Merge branch 'develop' of github.com:baidu/Paddle into feature/complete_variable_bind

paddle/framework/backward.cc

@@ -302,7 +302,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
     return grad_op_descs;  // empty vector
   }
 
-  grad_op_descs = OpRegistry::CreateGradOpDescs(*op_desc);
+  grad_op_descs = OpRegistry::CreateGradOpDescs(op_desc.get());
 
   std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
   for (auto& desc : grad_op_descs) {

paddle/framework/backward_test.cc

@@ -58,6 +58,8 @@ class MulOpMaker : public OpProtoAndCheckerMaker {
     AddInput("X", "A");
     AddInput("Y", "B");
     AddOutput("Out", "Out");
+    AddAttr<int>("x_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
+    AddAttr<int>("y_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
     AddComment("Mul");
   }
 };
@@ -440,6 +442,28 @@ TEST(Backward, simple_single_op) {
             std::vector<std::string>({f::GradVarName("b")}));
 }
 
+TEST(Backward, default_attribute) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op = block->AppendOp();
+  op->SetType("mul");
+  op->SetInput("X", {"x"});
+  op->SetInput("Y", {"y"});
+  op->SetOutput("Out", {"out"});
+
+  AppendBackward(program, {});
+
+  ASSERT_EQ(block->AllOps().size(), 2UL);
+  EXPECT_EQ(boost::get<int>(op->GetAttr("x_num_col_dims")), 1);
+  EXPECT_EQ(boost::get<int>(op->GetAttr("y_num_col_dims")), 1);
+
+  f::OpDescBind *grad_op = block->AllOps()[1];
+  ASSERT_EQ(grad_op->Type(), "mul_grad");
+  EXPECT_EQ(boost::get<int>(grad_op->GetAttr("x_num_col_dims")), 1);
+  EXPECT_EQ(boost::get<int>(grad_op->GetAttr("y_num_col_dims")), 1);
+}
+
 TEST(Backward, simple_mult_op) {
   f::ProgramDesc *program_desc = GetNewProgramDesc();
   f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);

paddle/framework/block_desc.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <deque>
+#include <memory>
 #include <unordered_map>
 #include <vector>
 #include "paddle/framework/op_desc.h"

paddle/framework/data_type.h

@@ -28,7 +28,6 @@ inline DataType ToDataType(std::type_index type) {
     return DataType::INT32;
   } else {
     PADDLE_THROW("Not supported");
-    return static_cast<DataType>(-1);
   }
 }
 

paddle/framework/framework.proto

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 syntax = "proto2";
+option optimize_for = LITE_RUNTIME;
 package paddle.framework;
 
 enum AttrType {

paddle/framework/op_desc.cc

@@ -25,6 +25,7 @@ OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs,
   inputs_ = inputs;
   outputs_ = outputs;
   attrs_ = attrs;
+  need_update_ = true;
 }
 
 OpDesc *OpDescBind::Proto() {

paddle/framework/op_desc.h

@@ -52,8 +52,6 @@ class OpDescBind {
   void SetOutput(const std::string &param_name,
                  const std::vector<std::string> &args);
 
-  std::string DebugString() { return this->Proto()->DebugString(); }
-
   bool HasAttr(const std::string &name) const {
     return attrs_.find(name) != attrs_.end();
   }
@@ -97,6 +95,11 @@ class OpDescBind {
 
   const VariableNameMap &Outputs() const { return outputs_; }
 
+  AttributeMap *MutableAttrMap() {
+    this->need_update_ = true;
+    return &this->attrs_;
+  }
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {

paddle/framework/op_registry.cc

@@ -60,9 +60,14 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDescBind& op_desc) {
 }
 
 std::vector<std::unique_ptr<OpDescBind>> OpRegistry::CreateGradOpDescs(
-    const OpDescBind& op_desc) {
-  auto& info = OpInfoMap::Instance().Get(op_desc.Type());
-  return info.grad_op_maker_(op_desc);
+    OpDescBind* op_desc) {
+  auto& info = OpInfoMap::Instance().Get(op_desc->Type());
+
+  if (info.Checker() != nullptr) {
+    info.Checker()->Check(*op_desc->MutableAttrMap());
+  }
+
+  return info.grad_op_maker_(*op_desc);
 }
 
 }  // namespace framework

paddle/framework/op_registry.h

@@ -80,7 +80,7 @@ class OpRegistry {
   static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
   static std::vector<std::unique_ptr<OpDescBind>> CreateGradOpDescs(
-      const OpDescBind& op_desc);
+      OpDescBind* op_desc);
 
   static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
 };

paddle/framework/program_desc.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
 #include <vector>
 #include "paddle/framework/framework.pb.h"
 #include "paddle/platform/macros.h"
@@ -31,8 +32,6 @@ class ProgramDescBind {
 
   BlockDescBind *Block(size_t idx) { return blocks_[idx].get(); }
 
-  std::string DebugString() { return Proto()->DebugString(); }
-
   size_t Size() const { return blocks_.size(); }
 
   ProgramDesc *Proto();

paddle/framework/tensor.h

@@ -95,6 +95,19 @@ class Tensor {
   template <typename T>
   inline void CopyFrom(const Tensor& src, const platform::Place& dst_place);
 
+  /**
+   * @brief   Copy the content of an external vector to a tensor.
+   *
+   * @param[in] src   The external vector.
+   * @param[in] ctx   The device context contains place where to store.
+   *
+   * * @note    CopyFromVector assumes that the tensor has been resized
+   *            before invoking.
+   */
+  template <typename T>
+  inline void CopyFromVector(const std::vector<T>& src,
+                             const platform::Place& dst_place);
+
   /**
    * @brief   Return the slice of the tensor.
    *

paddle/framework/tensor_impl.h

@@ -123,6 +123,29 @@ inline void Tensor::CopyFrom(const Tensor& src,
 #endif
 }
 
+template <typename T>
+inline void Tensor::CopyFromVector(const std::vector<T>& src,
+                                   const platform::Place& dst_place) {
+  auto src_ptr = static_cast<const void*>(src.data());
+  platform::CPUPlace src_place;
+  auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
+  auto size = src.size() * sizeof(T);
+
+  if (platform::is_cpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size);
+  }
+#ifdef PADDLE_WITH_CUDA
+  else if (platform::is_gpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size, 0);
+  }
+  PADDLE_ENFORCE(cudaStreamSynchronize(0),
+                 "cudaStreamSynchronize failed in Tensor CopyFromVector");
+
+#endif
+}
+
 template <typename T>
 inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   check_memory_size<T>();

paddle/framework/tensor_test.cc

@@ -263,6 +263,93 @@ TEST(Tensor, CopyFrom) {
 #endif
 }
 
+TEST(Tensor, CopyFromVector) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Compare Tensors
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    cpu_ptr = cpu_tensor.data<int>();
+    src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    delete cpu_place;
+  }
+
+#ifdef PADDLE_WITH_CUDA
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+    Tensor gpu_tensor;
+    Tensor dst_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Copy to GPUTensor
+    gpu_tensor.Resize(make_ddim({3, 3}));
+    auto gpu_place = new paddle::platform::GPUPlace();
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    // Copy from GPU to CPU tensor for comparison
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    // Compare Tensors
+    const int* src_ptr = src_vec.data();
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    gpu_tensor.Resize(make_ddim({2, 2}));
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    src_ptr = src_vec.data();
+    cpu_ptr = cpu_tensor.data<int>();
+    dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    delete cpu_place;
+    delete gpu_place;
+  }
+#endif
+}
+
 TEST(Tensor, ReshapeToMatrix) {
   using namespace paddle::framework;
   using namespace paddle::platform;

paddle/framework/type_defs.h

@@ -15,6 +15,7 @@
 #pragma once
 #include <functional>
 #include <map>
+#include <memory>
 #include "paddle/platform/variant.h"
 
 namespace paddle {

paddle/math/tests/test_GpuProfiler.cpp

@@ -162,4 +162,4 @@ int main(int argc, char** argv) {
   return RUN_ALL_TESTS();
 }
 
-#endif /* PADDLE_ONLY_CPU */
+#endif

paddle/memory/detail/buddy_allocator.cc

@@ -182,7 +182,7 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
       max_chunk_size_ = platform::GpuMaxChunkSize();
     }
   }
-#endif  // PADDLE_ONLY_CPU
+#endif
 
   // Allocate a new maximum sized block
   size_t index = 0;

paddle/memory/detail/system_allocator.cc

@@ -134,7 +134,7 @@ void GPUAllocator::Free(void* p, size_t size, size_t index) {
 
 bool GPUAllocator::UseGpu() const { return true; }
 
-#endif  // PADDLE_ONLY_CPU
+#endif
 
 }  // namespace detail
 }  // namespace memory

paddle/memory/detail/system_allocator.h

@@ -51,7 +51,7 @@ class GPUAllocator : public SystemAllocator {
   size_t gpu_alloc_size_ = 0;
   size_t fallback_alloc_size_ = 0;
 };
-#endif  // PADDLE_ONLY_CPU
+#endif
 
 }  // namespace detail
 }  // namespace memory

paddle/memory/detail/system_allocator_test.cc

@@ -62,4 +62,4 @@ TEST(GPUAllocator, Alloc) {
   TestAllocator(a, 2048);
   TestAllocator(a, 0);
 }
-#endif  // PADDLE_ONLY_CPU
+#endif

paddle/memory/memcpy.cc

@@ -89,7 +89,7 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
   platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToDevice);
 }
 
-#endif  // PADDLE_ONLY_CPU
+#endif
 
 }  // namespace memory
 }  // namespace paddle

paddle/memory/memcpy.h

@@ -53,7 +53,7 @@ template <typename DstPlace, typename SrcPlace>
 void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num,
           cudaStream_t stream);
 
-#endif  // PADDLE_ONLY_CPU
+#endif
 
 }  // namespace memory
 }  // namespace paddle

paddle/memory/memory.cc

@@ -111,7 +111,7 @@ size_t Used<platform::GPUPlace>(platform::GPUPlace place) {
   return GetGPUBuddyAllocator(place.device)->Used();
 }
 
