Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into dev_backward_for_op_desc

paddle/framework/backward.cc

@@ -141,9 +141,27 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
         net->ops_[op_offset]->Rename(name, dup_outputs.back());
       }
       // collect all the offset to append `add` op for each alias
+      //
+      // one variable is shared between multiple operators.
+      // insert add operator one by one, then add it to output
+      for (size_t output_idx = 0; output_idx < dup_outputs.size() - 1;
+           ++output_idx) {
+        auto insert_add_x = dup_outputs[output_idx];
+        auto insert_add_y = dup_outputs[output_idx];
+        auto insert_add_out = name + "@SHARED@" + std::to_string(output_idx);
+        // first add op inserted
+        if (output_idx == dup_outputs.size() - 2) {
+          insert_add_out = name;
+        }
+        if (output_idx != 0) {
+          insert_add_y = name + "@SHARED@" + std::to_string(output_idx - 1);
+        }
         insert_position.push_back(
-          {dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}},
-                                               {{"Out", {name}}}, {})});
+            {dup_op.back(),
+             OpRegistry::CreateOp(
+                 "sum", {{"X", {insert_add_x}}, {"X", {insert_add_y}}},
+                 {{"Out", {insert_add_out}}}, {})});
+      }
     }
 
     // make sure the inserted `add` ops follow the BFS order.

paddle/framework/backward_test.cc

@@ -133,15 +133,18 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-class AddOpMaker : public OpProtoAndCheckerMaker {
+class SumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+  SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "x").AsDuplicable();
-    AddOutput("Out", "out");
+    AddInput("X", "the input tensors of sum operator.")
+        .AsDuplicable()
+        .NotInGradient();
+    AddOutput("Out", "the output tensor of sum operator.").NotInGradient();
     AddComment("");
   }
 };
+
 }  // namespace framework
 }  // namespace paddle
 
@@ -154,7 +157,7 @@ REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
 REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
-REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
+REGISTER_OP(sum, f::NOP, f::SumOpMaker, sum_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
 REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
             f::NOP);
@@ -283,7 +286,7 @@ TEST(Backward, net_shared_weight) {
   ASSERT_TRUE(bwd->IsNetOp());
   auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
   ASSERT_EQ(3UL, bwd_net->ops_.size());
-  ASSERT_EQ("add", bwd_net->ops_[2]->Type());
+  ASSERT_EQ("sum", bwd_net->ops_[2]->Type());
 }
 
 TEST(Backward, op_register_grad_not_for_network) {

paddle/operators/cond_op.cc

@@ -126,8 +126,7 @@ void CondOp::PrepareDataForSubnet(
       dim[0] = index_tensors[i].dims()[0];
       tensor_child->mutable_data<float>(dim, platform::CPUPlace());
 
-      Gather<float>(dev_ctx.GetPlace(), tensor_parent, &index_tensors[i],
-                    tensor_child);
+      CPUGather<float>(dev_ctx, *tensor_parent, index_tensors[i], tensor_child);
     }
   }
 
@@ -188,7 +187,7 @@ void CondOp::MergeDataFromSubnet(const framework::Scope& scope,
       Variable* var_child = sub_scopes[i]->FindVar(output);
       PADDLE_ENFORCE_NOT_NULL(var_child);
       auto* tensor_child = &var_child->Get<LoDTensor>();
-      ScatterUpdate<float>(dev_ctx.GetPlace(), tensor_child, &index_tensors[i],
+      ScatterAssign<float>(dev_ctx, *tensor_child, index_tensors[i],
                            tensor_parent);
     }
   }

