ddim({1, 4}), CPUPlace());
+  p_index = index->mutable_data<int>(make_ddim({1}), CPUPlace());
+
+  for (size_t i = 0; i < 4; ++i) p_src[i] = float(i);
+  p_index[0] = 1;
+
+  float* p_output = output->mutable_data<float>(make_ddim({4, 4}), CPUPlace());
+
+  ScatterUpdate<float>(CPUPlace(), src, index, output);
+
+  for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], float(0));
+  for (size_t i = 0; i < 4; ++i) EXPECT_EQ(output->data<float>()[i], float(0));
+  for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], float(i - 4));
+  for (size_t i = 4; i < 8; ++i)
+    EXPECT_EQ(output->data<float>()[i], float(i - 4));
+  for (size_t i = 8; i < 16; ++i) EXPECT_EQ(p_output[i], float(0));
+  for (size_t i = 8; i < 16; ++i) EXPECT_EQ(output->data<float>()[i], float(0));
+}