-#endif  // PADDLE_ONLY_CPU
+#endif
 
 }  // namespace memory
 }  // namespace paddle

paddle/memory/memory_test.cc

@@ -135,4 +135,4 @@ TEST(BuddyAllocator, GPUMultAlloc) {
   }
 }
 
-#endif  // PADDLE_ONLY_CPU
+#endif

paddle/operators/CMakeLists.txt

@@ -55,12 +55,20 @@ function(op_library TARGET)
         set(pybind_flag 1)
     endif()
 
+    # pool_op contains several operators
     if ("${TARGET}" STREQUAL "pool_op")
         set(pybind_flag 1)
         # It's enough to just adding one operator to pybind
         file(APPEND ${pybind_file} "USE_OP(pool2d);\n")
     endif()
 
+    # pool_with_index_op contains several operators
+    if ("${TARGET}" STREQUAL "pool_with_index_op")
+        set(pybind_flag 1)
+        # It's enough to just adding one operator to pybind
+        file(APPEND ${pybind_file} "USE_OP(max_pool2d_with_index);\n")
+    endif()
+
     # activation_op contains several operators
     if ("${TARGET}" STREQUAL "activation_op")
         set(pybind_flag 1)

paddle/operators/activation_op.cc

@@ -201,6 +201,27 @@ class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker {
   }
 };
 
+template <typename AttrType>
+class ELUOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ELUOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(Tensor) The input of ELU operator, it shouldn't be empty. Input "
+             "is flattened and treated as a 1D array.");
+    AddOutput("Y",
+              "(Tensor) The output of ELU operator. It has the same shape as "
+              "the input.");
+    AddAttr<AttrType>(
+        "alpha", "(float, default 1.0) Alpha value in the elu formulation.")
+        .SetDefault(static_cast<AttrType>(1.));
+    AddComment(R"DOC(
+        ELU activation operator. It applies this element-wise computation on
+        the input: f(x) = max(0, x) + min(0, alpha * (exp(x) - 1)).
+        Check .. _Link: https://arxiv.org/abs/1511.07289 for more details.)DOC");
+  }
+};
+
 template <typename AttrType>
 class Relu6OpMaker : public framework::OpProtoAndCheckerMaker {
  public:
@@ -289,6 +310,9 @@ REGISTER_OP(leaky_relu, ops::ActivationOp, ops::LeakyReluOpMaker<float>,
 REGISTER_OP(soft_relu, ops::ActivationOp, ops::SoftReluOpMaker<float>,
             soft_relu_grad, ops::ActivationOpGrad);
 
+REGISTER_OP(elu, ops::ActivationOp, ops::ELUOpMaker<float>, elu_grad,
+            ops::ActivationOpGrad);
+
 REGISTER_OP(relu6, ops::ActivationOp, ops::Relu6OpMaker<float>, relu6_grad,
             ops::ActivationOpGrad);
 

paddle/operators/activation_op.h

@@ -384,6 +384,35 @@ struct LeakyReluGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+template <typename T>
+struct ELUFunctor : public BaseActivationFunctor<T> {
+  float alpha;
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"alpha", &alpha}};
+  }
+
+  template <typename Device, typename X, typename Y>
+  void operator()(Device d, X x, Y y) const {
+    y.device(d) =
+        x.cwiseMax(static_cast<T>(0)) +
+        (alpha * (x.exp() - static_cast<T>(1))).cwiseMin(static_cast<T>(0));
+  }
+};
+
+template <typename T>
+struct ELUGradFunctor : public BaseActivationFunctor<T> {
+  float alpha;
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"alpha", &alpha}};
+  }
+  template <typename Device, typename X, typename Y, typename dY, typename dX>
+  void operator()(Device d, X x, Y y, dY dy, dX dx) const {
+    dx.device(d) =
+        dy * (x > static_cast<T>(0)).template cast<T>() +
+        dy * (y + alpha) * (x < static_cast<T>(0)).template cast<T>();
+  }
+};
+
 template <typename T>
 struct PowFunctor : public BaseActivationFunctor<T> {
   float factor;
@@ -455,6 +484,7 @@ struct STanhGradFunctor : public BaseActivationFunctor<T> {
   __macro(pow, PowFunctor, PowGradFunctor);                       \
   __macro(stanh, STanhFunctor, STanhGradFunctor);                 \
   __macro(softsign, SoftsignFunctor, SoftsignGradFunctor);        \
-  __macro(relu6, Relu6Functor, Relu6GradFunctor);                \
   __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor);    \
-  __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor)
+  __macro(relu6, Relu6Functor, Relu6GradFunctor);                 \
+  __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor); \
+  __macro(elu, ELUFunctor, ELUGradFunctor)

paddle/operators/conv_shift_op.cc

+/* Copyright (c) 2017 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/conv_shift_op.h"
+#include "paddle/framework/eigen.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+class ConvShiftOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(y_dims.size(), 2, "Input(Y)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(x_dims[0], y_dims[0],
+                      "The 1st dimension of Input(X) and Input(Y) should "
+                      "be equal.");
+    PADDLE_ENFORCE_EQ(y_dims[1] % 2, 1,
+                      "The 2nd dimension of Input(Y) should be odd.");
+    PADDLE_ENFORCE_LE(y_dims[1], x_dims[1],
+                      "The 2nd dimension of Input(Y) should be less than or "
+                      "equal to the 2nd dimension of Input(X).");
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class ConvShiftGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should be not null.");
+
+    auto x_grad_name = framework::GradVarName("X");
+    if (ctx->HasOutput(x_grad_name)) {
+      auto x_dims = ctx->GetInputDim("X");
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+
+    auto y_grad_name = framework::GradVarName("Y");
+    if (ctx->HasOutput(y_grad_name)) {
+      auto y_dims = ctx->GetInputDim("Y");
+      ctx->SetOutputDim(y_grad_name, y_dims);
+    }
+  }
+};
+
+class ConvShiftOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ConvShiftOpMaker(framework::OpProto *proto,
+                   framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(Tensor, default Tensor<float>), a 2-D tensor with shape B x M, "
+             "where B is the batch size and M is the data dimension.");
+    AddInput("Y",
+             "(Tensor, default Tensor<float>), a 2-D tensor with shape B x N, "
+             "where B is the batch size and N is the data dimension. N must "
+             "be odd.");
+    AddOutput("Out",
+              "(Tensor, default Tensor<float>), a 2-D tensor with shape B x M, "
+              "i.e., the same shape as X.");
+    AddComment(R"DOC(
+ConvShift Operator.
+
+A layer for circular convolution of two vectors,
+as used in the Neural Turing Machine: https://arxiv.org/abs/1410.5401
+
+The equation is:
+
+  \f[
+      Out[i] = \sum_{j=-(N-1)/2}^{(N-1)/2} X_{i+j} * Y_{j}
+  \f]
+
+where X's index is computed modulo M, and b's index is computed modulo N.
+
+Both of the input `X` and `Y` can carry LoD (Level of Details) information.
+However, the output only shares the LoD information with input `X`.
+)DOC");
+  }
+};
+
+template <typename T>
+class ConvShiftKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *X = context.Input<Tensor>("X");
+    auto *Y = context.Input<Tensor>("Y");
+    auto *Out = context.Output<Tensor>("Out");
+    Out->mutable_data<T>(context.GetPlace());
+
+    auto x = EigenMatrix<T>::From(*X);
+    auto y = EigenMatrix<T>::From(*Y);
+    auto out = EigenMatrix<T>::From(*Out);
+    out.setZero();
+
+    size_t batch_size = X->dims()[0];
+    size_t x_width = X->dims()[1];
+    size_t y_width = Y->dims()[1];
+    size_t y_half_width = (y_width - 1) / 2;
+
+    for (size_t k = 0; k < batch_size; ++k) {
+      for (size_t i = 0; i < x_width; ++i) {
+        for (size_t j = 0; j < y_width; ++j) {
+          int index = (i + j - y_half_width + x_width) % x_width;
+          out(k, i) += x(k, index) * y(k, j);
+        }
+      }
+    }
+  }
+};
+
+template <typename T>
+class ConvShiftGradKernel<platform::CPUPlace, T>
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *X = context.Input<Tensor>("X");
+    auto *Y = context.Input<Tensor>("Y");
+    auto *dOut = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto *dX = context.Output<Tensor>(framework::GradVarName("X"));
+    auto *dY = context.Output<Tensor>(framework::GradVarName("Y"));
+
+    auto x = EigenMatrix<T>::From(*X);
+    auto y = EigenMatrix<T>::From(*Y);
+    auto dout = EigenMatrix<T>::From(*dOut);
+
+    auto x_dims = X->dims();
+    auto y_dims = Y->dims();
+    size_t batch_size = x_dims[0];
+    size_t x_width = x_dims[1];
+    size_t y_width = y_dims[1];
+    size_t y_half_width = (y_width - 1) / 2;
+
+    // The below trades code duplication for efficiency (keeping the if
+    // statement outside of the loop).
+    if (dX) {
+      dX->mutable_data<T>(context.GetPlace());
+      auto dx = EigenMatrix<T>::From(*dX);
+      dx.setZero();
+      for (size_t k = 0; k < batch_size; ++k) {
+        for (size_t i = 0; i < x_width; ++i) {
+          for (size_t j = 0; j < y_width; ++j) {
+            int index = (i + j - y_half_width + x_width) % x_width;
+            dx(k, index) += dout(k, i) * y(k, j);
+          }
+        }
+      }
+    }
+
+    if (dY) {
+      dY->mutable_data<T>(context.GetPlace());
+      auto dy = EigenMatrix<T>::From(*dY);
+      dy.setZero();
+      for (size_t k = 0; k < batch_size; ++k) {
+        for (size_t i = 0; i < x_width; ++i) {
+          for (size_t j = 0; j < y_width; ++j) {
+            int index = (i + j - y_half_width + x_width) % x_width;
+            dy(k, j) += x(k, index) * dout(k, i);
+          }
+        }
+      }
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(conv_shift, ops::ConvShiftOp, ops::ConvShiftOpMaker,
+            conv_shift_grad, ops::ConvShiftGradOp);
+REGISTER_OP_CPU_KERNEL(conv_shift,
+                       ops::ConvShiftKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    conv_shift_grad,
+    ops::ConvShiftGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/conv_shift_op.cu