paddle/operators/gather.cu.h

+/* 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 "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+using platform::Place;
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                              \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
+       i += blockDim.x * gridDim.x)
+
+template <typename T>
+__global__ void GatherCUDAKernel(const T* params, const int* indices, T* output,
+                                 size_t index_size, size_t slice_size) {
+  CUDA_1D_KERNEL_LOOP(i, index_size * slice_size) {
+    int indices_i = i / slice_size;
+    int slice_i = i - indices_i * slice_size;  // offset inside the slice
+    int gather_i = indices[indices_i];
+    int params_i = gather_i * slice_size + slice_i;
+    *(output + i) = *(params + params_i);
+  }
+}
+
+/**
+ * A thin wrapper on gpu tensor
+ * Return a new tensor from source tensor, gathered according to index
+ * input[src]: type-T source Tensor
+ * input[index]: type-int index Tensor (1-D)
+ * return: output tensor
+ */
+template <typename T>
+void GPUGather(const platform::DeviceContext& ctx, const Tensor& src,
+               const Tensor& index, Tensor* output) {
+  // PADDLE_ENFORCE(platform::is_gpu_place(place));
+  // check index of shape 1-D
+  PADDLE_ENFORCE(index.dims().size() == 1);
+  int index_size = index.dims()[0];
+
+  auto src_dims = src.dims();
+  framework::DDim output_dims(src_dims);
+  output_dims[0] = index_size;
+
+  // slice size
+  int slice_size = 1;
+  for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
+
+  const T* p_src = src.data<T>();
+  const int* p_index = index.data<int>();
+  T* p_output = output->data<T>();
+
+  int block = 512;
+  int n = slice_size * index_size;
+  int grid = (n + block - 1) / block;
+
+  GatherCUDAKernel<T><<<
+      grid, block, 0,
+      reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream()>>>(
+      p_src, p_index, p_output, index_size, slice_size);
+}
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/gather.h

@@ -24,49 +24,40 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-// Implementation of CPU copy
-template <typename T>
-void CPUGather(const T* src, const int* indices, const int slice_size,
-               const int index_size, T* output) {
-  const size_t slice_bytes = slice_size * sizeof(T);
-
-  for (int i = 0; i < index_size; ++i) {
-    int index_ = indices[i];
-    memcpy(output + i * slice_size, src + index_ * slice_size, slice_bytes);
-  }
-}
-
-// Implementation of GPU copy:
-template <typename T>
-void GPUGather(const T* src, const int* index, const int slice_size,
-               const int index_size, T* output);
+using framework::Tensor;
 
 /**
+ * A thin wrapper for gathering on cpu tensor
  * Return a new tensor from source tensor, gathered according to index
  * input[src]: type-T source Tensor
  * input[index]: type-int index Tensor (1-D)
  * return: output tensor
  */
 template <typename T>
-void Gather(const platform::Place& place, const paddle::framework::Tensor* src,
-            const paddle::framework::Tensor* index,
-            paddle::framework::Tensor* output) {
+void CPUGather(const platform::DeviceContext& ctx, const Tensor& src,
+               const Tensor& index, Tensor* output) {
+  PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index->dims().size() == 1);
-  int index_size = index->dims()[0];
+  PADDLE_ENFORCE(index.dims().size() == 1);
+  int index_size = index.dims()[0];
 
-  auto src_dims = src->dims();
+  auto src_dims = src.dims();
   framework::DDim output_dims(src_dims);
   output_dims[0] = index_size;
 
+  const T* p_src = src.data<T>();
+  const int* p_index = index.data<int>();
+  T* p_output = output->data<T>();
+
   // slice size
   int slice_size = 1;
   for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
 
-  // Gathering
-  if (platform::is_cpu_place(place)) {
-    CPUGather<T>(src->data<T>(), index->data<int>(), slice_size, index_size,
-                 output->data<T>());
+  const size_t slice_bytes = slice_size * sizeof(T);
+
+  for (int i = 0; i < index_size; ++i) {
+    int index_ = p_index[i];
+    memcpy(p_output + i * slice_size, p_src + index_ * slice_size, slice_bytes);
   }
 }
 

paddle/operators/gather_op.cc

@@ -31,6 +31,8 @@ class GatherOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of GatherOp should not be null.");
 
