diff --git a/paddle/operators/math/selected_rows_functor.cc b/paddle/operators/math/selected_rows_functor.cc
index 075196b47eeaf118a588b96532d87a05e4e600c6..514f2adef284c8877e2e74b943b4e6419c6ae721 100644
--- a/paddle/operators/math/selected_rows_functor.cc
+++ b/paddle/operators/math/selected_rows_functor.cc
@@ -145,6 +145,8 @@ struct SelectedRowsAddTo<platform::CPUPlace, T> {
 
 template struct SelectedRowsAddTo<platform::CPUPlace, float>;
 template struct SelectedRowsAddTo<platform::CPUPlace, double>;
+template struct SelectedRowsAddTo<platform::CPUPlace, int>;
+template struct SelectedRowsAddTo<platform::CPUPlace, int64_t>;
 
 template <typename T>
 struct SelectedRowsAddToTensor<platform::CPUPlace, T> {
@@ -175,6 +177,8 @@ struct SelectedRowsAddToTensor<platform::CPUPlace, T> {
 
 template struct SelectedRowsAddToTensor<platform::CPUPlace, float>;
 template struct SelectedRowsAddToTensor<platform::CPUPlace, double>;
+template struct SelectedRowsAddToTensor<platform::CPUPlace, int>;
+template struct SelectedRowsAddToTensor<platform::CPUPlace, int64_t>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/selected_rows_functor.cu b/paddle/operators/math/selected_rows_functor.cu
index 47fe3b44a50fee9f41ae807793187258159b9f29..c40649e55ef93dec852ff6949b5cb134495e4ebf 100644
--- a/paddle/operators/math/selected_rows_functor.cu
+++ b/paddle/operators/math/selected_rows_functor.cu
@@ -173,6 +173,8 @@ struct SelectedRowsAddTo<platform::GPUPlace, T> {
 
 template struct SelectedRowsAddTo<platform::GPUPlace, float>;
 template struct SelectedRowsAddTo<platform::GPUPlace, double>;
+template struct SelectedRowsAddTo<platform::GPUPlace, int>;
+template struct SelectedRowsAddTo<platform::GPUPlace, int64_t>;
 
 namespace {
 template <typename T, int block_size>
@@ -223,6 +225,8 @@ struct SelectedRowsAddToTensor<platform::GPUPlace, T> {
 
 template struct SelectedRowsAddToTensor<platform::GPUPlace, float>;
 template struct SelectedRowsAddToTensor<platform::GPUPlace, double>;
+template struct SelectedRowsAddToTensor<platform::GPUPlace, int>;
+template struct SelectedRowsAddToTensor<platform::GPUPlace, int64_t>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
index c2b7632b2865a3ef66051d815d7722a08c6a8cbd..ddc210c26e69566fef9baa20f49ba1052e993b3f 100644
--- a/paddle/operators/sum_op.cc
+++ b/paddle/operators/sum_op.cc
@@ -176,4 +176,6 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                   ops::SumOpVarTypeInference);
 REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel<paddle::platform::CPUPlace, float>,
-                       ops::SumKernel<paddle::platform::CPUPlace, double>);
+                       ops::SumKernel<paddle::platform::CPUPlace, double>,
+                       ops::SumKernel<paddle::platform::CPUPlace, int>,
+                       ops::SumKernel<paddle::platform::CPUPlace, int64_t>);
diff --git a/paddle/operators/sum_op.cu b/paddle/operators/sum_op.cu
index 5cf05b876b6d6a2ce61d9e10b7ec52ed3cef57d7..5c30dd4d470c2e0acecef18524a4a81f9eb786a9 100644
--- a/paddle/operators/sum_op.cu
+++ b/paddle/operators/sum_op.cu
@@ -14,4 +14,6 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(sum, ops::SumKernel<paddle::platform::GPUPlace, float>,
-                       ops::SumKernel<paddle::platform::GPUPlace, double>);
+                       ops::SumKernel<paddle::platform::GPUPlace, double>,
+                       ops::SumKernel<paddle::platform::GPUPlace, int>,
+                       ops::SumKernel<paddle::platform::GPUPlace, int64_t>);
diff --git a/paddle/platform/cuda_helper.h b/paddle/platform/cuda_helper.h
index a7d99cde106a0a66f122a8c43f49717c03e60dec..376bb0e6887c797c3c1019e92f738a62d01a9c51 100644
--- a/paddle/platform/cuda_helper.h
+++ b/paddle/platform/cuda_helper.h
@@ -31,6 +31,16 @@ constexpr int PADDLE_CUDA_NUM_THREADS = 512;
 
 // For atomicAdd.
 USE_CUDA_ATOMIC(Add, float);
+USE_CUDA_ATOMIC(Add, int);
+USE_CUDA_ATOMIC(Add, unsigned int);
+USE_CUDA_ATOMIC(Add, unsigned long long int);
+
+CUDA_ATOMIC_WRAPPER(Add, int64_t) {
+  static_assert(sizeof(int64_t) == sizeof(long long int),
+                "long long should be int64");
+  return CudaAtomicAdd(reinterpret_cast<unsigned long long int*>(address),
+                       static_cast<unsigned long long int>(val));
+}
 
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600
 USE_CUDA_ATOMIC(Add, double);