+/* Copyright (c) 2017 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/conv_shift_op.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+namespace {
+
+inline int div_up(int x, int y) { return (x + y - 1) / y; }
+
+// Some notes on the design:
+//
+// Each thread is responsible for computing a single output out[k, i].
+// Thread blocks are based on tiles of x with height 1 in the batch dimension.
+//
+// This design is based on the typical use case where the filter
+// y is fairly small. For large y, it would probably be more efficient
+// to also tile across y.
+template <typename T>
+__global__ void conv_shift_forward(const T *x, const T *y, T *out, int x_width,
+                                   int y_width, int y_half_width,
+                                   int batch_size) {
+  extern __shared__ T mem[];
+
+  int tx = threadIdx.x;
+  int i = blockIdx.x * blockDim.x + tx;  // global x index
+  int k = blockIdx.y;                    // batch index
+
+  // Check if we are in a boundary block with fewer x's to process than
+  // blockDim.x.
+  int num_x =
+      (blockIdx.x == gridDim.x - 1) ? (x_width % blockDim.x) : blockDim.x;
+
+  T *sx = mem;
+  T *sx_pad = &mem[num_x];
+  T *sy = &mem[blockDim.x + y_width];
+
+  // Collaboratively load y[k, :] and length-y padding of x into shared memory.
+  int pad_start = blockIdx.x * blockDim.x + num_x + x_width - y_half_width;
+  for (int j = tx; j < y_width; j += blockDim.x) {
+    sy[j] = y[k * y_width + j];
+    sx_pad[j] = x[k * x_width + (pad_start + j) % x_width];
+  }
+
+  // Load a cyclically shifted slice of x into shared memory.
+  if (tx < num_x) {
+    int load_i = (i - y_half_width + x_width) % x_width;
+    sx[tx] = x[k * x_width + load_i];
+  } else {
+    return;
+  }
+  __syncthreads();
+
+  // Compute dot product of sx[tx:tx + y_width] and sy.
+  T sum = 0;
+  for (int j = 0; j < y_width; ++j) {
+    sum += sx[tx + j] * sy[j];
+  }
+
+  // Save to out[k, i].
+  out[k * x_width + i] = sum;
+}
+
+// Compute x gradient - initial naive implementation with atomic add.
+template <typename T>
+__global__ void conv_shift_dx(const T *dout, const T *y, T *dx, int x_width,
+                              int y_width, int y_half_width, int batch_size) {
+  int i = blockIdx.x * blockDim.x + threadIdx.x;  // x index
+  int j = blockIdx.y;                             // y index
+  int k = blockIdx.z;                             // batch index
+
+  if (i < x_width) {
+    int index = (i + j - y_half_width + x_width) % x_width;
+    atomicAdd(&dx[k * x_width + index],
+              dout[k * x_width + i] * y[k * y_width + j]);
+  }
+}
+
+// Compute y gradient - initial naive implementation with atomic add.
+template <typename T>
+__global__ void conv_shift_dy(const T *x, const T *dout, T *dy, int x_width,
+                              int y_width, int y_half_width, int batch_size) {
+  int i = blockIdx.x * blockDim.x + threadIdx.x;  // x index
+  int j = blockIdx.y;                             // y index
+  int k = blockIdx.z;                             // batch index
+
+  if (i < x_width) {
+    int index = (i + j - y_half_width + x_width) % x_width;
+    atomicAdd(&dy[k * y_width + j],
+              x[k * x_width + index] * dout[k * x_width + i]);
+  }
+}
+}  // namespace
+
+template <typename T>
+class ConvShiftKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    const Tensor *X = context.Input<Tensor>("X");
+    const Tensor *Y = context.Input<Tensor>("Y");
+    Tensor *Out = context.Output<Tensor>("Out");
+    const T *x_data = X->data<T>();
+    const T *y_data = Y->data<T>();
+    T *out_data = Out->mutable_data<T>(context.GetPlace());
+
+    int batch_size = X->dims()[0];
+    int x_width = X->dims()[1];
+    int y_width = Y->dims()[1];
+    int y_half_width = (y_width - 1) / 2;
+
+    const int x_per_block = 256;
+    int num_x_blocks = div_up(x_width, x_per_block);
+    int mem_per_block = (x_per_block + 2 * y_width) * sizeof(T);
+
+    dim3 grid_dim(num_x_blocks, batch_size);
+
+    auto stream = reinterpret_cast<const platform::CUDADeviceContext &>(
+                      context.device_context())
+                      .stream();
+
+    conv_shift_forward<T><<<grid_dim, x_per_block, mem_per_block, stream>>>(
+        x_data, y_data, out_data, x_width, y_width, y_half_width, batch_size);
+  }
+};
+
+template <typename T>
+class ConvShiftGradKernel<platform::GPUPlace, T>
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    const Tensor *X = context.Input<Tensor>("X");
+    const Tensor *Y = context.Input<Tensor>("Y");
+    const Tensor *dOut = context.Input<Tensor>(framework::GradVarName("Out"));
+    const T *x_data = X->data<T>();
+    const T *y_data = Y->data<T>();
+    const T *dout_data = dOut->data<T>();
+
+    Tensor *dX = context.Output<Tensor>(framework::GradVarName("X"));
+    Tensor *dY = context.Output<Tensor>(framework::GradVarName("Y"));
+
+    int batch_size = X->dims()[0];
+    int x_width = X->dims()[1];
+    int y_width = Y->dims()[1];
+    int y_half_width = (y_width - 1) / 2;
+
+    auto stream = reinterpret_cast<const platform::CUDADeviceContext &>(
+                      context.device_context())
+                      .stream();
+
+    const int x_per_block = 256;
+    int num_x_blocks = div_up(x_width, x_per_block);
+    dim3 grid_dim(num_x_blocks, y_width, batch_size);
+
+    if (dX) {
+      T *dx_data = dX->mutable_data<T>(context.GetPlace());
+      cudaMemsetAsync(dx_data, 0, dX->numel() * sizeof(T), stream);
+      conv_shift_dx<T><<<grid_dim, x_per_block, 0, stream>>>(
+          dout_data, y_data, dx_data, x_width, y_width, y_half_width,
+          batch_size);
+    }
+    if (dY) {
+      T *dy_data = dY->mutable_data<T>(context.GetPlace());
+      cudaMemsetAsync(dy_data, 0, dY->numel() * sizeof(T), stream);
+      conv_shift_dy<T><<<grid_dim, x_per_block, 0, stream>>>(
+          x_data, dout_data, dy_data, x_width, y_width, y_half_width,
+          batch_size);
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(conv_shift,
+                       ops::ConvShiftKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    conv_shift_grad,
+    ops::ConvShiftGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/conv_shift_op.h

+/* Copyright (c) 2017 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/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ConvShiftKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override;
+};
+
+template <typename Place, typename T>
+class ConvShiftGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override;
+};
+}  // namespace operators
+}  // namespace paddle

paddle/operators/fill_constant_op.cc

+/* 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/fill_constant_op.h"
+
+namespace paddle {
+namespace operators {
+
+class FillConstantOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of FillConstantOp should not be null.");
+    auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    std::vector<int64_t> shape_int64(shape.size(), 0);
+    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
+                   [](int a) { return static_cast<int64_t>(a); });
+    auto dims = framework::make_ddim(shape_int64);
+    ctx->SetOutputDim("Out", dims);
+  }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext &ctx) const override {
+    return static_cast<framework::DataType>(ctx.Attr<int>("dataType"));
+  }
+};
+
+class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FillConstantOpMaker(framework::OpProto *proto,
+                      framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddAttr<int>("dataType",
+                 "(int, default 5 (FP32)) "
+                 "Output data type")
+        .SetDefault(framework::DataType::FP32);
+    AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
+    AddAttr<float>("value", "(float, default 0) The value to be filled")
+        .SetDefault(0.0f);
+    AddOutput("Out",
+              "(Tensor) Tensor of specified shape will be filled "
+              "with the specified value");
+    AddComment(R"DOC(Fill up a variable with specified constant value.)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(fill_constant, ops::FillConstantOp,
+                             ops::FillConstantOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    fill_constant,
+    ops::FillConstantOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/fill_constant_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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/fill_constant_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    fill_constant,
+    ops::FillConstantOpKernel<paddle::platform::GPUPlace, float>);

paddle/operators/fill_constant_op.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/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class FillConstantOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    out->mutable_data<T>(ctx.GetPlace());
+    auto value = ctx.Attr<T>("value");
+
+    auto out_eigen = framework::EigenVector<T>::Flatten(*out);
+    auto place = ctx.GetEigenDevice<Place>();
+    out_eigen.device(place) = out_eigen.constant(static_cast<T>(value));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/interp_op.cc