+    auto index_dims = ctx->GetInputDim("Index");
+    PADDLE_ENFORCE(index_dims.size() == 1);
     int batch_size = ctx->GetInputDim("Index")[0];
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
     framework::DDim output_dims(ctx->GetInputDim("X"));
@@ -79,8 +81,5 @@ Out = X[Index]
 namespace ops = paddle::operators;
 REGISTER_OP(gather, ops::GatherOp, ops::GatherOpMaker, gather_grad,
             ops::GatherGradOp);
-REGISTER_OP_CPU_KERNEL(gather,
-                       ops::GatherOpKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(
-    gather_grad,
-    ops::GatherGradientOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(gather, ops::GatherOpKernel<float>);
+REGISTER_OP_CPU_KERNEL(gather_grad, ops::GatherGradientOpKernel<float>);

paddle/operators/gather_op.cu

+/* 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 "gather.cu.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/operators/gather_op.h"
+#include "scatter.cu.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class GatherOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "This kernel only runs on GPU device.");
+    auto *x = ctx.Input<Tensor>("X");
+    auto *index = ctx.Input<Tensor>("Index");
+    auto *output = ctx.Output<Tensor>("Out");
+
+    output->mutable_data<T>(ctx.GetPlace());
+
+    GPUGather<T>(ctx.device_context(), *x, *index, output);
+  }
+};
+
+template <typename T>
+class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "This kernel only runs on GPU device.");
+    auto *Index = ctx.Input<Tensor>("Index");
+    auto *dX = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *dO = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto *x = ctx.Input<Tensor>("X");
+
+    dX->mutable_data<T>(ctx.GetPlace());
+    auto dxt = framework::EigenVector<T>::Flatten(*dX);
+    auto place = ctx.GetEigenDevice<platform::GPUPlace>();
+    dxt.device(place) = dxt.constant(static_cast<T>(0));
+
+    GPUScatterAssign<T>(ctx.device_context(), *dO, *Index, dX);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(gather, ops::GatherOpCUDAKernel<float>);
+REGISTER_OP_GPU_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel<float>);

paddle/operators/gather_op.h

@@ -23,29 +23,40 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
+template <typename T>
 class GatherOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
-    auto *X = ctx.Input<Tensor>("X");
-    auto *Index = ctx.Input<Tensor>("Index");
-    auto *Y = ctx.Output<Tensor>("Out");
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "This kernel only runs on CPU.");
+
+    auto *x = ctx.Input<Tensor>("X");
+    auto *index = ctx.Input<Tensor>("Index");
+    auto *output = ctx.Output<Tensor>("Out");
+
+    output->mutable_data<T>(ctx.GetPlace());
 
-    Y->mutable_data<T>(ctx.GetPlace());
-    Gather<T>(ctx.GetPlace(), X, Index, Y);
+    CPUGather<T>(ctx.device_context(), *x, *index, output);
   }
 };
 
-template <typename Place, typename T>
+template <typename T>
 class GatherGradientOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "This kernel only runs on CPU.");
+
     auto *Index = ctx.Input<Tensor>("Index");
     auto *dX = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto *dO = ctx.Input<Tensor>(framework::GradVarName("Out"));
 
     dX->mutable_data<T>(ctx.GetPlace());
-    ScatterUpdate<T>(ctx.GetPlace(), dO, Index, dX);
+    auto dxt = framework::EigenVector<T>::Flatten(*dX);
+    auto place = ctx.GetEigenDevice<platform::CPUPlace>();
+    dxt.device(place) = dxt.constant(static_cast<T>(0));
+
+    ScatterAssign<T>(ctx.device_context(), *dO, *Index, dX);
   }
 };
 

paddle/operators/gather_test.cc

@@ -41,7 +41,9 @@ TEST(Gather, GatherData) {
 
   int* p_output = output->mutable_data<int>(make_ddim({2, 4}), CPUPlace());
 
-  Gather<int>(CPUPlace(), src, index, output);
+  auto* cpu_place = new paddle::platform::CPUPlace();
+  paddle::platform::CPUDeviceContext ctx(*cpu_place);
+  CPUGather<int>(ctx, *src, *index, output);
 