+/* 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/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+class InterpOp : public NetOp {
+ public:
+  InterpOp(const std::string &type, const framework::VariableNameMap &inputs,
+           const framework::VariableNameMap &outputs,
+           const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE_NE(Input("X"), framework::kEmptyVarName,
+                      "Input(X) of InterpOp should not be null.");
+    PADDLE_ENFORCE_NE(Input("Y"), framework::kEmptyVarName,
+                      "Input(Y) of InterpOp should not be null.");
+    PADDLE_ENFORCE_NE(Input("W"), framework::kEmptyVarName,
+                      "Input(W) of InterpOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("SubOut"), framework::kEmptyVarName,
+                      "Output(SubOut) of InterpOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("MulOut"), framework::kEmptyVarName,
+                      "Output(MulOut) of InterpOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Out"), framework::kEmptyVarName,
+                      "Output(Out) of InterpOp should not be null.");
+
+    // SubOut = X - Y
+    auto x = Input("X");
+    auto y = Input("Y");
+    auto sub_out = Output("SubOut");
+    AppendOp(framework::OpRegistry::CreateOp(
+        "elementwise_sub", {{"X", {x}}, {"Y", {y}}}, {{"Out", {sub_out}}}, {}));
+
+    // MulOut = SubOut * W = (X - Y) * W
+    auto w = Input("W");
+    auto mul_out = Output("MulOut");
+    AppendOp(framework::OpRegistry::CreateOp(
+        "elementwise_mul", {{"X", {sub_out}}, {"Y", {w}}}, {{"Out", {mul_out}}},
+        {{"axis", 0}}));
+
+    // Out = MulOut + Y = (X - Y) * W + Y = X * W + Y * (1 - W)
+    AppendOp(framework::OpRegistry::CreateOp("elementwise_add",
+                                             {{"X", {mul_out}}, {"Y", {y}}},
+                                             {{"Out", {Output("Out")}}}, {}));
+
+    CompleteAddOp(false);
+  }
+};
+
+class InterpOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  InterpOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(Tensor), 2-D Matrix of shape [batch_size, data_dim]"
+             "containing data samples, the first input of interp_op");
+    AddInput("Y",
+             "(Tensor), 2-D Matrix of shape `[batch_size, data_dim]`"
+             "containing data samples, the second input of interp_op");
+    AddInput("W",
+             "(Tensor), 1-D Vector of shape [batch_size],"
+             "the interpolated values in the half-open interval [0.0, 1.0)");
+    AddOutput("SubOut",
+              "(Tensor), the intermediate subtraction outputs, saving X - Y.")
+        .AsIntermediate();
+    AddOutput("MulOut",
+              "(Tensor), the intermediate multiplication outputs,"
+              "saving the elementwise multiplication of (X - Y) and W.")
+        .AsIntermediate();
+    AddOutput("Out",
+              "(Tensor), the output of interp_op, same shape with X,"
+              "returns the first-dimensional piecewise linear interpolant "
+              "between X and Y");
+    AddComment(R"DOC(
+    Linear Interpolation with two inputs, used in NEURAL TURING MACHINE.
+
+    Equation:
+      Out.row[i] = X.row[i] * W[i] + Y.row[i] * (1 - W[i])
+                 = (X.row[i] - Y.row[i]) * W[i] + Y.row[i]
+
+    Example:
+      X = [[1,2],[3,4]],
+      Y = [[2,1],[4,3]],
+      W = [0.3, 0.4]
+
+      Then, Out = [[1.7,1.3],[3.6,3.4]]
+
+      where 1.7 = 1*0.3+2*(1-0.3),
+            1.3 = 2*0.3+1*(1-0.3),
+            3.6 = 3*0.4+4*(1-0.4),
+            3.4 = 4*0.4+3*(1-0.4)
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(interp, ops::InterpOp, ops::InterpOpMaker);

paddle/operators/math/pooling.cc

@@ -18,6 +18,11 @@ namespace paddle {
 namespace operators {
 namespace math {
 
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
 template <typename PoolProcess, typename T>
 class Pool2dFunctor<platform::CPUPlace, PoolProcess, T> {
  public:
@@ -73,6 +78,11 @@ class Pool2dFunctor<platform::CPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+* All tensors are in NCHW format.
+* Ksize, strides, paddings are two elements. These two elements represent height
+* and width, respectively.
+*/
 template <typename PoolProcess, class T>
 class Pool2dGradFunctor<platform::CPUPlace, PoolProcess, T> {
  public:
@@ -135,6 +145,11 @@ class Pool2dGradFunctor<platform::CPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
 template <class T>
 class MaxPool2dGradFunctor<platform::CPUPlace, T> {
  public:
@@ -197,7 +212,7 @@ class MaxPool2dGradFunctor<platform::CPUPlace, T> {
 };
 
 template class MaxPool2dGradFunctor<platform::CPUPlace, float>;
-// template class MaxPool2dGradFunctor<platform::CPUPlace, double>;
+template class MaxPool2dGradFunctor<platform::CPUPlace, double>;
 
 template class Pool2dFunctor<platform::CPUPlace,
                              paddle::operators::math::MaxPool<float>, float>;
@@ -216,6 +231,11 @@ template class Pool2dGradFunctor<
 template class Pool2dGradFunctor<
     platform::CPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
 
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
 template <typename PoolProcess, class T>
 class Pool3dFunctor<platform::CPUPlace, PoolProcess, T> {
  public:
@@ -286,6 +306,11 @@ class Pool3dFunctor<platform::CPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
 template <typename PoolProcess, class T>
 class Pool3dGradFunctor<platform::CPUPlace, PoolProcess, T> {
  public:
@@ -364,6 +389,11 @@ class Pool3dGradFunctor<platform::CPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
 template <class T>
 class MaxPool3dGradFunctor<platform::CPUPlace, T> {
  public:
@@ -440,7 +470,7 @@ class MaxPool3dGradFunctor<platform::CPUPlace, T> {
 };
 
 template class MaxPool3dGradFunctor<platform::CPUPlace, float>;
-// template class MaxPool3dGradFunctor<platform::CPUPlace, double>;
+template class MaxPool3dGradFunctor<platform::CPUPlace, double>;
 
 template class Pool3dFunctor<platform::CPUPlace,
                              paddle::operators::math::MaxPool<float>, float>;
@@ -458,6 +488,253 @@ template class Pool3dGradFunctor<
     platform::CPUPlace, paddle::operators::math::MaxPoolGrad<double>, double>;
 template class Pool3dGradFunctor<
     platform::CPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
+
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
+template <typename T>
+class MaxPool2dWithIndexFunctor<platform::CPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor& output,
+                  framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input.dims()[0];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output.dims()[1];
+    const int output_height = output.dims()[2];
+    const int output_width = output.dims()[3];
+    const int ksize_height = ksize[0];
+    const int ksize_width = ksize[1];
+    const int stride_height = strides[0];
+    const int stride_width = strides[1];
+    const int padding_height = paddings[0];
+    const int padding_width = paddings[1];
+    const int input_stride = input_height * input_width;
+    const int output_stride = output_height * output_width;
+
+    const T* input_data = input.data<T>();
+    T* output_data = output.mutable_data<T>(context.GetPlace());
+    T* mask_data = mask.mutable_data<T>(context.GetPlace());
+
+    for (int i = 0; i < batch_size; i++) {
+      for (int c = 0; c < output_channels; ++c) {
+        for (int ph = 0; ph < output_height; ++ph) {
+          int hstart = ph * stride_height - padding_height;
+          int hend = std::min(hstart + ksize_height, input_height);
+          hstart = std::max(hstart, 0);
+          for (int pw = 0; pw < output_width; ++pw) {
+            int wstart = pw * stride_width - padding_width;
+            int wend = std::min(wstart + ksize_width, input_width);
+            wstart = std::max(wstart, 0);
+
+            T ele = static_cast<T>(-FLT_MAX);
+            int index = -1;
+            for (int h = hstart; h < hend; ++h) {
+              for (int w = wstart; w < wend; ++w) {
+                if (ele < input_data[h * input_width + w]) {
+                  ele = input_data[h * input_width + w];
+                  index = h * input_width + w;
+                }
+              }
+            }
+            output_data[ph * output_width + pw] = ele;
+            mask_data[ph * output_width + pw] = index;
+          }
+        }
+        // offset
+        input_data += input_stride;
+        output_data += output_stride;
+        mask_data += output_stride;
+      }
+    }
+  }
+};
+
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
+template <typename T>
+class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  framework::Tensor& input_grad,
+                  const framework::Tensor& output_grad,
+                  const framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input_grad.dims()[0];
+    const int input_height = input_grad.dims()[2];
+    const int input_width = input_grad.dims()[3];
+    const int output_channels = output_grad.dims()[1];
+    const int output_height = output_grad.dims()[2];
+    const int output_width = output_grad.dims()[3];
+    const int input_stride = input_height * input_width;
+    const int output_stride = output_height * output_width;
+
+    const T* mask_data = mask.data<T>();
+    const T* output_grad_data = output_grad.data<T>();
+    T* input_grad_data = input_grad.mutable_data<T>(context.GetPlace());
+
+    for (int n = 0; n < batch_size; ++n) {
+      for (int c = 0; c < output_channels; ++c) {
+        for (int ph = 0; ph < output_height; ++ph) {
+          for (int pw = 0; pw < output_width; ++pw) {
+            const int output_idx = ph * output_width + pw;
+            const int input_idx = static_cast<int>(mask_data[output_idx]);
+            input_grad_data[input_idx] += output_grad_data[output_idx];
+          }
+        }
+        // offset
+        input_grad_data += input_stride;
+        output_grad_data += output_stride;
+        mask_data += output_stride;
+      }
+    }
+  }
+};
+
+template class MaxPool2dWithIndexFunctor<platform::CPUPlace, float>;
+template class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, float>;
+template class MaxPool2dWithIndexFunctor<platform::CPUPlace, double>;
+template class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, double>;
+
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
+template <typename T>
+class MaxPool3dWithIndexFunctor<platform::CPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor& output,
+                  framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input.dims()[0];
+    const int input_depth = input.dims()[2];
+    const int input_height = input.dims()[3];
+    const int input_width = input.dims()[4];
+    const int output_channels = output.dims()[1];
+    const int output_depth = output.dims()[2];
+    const int output_height = output.dims()[3];
+    const int output_width = output.dims()[4];
+    const int ksize_depth = ksize[0];
+    const int ksize_height = ksize[1];
+    const int ksize_width = ksize[2];
+    const int stride_depth = strides[0];
+    const int stride_height = strides[1];
+    const int stride_width = strides[2];
+    const int padding_depth = paddings[0];
+    const int padding_height = paddings[1];
+    const int padding_width = paddings[2];
+    const int input_stride = input_depth * input_height * input_width;
+    const int output_stride = output_depth * output_height * output_width;
+
+    const T* input_data = input.data<T>();
+    T* output_data = output.mutable_data<T>(context.GetPlace());
+    T* mask_data = mask.mutable_data<T>(context.GetPlace());
+
+    for (int i = 0; i < batch_size; i++) {
+      for (int c = 0; c < output_channels; ++c) {
+        for (int pd = 0; pd < output_depth; ++pd) {
+          int dstart = pd * stride_depth - padding_depth;
+          int dend = std::min(dstart + ksize_depth, input_depth);
+          dstart = std::max(dstart, 0);
+          for (int ph = 0; ph < output_height; ++ph) {
+            int hstart = ph * stride_height - padding_height;
+            int hend = std::min(hstart + ksize_height, input_height);
+            hstart = std::max(hstart, 0);
+            for (int pw = 0; pw < output_width; ++pw) {
+              int wstart = pw * stride_width - padding_width;
+              int wend = std::min(wstart + ksize_width, input_width);
+              wstart = std::max(wstart, 0);
+
+              int output_idx = (pd * output_height + ph) * output_width + pw;
+              T ele = static_cast<T>(-FLT_MAX);
+              int index = -1;
+              for (int d = dstart; d < dend; ++d) {
+                for (int h = hstart; h < hend; ++h) {
+                  for (int w = wstart; w < wend; ++w) {
+                    int input_idx = (d * input_height + h) * input_width + w;
+                    if (ele < input_data[input_idx]) {
+                      index = input_idx;
+                      ele = input_data[input_idx];
+                    }
+                  }
+                }
+              }
+              output_data[output_idx] = ele;
+              mask_data[output_idx] = index;
+            }
+          }
+        }
+        // offset
+        input_data += input_stride;
+        output_data += output_stride;
+        mask_data += output_stride;
+      }
+    }
+  }
+};
+
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
+template <typename T>
+class MaxPool3dWithIndexGradFunctor<platform::CPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  framework::Tensor& input_grad,
+                  const framework::Tensor& output_grad,
+                  const framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input_grad.dims()[0];
+    const int input_depth = input_grad.dims()[2];
+    const int input_height = input_grad.dims()[3];
+    const int input_width = input_grad.dims()[4];
+    const int output_channels = output_grad.dims()[1];
+    const int output_depth = output_grad.dims()[2];
+    const int output_height = output_grad.dims()[3];
+    const int output_width = output_grad.dims()[4];
+    const int input_stride = input_depth * input_height * input_width;
+    const int output_stride = output_depth * output_height * output_width;
+
+    const T* mask_data = mask.data<T>();
+    const T* output_grad_data = output_grad.data<T>();
+    T* input_grad_data = input_grad.mutable_data<T>(context.GetPlace());
+
+    for (int n = 0; n < batch_size; ++n) {
+      for (int c = 0; c < output_channels; ++c) {
+        for (int pd = 0; pd < output_depth; ++pd) {
+          for (int ph = 0; ph < output_height; ++ph) {
+            for (int pw = 0; pw < output_width; ++pw) {
+              const int output_idx =
+                  (pd * output_height + ph) * output_width + pw;
+              const int input_idx = static_cast<int>(mask_data[output_idx]);
+              input_grad_data[input_idx] += output_grad_data[output_idx];
+            }
+          }
+        }
+        // offset
+        input_grad_data += input_stride;
+        output_grad_data += output_stride;
+        mask_data += output_stride;
+      }
+    }
+  }
+};
+
+template class MaxPool3dWithIndexFunctor<platform::CPUPlace, float>;
+template class MaxPool3dWithIndexGradFunctor<platform::CPUPlace, float>;
+template class MaxPool3dWithIndexFunctor<platform::CPUPlace, double>;
+template class MaxPool3dWithIndexGradFunctor<platform::CPUPlace, double>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/pooling.cu

@@ -144,11 +144,16 @@ __global__ void KernelMaxPool2DGrad(
 
     if (maxIndex != -1) {
       // atomic add
-      atomicAdd(input_grad + maxIndex, output_grad[index]);
+      platform::CudaAtomicAdd(input_grad + maxIndex, output_grad[index]);
     }
   }
 }
 
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
 template <typename PoolProcess, typename T>
 class Pool2dFunctor<platform::GPUPlace, PoolProcess, T> {
  public:
@@ -190,6 +195,11 @@ class Pool2dFunctor<platform::GPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
 template <typename PoolProcess, typename T>
 class Pool2dGradFunctor<platform::GPUPlace, PoolProcess, T> {
  public:
@@ -234,6 +244,11 @@ class Pool2dGradFunctor<platform::GPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
 template <typename T>
 class MaxPool2dGradFunctor<platform::GPUPlace, T> {
  public:
@@ -278,9 +293,7 @@ class MaxPool2dGradFunctor<platform::GPUPlace, T> {
 };
 
 template class MaxPool2dGradFunctor<platform::GPUPlace, float>;
-// template class MaxPool2dGradFunctor<platform::GPUPlace, double>; // The
-// 64-bit floating-point version of atomicAdd() is only supported by devices of
-// compute capability 6.x and higher.
+template class MaxPool2dGradFunctor<platform::GPUPlace, double>;
 
 template class Pool2dFunctor<platform::GPUPlace,
                              paddle::operators::math::MaxPool<float>, float>;
@@ -453,11 +466,16 @@ __global__ void KernelMaxPool3DGrad(
     }
     if (maxIdx != -1) {
       // atomic add
-      atomicAdd(input_grad + maxIdx, output_grad[index]);
+      platform::CudaAtomicAdd(input_grad + maxIdx, output_grad[index]);
     }
   }
 }
 
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
 template <typename PoolProcess, class T>
 class Pool3dFunctor<platform::GPUPlace, PoolProcess, T> {
  public:
@@ -506,6 +524,11 @@ class Pool3dFunctor<platform::GPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
 template <typename PoolProcess, class T>
 class Pool3dGradFunctor<platform::GPUPlace, PoolProcess, T> {
  public:
@@ -558,6 +581,11 @@ class Pool3dGradFunctor<platform::GPUPlace, PoolProcess, T> {
   }
 };
 
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
 template <class T>
 class MaxPool3dGradFunctor<platform::GPUPlace, T> {
  public:
@@ -609,9 +637,7 @@ class MaxPool3dGradFunctor<platform::GPUPlace, T> {
 };
 
 template class MaxPool3dGradFunctor<platform::GPUPlace, float>;
-// template class MaxPool3dGradFunctor<platform::GPUPlace, double>;  // The
-// 64-bit floating-point version of atomicAdd() is only supported by devices of
-// compute capability 6.x and higher.
+template class MaxPool3dGradFunctor<platform::GPUPlace, double>;
 
 template class Pool3dFunctor<platform::GPUPlace,
                              paddle::operators::math::MaxPool<float>, float>;
@@ -630,6 +656,404 @@ template class Pool3dGradFunctor<
 template class Pool3dGradFunctor<
     platform::GPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
 