   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);

paddle/operators/scatter.cu.h

+/* 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 "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                              \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
+       i += blockDim.x * gridDim.x)
+
+template <typename T>
+__global__ void ScatterCUDAKernel(const T* params, const int* indices,
+                                  T* output, size_t index_size,
+                                  size_t slice_size) {
+  CUDA_1D_KERNEL_LOOP(i, index_size * slice_size) {
+    int indices_i = i / slice_size;
+    int slice_i = i - indices_i * slice_size;  // offset inside the slice
+    int scatter_i = indices[indices_i];
+    int out_i = scatter_i * slice_size + slice_i;
+    *(output + out_i) = *(params + i);
+  }
+}
+
+/**
+ * A thin wrapper on gpu tensor
+ * Return a new updated tensor from source tensor, scatter-assigned according to
+ * index
+ * input[src]: type-T source Tensor
+ * input[index]: type-int index Tensor (1-D)
+ * return: output tensor
+ */
+template <typename T>
+void GPUScatterAssign(const platform::DeviceContext& ctx, const Tensor& src,
+                      const Tensor& index, Tensor* output) {
+  // PADDLE_ENFORCE(platform::is_gpu_place(place));
+  // check index of shape 1-D
+  PADDLE_ENFORCE(index.dims().size() == 1);
+  int index_size = index.dims()[0];
+
+  auto src_dims = src.dims();
+  framework::DDim output_dims(src_dims);
+  output_dims[0] = index_size;
+
+  // slice size
+  int slice_size = 1;
+  for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
+
+  const T* p_src = src.data<T>();
+  const int* p_index = index.data<int>();
+  T* p_output = output->data<T>();
+
+  int block = 512;
+  int n = slice_size * index_size;
+  int grid = (n + block - 1) / block;
+
+  ScatterCUDAKernel<T><<<
+      grid, block, 0,
+      reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream()>>>(
+      p_src, p_index, p_output, index_size, slice_size);
+}
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/scatter.h

@@ -24,63 +24,42 @@ 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]]
+ * 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) {
+void ScatterAssign(const platform::DeviceContext& ctx, const Tensor& src,
+                   const Tensor& index, Tensor* output) {
+  PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index->dims().size() == 1);
-  int index_size = index->dims()[0];
+  PADDLE_ENFORCE(index.dims().size() == 1);
+  int index_size = index.dims()[0];
 
-  auto src_dims = src->dims();
+  auto src_dims = src.dims();
   auto dst_dims = output->dims();
 
+  const T* p_src = src.data<T>();
+  const int* p_index = index.data<int>();
+  T* p_output = output->data<T>();
+
   // 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]);
 
-  if (platform::is_cpu_place(place)) {
-    CPUScatterUpdate<T>(src, index->data<int>(), index_size, output);
-  } else {
+  // slice size
+  size_t slice_size = 1;
+  for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
+
+  const size_t slice_bytes = slice_size * sizeof(T);
+
+  for (int i = 0; i < index_size; ++i) {
+    int index_ = p_index[i];
+    memcpy(p_output + index_ * slice_size, p_src + i * slice_size, slice_bytes);
   }
 }
 

paddle/operators/scatter_op.cc

@@ -97,8 +97,5 @@ Out[Index] = Ref[Index] + Updates
 namespace ops = paddle::operators;
 REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, scatter_grad,
             ops::ScatterGradOp);
-REGISTER_OP_CPU_KERNEL(scatter,
-                       ops::ScatterOpKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(
-    scatter_grad,
-    ops::ScatterGradientOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(scatter, ops::ScatterOpKernel<float>);
+REGISTER_OP_CPU_KERNEL(scatter_grad, ops::ScatterGradientOpKernel<float>);