+template <typename T>
+__global__ void KernelMaxPool2dWithIdx(
+    const int nthreads, const T* input_data, T* output_data, T* mask_data,
+    const int channels, const int input_height, const int input_width,
+    const int output_height, const int output_width, const int ksize_height,
+    const int ksize_width, const int stride_height, const int stride_width,
+    const int padding_height, const int padding_width) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads;
+       index += blockDim.x * gridDim.x) {
+    int pw = index % output_width;
+    int ph = (index / output_width) % output_height;
+    int c = (index / output_width / output_height) % channels;
+    int batch_idx = index / output_width / output_height / channels;
+
+    int hstart = ph * stride_height - padding_height;
+    int hend = min(hstart + ksize_height, input_height);
+    hstart = max(hstart, 0);
+
+    int wstart = pw * stride_width - padding_width;
+    int wend = min(wstart + ksize_width, input_width);
+    wstart = max(wstart, 0);
+
+    input_data += (batch_idx * channels + c) * input_height * input_width;
+    T ele = -FLT_MAX;
+    int max_index = -1;
+    for (int h = hstart; h < hend; ++h) {
+      for (int w = wstart; w < wend; ++w) {
+        int input_index = h * input_width + w;
+        if (ele < input_data[input_index]) {
+          max_index = input_index;
+          ele = input_data[input_index];
+        }
+      }
+    }
+    output_data[index] = ele;
+    mask_data[index] = max_index;
+  }
+}
+
+template <typename T>
+__global__ void KernelMaxPool2DWithIdxGrad(
+    const int nthreads, T* input_grad, const T* output_grad, const T* mask_data,
+    const int channels, const int input_height, const int input_width,
+    const int output_height, const int output_width, const int ksize_height,
+    const int ksize_width, const int stride_height, const int stride_width,
+    const int padding_height, const int padding_width) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads;
+       index += blockDim.x * gridDim.x) {
+    int w_offset = index % input_width;
+    int h_offset = (index / input_width) % input_height;
+    int c_offset = (index / input_width / input_height) % channels;
+    int batch_idx = index / input_width / input_height / channels;
+
+    int ph_start =
+        (h_offset + padding_height < ksize_height)
+            ? 0
+            : (h_offset + padding_height - ksize_height) / stride_height + 1;
+    int pw_start =
+        (w_offset + padding_width < ksize_width)
+            ? 0
+            : (w_offset + padding_width - ksize_width) / stride_width + 1;
+    int ph_end =
+        min((h_offset + padding_height) / stride_height + 1, output_height);
+    int pw_end =
+        min((w_offset + padding_width) / stride_width + 1, output_width);
+
+    T gradient = 0;
+    int input_current_featuremap_idx = h_offset * input_width + w_offset;
+    int output_idx =
+        (batch_idx * channels + c_offset) * output_height * output_width;
+
+    mask_data += output_idx;
+    output_grad += output_idx;
+    for (int ph = ph_start; ph < ph_end; ++ph) {
+      for (int pw = pw_start; pw < pw_end; ++pw) {
+        if (mask_data[ph * output_width + pw] == input_current_featuremap_idx)
+          gradient += output_grad[ph * output_width + pw];
+      }
+    }
+    input_grad[index] = gradient;
+  }
+}
+
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
+template <typename T>
+class MaxPool2dWithIndexFunctor<platform::GPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor& output,
+                  framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input.dims()[0];
+    const int input_channels = input.dims()[1];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output.dims()[1];
+    const int output_height = output.dims()[2];
+    const int output_width = output.dims()[3];
+    const int ksize_height = ksize[0];
+    const int ksize_width = ksize[1];
+    const int stride_height = strides[0];
+    const int stride_width = strides[1];
+    const int padding_height = paddings[0];
+    const int padding_width = paddings[1];
+
+    const T* input_data = input.data<T>();
+    T* output_data = output.mutable_data<T>(context.GetPlace());
+    T* mask_data = mask.mutable_data<T>(context.GetPlace());
+
+    int nthreads = batch_size * output_channels * output_height * output_width;
+    int blocks = (nthreads + 1024 - 1) / 1024;
+    dim3 threads(1024, 1);
+    dim3 grid(blocks, 1);
+
+    KernelMaxPool2dWithIdx<
+        T><<<grid, threads, 0,
+             reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                 .stream()>>>(nthreads, input_data, output_data, mask_data,
+                              input_channels, input_height, input_width,
+                              output_height, output_width, ksize_height,
+                              ksize_width, stride_height, stride_width,
+                              padding_height, padding_width);
+  }
+};
+
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
+template <typename T>
+class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  framework::Tensor& input_grad,
+                  const framework::Tensor& output_grad,
+                  const framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input_grad.dims()[0];
+    const int input_channels = input_grad.dims()[1];
+    const int input_height = input_grad.dims()[2];
+    const int input_width = input_grad.dims()[3];
+    const int output_height = output_grad.dims()[2];
+    const int output_width = output_grad.dims()[3];
+    const int ksize_height = ksize[0];
+    const int ksize_width = ksize[1];
+    const int stride_height = strides[0];
+    const int stride_width = strides[1];
+    const int padding_height = paddings[0];
+    const int padding_width = paddings[1];
+
+    const T* mask_data = mask.data<T>();
+    const T* output_grad_data = output_grad.data<T>();
+    T* input_grad_data = input_grad.mutable_data<T>(context.GetPlace());
+
+    int nthreads = batch_size * input_channels * input_height * input_width;
+    int blocks = (nthreads + 1024 - 1) / 1024;
+    dim3 threads(1024, 1);
+    dim3 grid(blocks, 1);
+
+    KernelMaxPool2DWithIdxGrad<
+        T><<<grid, threads, 0,
+             reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                 .stream()>>>(nthreads, input_grad_data, output_grad_data,
+                              mask_data, input_channels, input_height,
+                              input_width, output_height, output_width,
+                              ksize_height, ksize_width, stride_height,
+                              stride_width, padding_height, padding_width);
+  }
+};
+
+template class MaxPool2dWithIndexFunctor<platform::GPUPlace, float>;
+template class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, float>;
+template class MaxPool2dWithIndexFunctor<platform::GPUPlace, double>;
+template class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, double>;
+
+template <typename T>
+__global__ void KernelMaxPool3DWithIdx(
+    const int nthreads, const T* input_data, T* output_data, T* mask_data,
+    const int channels, const int input_depth, const int input_height,
+    const int input_width, const int output_depth, const int output_height,
+    const int output_width, const int ksize_depth, const int ksize_height,
+    const int ksize_width, const int stride_depth, const int stride_height,
+    const int stride_width, const int padding_depth, const int padding_height,
+    const int padding_width) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads;
+       index += blockDim.x * gridDim.x) {
+    int pw = index % output_width;
+    int ph = (index / output_width) % output_height;
+    int pd = (index / output_width / output_height) % output_depth;
+    int c = (index / output_width / output_height / output_depth) % channels;
+    int batch_idx =
+        index / output_width / output_height / output_depth / channels;
+
+    int dstart = pd * stride_depth - padding_depth;
+    int hstart = ph * stride_height - padding_height;
+    int wstart = pw * stride_width - padding_width;
+    int dend = min(dstart + ksize_depth, input_depth);
+    int hend = min(hstart + ksize_height, input_height);
+    int wend = min(wstart + ksize_width, input_width);
+    dstart = max(dstart, 0);
+    hstart = max(hstart, 0);
+    wstart = max(wstart, 0);
+
+    T ele = -FLT_MAX;
+    int max_index = -1;
+    input_data +=
+        (batch_idx * channels + c) * input_depth * input_height * input_width;
+
+    for (int d = dstart; d < dend; ++d) {
+      for (int h = hstart; h < hend; ++h) {
+        for (int w = wstart; w < wend; ++w) {
+          if (ele < input_data[(d * input_height + h) * input_width + w]) {
+            max_index = (d * input_height + h) * input_width + w;
+            ele = input_data[max_index];
+          }
+        }
+      }
+    }
+    output_data[index] = ele;
+    mask_data[index] = max_index;
+  }
+}
+
+template <typename T>
+__global__ void KernelMaxPool3DWithIdxGrad(
+    const int nthreads, T* input_grad, const T* output_grad, const T* mask,
+    const int channels, const int input_depth, const int input_height,
+    const int input_width, const int output_depth, const int output_height,
+    const int output_width, const int ksize_depth, const int ksize_height,
+    const int ksize_width, const int stride_depth, const int stride_height,
+    const int stride_width, const int padding_depth, const int padding_height,
+    const int padding_width) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads;
+       index += blockDim.x * gridDim.x) {
+    int w_offset = index % input_width;
+    int h_offset = (index / input_width) % input_height;
+    int d_offset = (index / input_width / input_height) % input_depth;
+    int c_offset =
+        (index / input_width / input_height / input_depth) % channels;
+    int batch_idx = index / input_width / input_height / input_depth / channels;
+
+    int pd_start =
+        (d_offset + padding_depth < ksize_depth)
+            ? 0
+            : (d_offset + padding_depth - ksize_depth) / stride_depth + 1;
+    int ph_start =
+        (h_offset + padding_height < ksize_height)
+            ? 0
+            : (h_offset + padding_height - ksize_height) / stride_height + 1;
+    int pw_start =
+        (w_offset + padding_width < ksize_width)
+            ? 0
+            : (w_offset + padding_width - ksize_width) / stride_width + 1;
+    int pd_end =
+        min((d_offset + padding_depth) / stride_depth + 1, output_depth);
+    int ph_end =
+        min((h_offset + padding_height) / stride_height + 1, output_height);
+    int pw_end =
+        min((w_offset + padding_width) / stride_width + 1, output_width);
+
+    T gradient = 0;
+    int input_current_feature_map_idx =
+        (d_offset * input_height + h_offset) * input_width + w_offset;
+    int output_idx = (batch_idx * channels + c_offset) * output_depth *
+                     output_height * output_width;
+    mask += output_idx;
+    output_grad += output_idx;
+
+    for (int pd = pd_start; pd < pd_end; ++pd) {
+      for (int ph = ph_start; ph < ph_end; ++ph) {
+        for (int pw = pw_start; pw < pw_end; ++pw) {
+          if (mask[(pd * output_height + ph) * output_width + pw] ==
+              input_current_feature_map_idx)
+            gradient +=
+                output_grad[(pd * output_height + ph) * output_width + pw];
+        }
+      }
+    }
+    input_grad[index] = gradient;
+  }
+}
+
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
+template <typename T>
+class MaxPool3dWithIndexFunctor<platform::GPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor& output,
+                  framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input.dims()[0];
+    const int input_channels = input.dims()[1];
+    const int input_depth = input.dims()[2];
+    const int input_height = input.dims()[3];
+    const int input_width = input.dims()[4];
+    const int output_channels = output.dims()[1];
+    const int output_depth = output.dims()[2];
+    const int output_height = output.dims()[3];
+    const int output_width = output.dims()[4];
+    const int ksize_depth = ksize[0];
+    const int ksize_height = ksize[1];
+    const int ksize_width = ksize[2];
+    const int stride_depth = strides[0];
+    const int stride_height = strides[1];
+    const int stride_width = strides[2];
+    const int padding_depth = paddings[0];
+    const int padding_height = paddings[1];
+    const int padding_width = paddings[2];
+
+    const T* input_data = input.data<T>();
+    T* output_data = output.mutable_data<T>(context.GetPlace());
+    T* mask_data = mask.mutable_data<T>(context.GetPlace());
+
+    int nthreads = batch_size * output_channels * output_depth * output_height *
+                   output_width;
+    int blocks = (nthreads + 1024 - 1) / 1024;
+    dim3 threads(1024, 1);
+    dim3 grid(blocks, 1);
+
+    KernelMaxPool3DWithIdx<
+        T><<<grid, threads, 0,
+             reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                 .stream()>>>(
+        nthreads, input_data, output_data, mask_data, input_channels,
+        input_depth, input_height, input_width, output_depth, output_height,
+        output_width, ksize_depth, ksize_height, ksize_width, stride_depth,
+        stride_height, stride_width, padding_depth, padding_height,
+        padding_width);
+  }
+};
+
+/*
+ * All tensors are in NCDHW format.
+ * Ksize, strides, paddings are three elements. These three elements represent
+ * depth, height and width, respectively.
+ */
+template <typename T>
+class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  framework::Tensor& input_grad,
+                  const framework::Tensor& output_grad,
+                  const framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings) {
+    const int batch_size = input_grad.dims()[0];
+    const int input_channels = input_grad.dims()[1];
+    const int input_depth = input_grad.dims()[2];
+    const int input_height = input_grad.dims()[3];
+    const int input_width = input_grad.dims()[4];
+    const int output_depth = output_grad.dims()[2];
+    const int output_height = output_grad.dims()[3];
+    const int output_width = output_grad.dims()[4];
+    const int ksize_depth = ksize[0];
+    const int ksize_height = ksize[1];
+    const int ksize_width = ksize[2];
+    const int stride_depth = strides[0];
+    const int stride_height = strides[1];
+    const int stride_width = strides[2];
+    const int padding_depth = paddings[0];
+    const int padding_height = paddings[1];
+    const int padding_width = paddings[2];
+
+    const T* output_grad_data = output_grad.data<T>();
+    const T* mask_data = mask.data<T>();
+    T* input_grad_data = input_grad.mutable_data<T>(context.GetPlace());
+
+    int nthreads =
+        batch_size * input_channels * input_depth * input_height * input_width;
+    int blocks = (nthreads + 1024 - 1) / 1024;
+    dim3 threads(1024, 1);
+    dim3 grid(blocks, 1);
+
+    KernelMaxPool3DWithIdxGrad<
+        T><<<grid, threads, 0,
+             reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                 .stream()>>>(
+        nthreads, input_grad_data, output_grad_data, mask_data, input_channels,
+        input_depth, input_height, input_width, output_depth, output_height,
+        output_width, ksize_depth, ksize_height, ksize_width, stride_depth,
+        stride_height, stride_width, padding_depth, padding_height,
+        padding_width);
+  }
+};
+
+template class MaxPool3dWithIndexFunctor<platform::GPUPlace, float>;
+template class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, float>;
+template class MaxPool3dWithIndexFunctor<platform::GPUPlace, double>;
+template class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, double>;
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/pooling.h