paddle/operators/scatter_op.cu

+/* 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 "gather.cu.h"
+#include "paddle/operators/gather_op.h"
+#include "scatter.cu.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class ScatterOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "This kernel only runs on GPU device.");
+    auto *Ref = ctx.Input<Tensor>("Ref");
+    auto *Index = ctx.Input<Tensor>("Index");
+    auto *Updates = ctx.Input<Tensor>("Updates");
+    auto *Out = ctx.Output<Tensor>("Out");
+
+    Out->ShareDataWith<T>(*Ref);
+
+    GPUScatterAssign<T>(ctx.device_context(), *Updates, *Index, Out);
+  }
+};
+
+template <typename T>
+class ScatterGradOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "This kernel only runs on GPU device.");
+    auto *dRef = ctx.Output<Tensor>(framework::GradVarName("Ref"));
+    auto *dUpdates = ctx.Output<Tensor>(framework::GradVarName("Updates"));
+    auto *Index = ctx.Input<Tensor>("Index");
+    auto *dOut = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    // In place gradient: dRef = dO
+    dRef->ShareDataWith<T>(*dOut);
+    dUpdates->mutable_data<T>(ctx.GetPlace());
+    // Gradient by Gather: dUpdates = dO[Index]
+    GPUGather<T>(ctx.device_context(), *dOut, *Index, dUpdates);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(scatter, ops::ScatterOpCUDAKernel<float>);
+REGISTER_OP_GPU_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel<float>);

paddle/operators/scatter_op.h

@@ -23,10 +23,12 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
+template <typename T>
 class ScatterOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "This kernel only runs on CPU.");
     auto *Ref = ctx.Input<Tensor>("Ref");
     auto *Index = ctx.Input<Tensor>("Index");
     auto *Updates = ctx.Input<Tensor>("Updates");
@@ -35,14 +37,16 @@ class ScatterOpKernel : public framework::OpKernel<T> {
     // In place output: Out = Ref, Out[Index] += Updates
     Out->ShareDataWith<T>(*Ref);
     // Apply ScatterUpdate: Out[index] += Updates[:]
-    ScatterUpdate<T>(ctx.GetPlace(), Updates, Index, Out);
+    ScatterAssign<T>(ctx.device_context(), *Updates, *Index, Out);
   }
 };
 
-template <typename Place, typename T>
+template <typename T>
 class ScatterGradientOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "This kernel only runs on CPU.");
     auto *dRef = ctx.Output<Tensor>(framework::GradVarName("Ref"));
     auto *dUpdates = ctx.Output<Tensor>(framework::GradVarName("Updates"));
     auto *Index = ctx.Input<Tensor>("Index");
@@ -52,7 +56,7 @@ class ScatterGradientOpKernel : public framework::OpKernel<T> {
     dRef->ShareDataWith<T>(*dOut);
     dUpdates->mutable_data<T>(ctx.GetPlace());
     // Gradient by Gather: dUpdates += dO[Index]
-    Gather<T>(ctx.GetPlace(), dOut, Index, dUpdates);
+    CPUGather<T>(ctx.device_context(), *dOut, *Index, dUpdates);
   }
 };
 

paddle/operators/scatter_test.cc

@@ -40,7 +40,9 @@ TEST(scatter, ScatterUpdate) {
 
   float* p_output = output->mutable_data<float>(make_ddim({4, 4}), CPUPlace());
 
-  ScatterUpdate<float>(CPUPlace(), src, index, output);
+  auto* cpu_place = new paddle::platform::CPUPlace();
+  paddle::platform::CPUDeviceContext ctx(*cpu_place);
+  ScatterAssign<float>(ctx, *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));

python/paddle/v2/framework/tests/test_scatter_op.py

@@ -10,7 +10,7 @@ class TestScatterOp(OpTest):
         index_np = np.array([1, 2]).astype("int32")
         updates_np = np.random.random((2, 3)).astype("float32")
         output_np = np.copy(ref_np)
-        output_np[index_np] += updates_np
+        output_np[index_np] = updates_np
         self.inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np}
         self.outputs = {'Out': output_np}
 
@@ -18,7 +18,7 @@ class TestScatterOp(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['Updates', 'Ref'], 'Out', in_place=True)
+        self.check_grad(['Updates'], 'Out', in_place=True)
 
 
 if __name__ == "__main__":