@@ -21,15 +21,27 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 namespace math {
-//////////////////////
-#define FLT_MAX __FLT_MAX__  //
 
+#define FLT_MAX \
+  __FLT_MAX__  // It might need to be placed in another file, but I'm still
+               // wondering where to put it.
+
+/*
+ * \brief Extracting simple operations from pooling.
+ *        Both MaxPool and AvgPool need "initial", "compute" and "finalize"
+ * operation.
+ *        MaxPool initializes temp variable to the negative maximum to find the
+ * maximum value in the pooling field.
+ *        AvgPool initializes temp variable to the zero to accumulate all values
+ * in pool pooling, and finally takes the average.
+ *        MaxPoolGrad and AvgPoolGrad are gradient operations respectively.
+ */
 template <class T>
 class MaxPool {
  public:
   DEVICE inline T initial() { return static_cast<T>(-FLT_MAX); }
   DEVICE inline void compute(T& y, const T& x) { y = y > x ? y : x; }
-  DEVICE inline void finalize(T& y, const T& poo_size) {}
+  DEVICE inline void finalize(T& y, const T& pool_field) {}
 };
 
 template <class T>
@@ -37,8 +49,9 @@ class AvgPool {
  public:
   DEVICE inline T initial() { return static_cast<T>(0); }
   DEVICE inline void compute(T& y, const T& x) { y += x; }
-  DEVICE inline void finalize(T& y, const T& poo_size) { y /= poo_size; }
+  DEVICE inline void finalize(T& y, const T& pool_field) { y /= pool_field; }
 };
+
 template <class T>
 class MaxPoolGrad {
  public:
@@ -57,6 +70,20 @@ class AvgPoolGrad {
   }
 };
 
+/*
+ * \brief Getting pooling results, and calculating gradient.
+ *
+ * In pool2d, all tensors are in NCHW format. Where N is batch size, C is the
+ * number of channels, H and W is the height and width of feature.
+ * In pool3d, all tensors are in NCDHW format. Where N is batch size, C is the
+ * number of channels, D, H and W is the depth, height and width of feature.
+ *
+ * In max pooling, it is possible that the pooling region has multiple maximum
+ * elements. In this case, we should compute the gradient of the first maximum
+ * element.
+ * This is different from average pooling. So we rewrite the max_pool_grad:
+ * MaxPool2dGradFunctor, MaxPool3dGradFunctor.
+ */
 template <typename Place, typename PoolProcess, typename T>
 class Pool2dFunctor {
  public:
@@ -117,6 +144,51 @@ class MaxPool3dGradFunctor {
                   std::vector<int>& strides, std::vector<int>& paddings);
 };
 
+/*
+ * \brief Getting max pooling results and corresponding max index, and
+ * calculating gradient.
+ * In up-sampling-pooling, it is necessary to know max element index.
+ * In pool2d, all tensors are in NCHW format. In pool3d, all tensors are in
+ * NCDHW format.
+ */
+template <typename Place, typename T>
+class MaxPool2dWithIndexFunctor {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor& output,
+                  framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings);
+};
+
+template <typename Place, typename T>
+class MaxPool2dWithIndexGradFunctor {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  framework::Tensor& input_grad,
+                  const framework::Tensor& output_grad,
+                  const framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings);
+};
+
+template <typename Place, typename T>
+class MaxPool3dWithIndexFunctor {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor& output,
+                  framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings);
+};
+
+template <typename Place, typename T>
+class MaxPool3dWithIndexGradFunctor {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  framework::Tensor& input_grad,
+                  const framework::Tensor& output_grad,
+                  const framework::Tensor& mask, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings);
+};
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/net_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <set>
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/op_registry.h"
 

paddle/operators/pool_with_index_op.cc

+/* 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/pool_with_index_op.h"
+
+namespace paddle {
+namespace operators {
+
+inline int OutputSizeMaxPool(int input_size, int filter_size, int padding,
+                             int stride) {
+  int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
+  return output_size;
+}
+
+class MaxPoolWithIndexOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "X(Input) of Pooling should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Out(Output) of Pooling should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Mask"),
+                   "Mask(Output) of Pooling should not be null.");
+
+    auto in_x_dims = ctx->GetInputDim("X");
+
+    std::vector<int> ksize = ctx->Attrs().Get<std::vector<int>>("ksize");
+    std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
+    std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+
+    PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
+                   "Pooling intput should be 4-D or 5-D");
+
+    if (ctx->Attrs().Get<bool>("globalPooling")) {
+      ksize.resize(static_cast<size_t>(in_x_dims.size()) - 2);
+      for (size_t i = 0; i < ksize.size(); ++i)
+        ksize[i] = static_cast<int>(in_x_dims[i + 2]);
+    }
+
+    PADDLE_ENFORCE(in_x_dims.size() - ksize.size() == 2U,
+                   "Intput size and pooling size should be consistent.");
+    PADDLE_ENFORCE_EQ(ksize.size(), strides.size(),
+                      "Strides size and pooling size should be the same.");
+    PADDLE_ENFORCE_EQ(ksize.size(), paddings.size(),
+                      "Paddings size and pooling size should be the same.");
+
+    std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
+    for (size_t i = 0; i < ksize.size(); ++i) {
+      output_shape.push_back(OutputSizeMaxPool(in_x_dims[i + 2], ksize[i],
+                                               paddings[i], strides[i]));
+    }
+    ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
+    ctx->SetOutputDim("Mask", framework::make_ddim(output_shape));
+  }
+};
+
+class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
+                   "Input(X@GRAD) should not be null.");
+    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+  }
+};
+
+class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MaxPool2dWithIndexOpMaker(framework::OpProto *proto,
+                            framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(
+        "X",
+        "The input tensor of pooling operator. "
+        "The format of input tensor is NCHW. Where N is batch size, C is the "
+        "number of channels, H and W is the height and width of image.");
+    AddOutput("Out",
+              "The output tensor of pooling operator."
+              "The format of output tensor is also NCHW."
+              "Where N is batch size, C is "
+              "the number of channels, H and W is the height and "
+              "width of image.");
+    AddOutput("Mask",
+              "The Mask tensor of pooling operator."
+              "The format of output tensor is also NCHW."
+              "Where N is batch size, C is the number of channels, H and W "
+              "is the height and width of image."
+              "The value in it is the index in current feature map");
+
+    AddAttr<std::vector<int>>(
+        "ksize",
+        "The pooling size(height, width) of pooling operator."
+        "If globalPooling = true, ksize is ignored and need not be "
+        "specified.");  // TODO(Chengduo): Add checker. (Currently,
+                        // TypedAttrChecker don't support vector type.)
+    AddAttr<bool>(
+        "globalPooling",
+        "Whether to use the globalPooling."
+        "Bool constant equal to false or true."
+        "Default false."
+        "If globalPooling = true, ksize is ignored and need not be specified.")
+        .SetDefault(false);
+    AddAttr<std::vector<int>>("strides",
+                              "Strides(height, width) of pooling operator."
+                              "Default {1,1}.")
+        .SetDefault({1, 1});  // TODO(Chengduo): Add checker. (Currently,
+                              // TypedAttrChecker don't support vector type.)
+    AddAttr<std::vector<int>>("paddings",
+                              "Paddings(height, width) of pooling operator."
+                              "Default {0,0}.")
+        .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
+                              // TypedAttrChecker don't support vector type.)
+
+    AddComment(R"DOC(
+The maxPooling2d with index operation calculates the output and the mask
+based on the input and ksize, strides, paddings parameters. Input(X) and
+output(Out, Mask) are in NCHW format. Where N is batch size, C is the
+number of channels, H and W is the height and width of feature.
+Parameters(ksize, strides, paddings) are two elements.
+These two elements represent height and width, respectively.
+)DOC");
+  }
+};
+
+class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MaxPool3dWithIndexOpMaker(framework::OpProto *proto,
+                            framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(
+        "X",
+        "The input tensor of pooling operator. "
+        "The format of input tensor is NCDHW. Where N is batch size, C is "
+        "the number of channels, D, H and W is the depth, height and width of "
+        "image.");
+    AddOutput("Out",
+              "The output tensor of pooling operator."
+              "The format of output tensor is also NCDHW."
+              "Where N is batch size, C is "
+              "the number of channels, D, H and W is the depth, height and "
+              "width of image.");
+    AddOutput("Mask",
+              "The Mask tensor of pooling operator."
+              "The format of output tensor is also NCDHW."
+              "Where N is batch size, C is the number of channels, D, H and W "
+              "is the depth, height and width of image."
+              "The value in it is the index in current feature map");
+
+    AddAttr<std::vector<int>>(
+        "ksize",
+        "The pooling size(depth, height, width) of pooling operator."
+        "If globalPooling = true, ksize is ignored and need not be "
+        "specified.");  // TODO(Chengduo): Add checker. (Currently,
+                        // TypedAttrChecker don't support vector type.)
+    AddAttr<bool>(
+        "globalPooling",
+        "Whether to use the globalPooling."
+        "Bool constant equal to false or true."
+        "Default false."
+        "If globalPooling = true, ksize is ignored and need not be specified.")
+        .SetDefault(false);
+    AddAttr<std::vector<int>>(
+        "strides",
+        "Strides(depth, height, width) of pooling operator."
+        "Default {1,1,1}.")
+        .SetDefault({1, 1, 1});  // TODO(Chengduo): Add checker. (Currently,
+                                 // TypedAttrChecker don't support vector type.)
+    AddAttr<std::vector<int>>(
+        "paddings",
+        "Paddings(depth, height, width) of pooling operator."
+        "Default {0,0,0}.")
+        .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
+                                 // TypedAttrChecker don't support vector type.)
+
+    AddComment(R"DOC(
+The maxpooling3d with index operation calculates the output and the mask
+based on the input and ksize, strides, paddings parameters.
+Input(X) and output(Out, Mask) are in NCDHW format. Where N is batch
+size, C is the number of channels, D, H and W is the depth, height and
+width of feature. Parameters(ksize, strides, paddings) are three elements.
+These three elements represent depth, height and width, respectively.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP(max_pool2d_with_index, ops::MaxPoolWithIndexOp,
+            ops::MaxPool2dWithIndexOpMaker, max_pool2d_with_index_grad,
+            ops::MaxPoolWithIndexOpGrad);
+
+REGISTER_OP_CPU_KERNEL(
+    max_pool2d_with_index,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    max_pool2d_with_index_grad,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float>)
+
+REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp,
+            ops::MaxPool3dWithIndexOpMaker, max_pool3d_with_index_grad,
+            ops::MaxPoolWithIndexOpGrad);
+
+REGISTER_OP_CPU_KERNEL(
+    max_pool3d_with_index,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    max_pool3d_with_index_grad,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float>)

paddle/operators/pool_with_index_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 "paddle/operators/pool_with_index_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    max_pool2d_with_index,
+    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    max_pool2d_with_index_grad,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float>)
+
+REGISTER_OP_GPU_KERNEL(
+    max_pool3d_with_index,
+    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    max_pool3d_with_index_grad,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float>)

paddle/operators/pool_with_index_op.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/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+#include "paddle/operators/math/pooling.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename Place, typename T>
+class MaxPoolWithIndexKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* in_x = context.Input<Tensor>("X");
+    Tensor* out = context.Output<Tensor>("Out");
+    Tensor* mask = context.Output<Tensor>("Mask");
+
+    std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
+    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+    if (context.Attr<bool>("globalPooling")) {
+      for (size_t i = 0; i < ksize.size(); ++i) {
+        ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
+      }
+    }
+
+    switch (ksize.size()) {
+      case 2: {
+        paddle::operators::math::MaxPool2dWithIndexFunctor<Place, T>
+            pool2d_forward;
+        pool2d_forward(context.device_context(), *in_x, *out, *mask, ksize,
+                       strides, paddings);
+      } break;
+      case 3: {
+        paddle::operators::math::MaxPool3dWithIndexFunctor<Place, T>
+            pool3d_forward;
+        pool3d_forward(context.device_context(), *in_x, *out, *mask, ksize,
+                       strides, paddings);
+      } break;
+    }
+  }
+};
+
+template <typename Place, typename T>
+class MaxPoolWithIndexGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* mask = context.Input<Tensor>("Mask");
+    const Tensor* out_grad =
+        context.Input<Tensor>(framework::GradVarName("Out"));
+    Tensor* in_x_grad = context.Output<Tensor>(framework::GradVarName("X"));
+
+    std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
+    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+    if (context.Attr<bool>("globalPooling")) {
+      for (size_t i = 0; i < ksize.size(); ++i) {
+        ksize[i] = static_cast<int>(in_x_grad->dims()[i + 2]);
+      }
+    }
+
+    if (in_x_grad) {
+      in_x_grad->mutable_data<T>(context.GetPlace());
+      auto temp = framework::EigenVector<T>::Flatten(*in_x_grad);
+      temp.device(context.GetEigenDevice<Place>()) =
+          temp.constant(static_cast<T>(0));
+
+      switch (ksize.size()) {
+        case 2: {
+          paddle::operators::math::MaxPool2dWithIndexGradFunctor<Place, T>
+              pool2d_backward;
+          pool2d_backward(context.device_context(), *in_x_grad, *out_grad,
+                          *mask, ksize, strides, paddings);
+        } break;
+        case 3: {
+          paddle::operators::math::MaxPool3dWithIndexGradFunctor<Place, T>
+              pool3d_backward;
+          pool3d_backward(context.device_context(), *in_x_grad, *out_grad,
+                          *mask, ksize, strides, paddings);
+        } break;
+      }
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/platform/device_context.cc

@@ -136,7 +136,7 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
 cudaStream_t CUDADeviceContext::stream() const { return stream_; }
 
-#endif  // PADDLE_ONLY_CPU
+#endif
 
 }  // namespace platform
 }  // namespace paddle

paddle/platform/enforce.h

@@ -41,7 +41,7 @@ limitations under the License. */
 #include <thrust/system/cuda/error.h>
 #include <thrust/system_error.h>
 
-#endif  // PADDLE_ONLY_CPU
+#endif
 
 namespace paddle {
 namespace platform {

paddle/platform/gpu_info.h

@@ -63,4 +63,4 @@ void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device,
 }  // namespace platform
 }  // namespace paddle
 
-#endif  // PADDLE_ONLY_CPU
+#endif

paddle/pybind/protobuf.cc

@@ -117,7 +117,6 @@ void BindProgramDesc(py::module &m) {
       .def("append_block", &ProgramDescBind::AppendBlock,
            py::return_value_policy::reference)
       .def("block", &ProgramDescBind::Block, py::return_value_policy::reference)
-      .def("__str__", &ProgramDescBind::DebugString)
       .def("num_blocks", &ProgramDescBind::Size);
 }
 
@@ -193,8 +192,6 @@ void BindOpDesc(py::module &m) {
       .def("output", &OpDescBind::Output)
       .def("output_names", &OpDescBind::OutputNames)
       .def("set_output", &OpDescBind::SetOutput)
-      .def("__str__", &OpDescBind::DebugString)
-      .def("__repr__", &OpDescBind::DebugString)
       .def("has_attr", &OpDescBind::HasAttr)
       .def("attr_type", &OpDescBind::GetAttrType)
       .def("attr_names", &OpDescBind::AttrNames)

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

@@ -181,6 +181,26 @@ class TestSoftRelu(OpTest):
         self.check_grad(['X'], 'Y', max_relative_error=0.02)
 
 
+class TestELU(OpTest):
+    def setUp(self):
+        self.op_type = "elu"
+        x = np.random.uniform(-3, 3, [4, 4]).astype("float32")
+        alpha = 1.
+        # Note: unlike other Relu extensions, point 0 on standard ELU function (i.e. alpha = 1)
+        # is differentiable, so we can skip modifications like x[np.abs(x) < 0.005] = 0.02 here
+        self.inputs = {'X': x}
+        self.attrs = {'alpha': alpha}
+        self.outputs = {
+            'Y': np.maximum(0, x) + np.minimum(0, alpha * (np.exp(x) - 1))
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y', max_relative_error=0.02)
+
+
 class TestReciprocal(OpTest):
     def setUp(self):
         self.op_type = "reciprocal"

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def conv_shift_forward(x, y):
+    out = np.zeros_like(x)
+    M = x.shape[1]
+    N = y.shape[1]
+    y_half_width = (N - 1) / 2
+    for i in xrange(M):
+        for j in xrange(N):
+            out[:, i] += x[:, (i + j + M - y_half_width) % M] * y[:, j]
+    return out
+
+
+class TestConvShiftOp(OpTest):
+    def setUp(self):
+        self.op_type = "conv_shift"
+
+        batch_size = 4
+        x_dim = 17
+        y_dim = 3  # must be odd and <= x_dim
+        x = np.random.random((batch_size, x_dim)).astype("float32")
+        y = np.random.random((batch_size, y_dim)).astype("float32")
+        self.inputs = {'X': x, 'Y': y}
+
+        out = conv_shift_forward(x, y)
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05)
+
+    def test_check_grad_ignore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set("X"))
+
+    def test_check_grad_ignore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
+
+
+if __name__ == '__main__':
+    unittest.main()

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestFillConstantOp1(OpTest):
+    def setUp(self):
+        '''Test fill_constant op with specified value
+        '''
+        self.op_type = "fill_constant"
+
+        self.inputs = {}
+        self.attrs = {'shape': [123, 92], 'value': 3.8}
+        self.outputs = {'Out': np.full((123, 92), 3.8)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestFillConstantOp2(OpTest):
+    def setUp(self):
+        '''Test fill_constant op with default value
+        '''
+        self.op_type = "fill_constant"
+
+        self.inputs = {}
+        self.attrs = {'shape': [123, 92]}
+        self.outputs = {'Out': np.full((123, 92), 0.0)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestInterpOp(OpTest):
+    def setUp(self):
+        self.op_type = "interp"
+        x = np.random.random((2, 3)).astype("float32")
+        y = np.random.random((2, 3)).astype("float32")
+        w = np.random.random(2).astype("float32")
+
+        sub_out = x - y
+        mul_out = sub_out * w.reshape(2, 1)
+        out = mul_out + y
+
+        self.inputs = {'X': x, 'Y': y, 'W': w}
+        self.outputs = {'Out': out, 'SubOut': sub_out, 'MulOut': mul_out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out')
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def max_pool3D_forward_naive(x,
+                             ksize,
+                             strides,
+                             paddings=[0, 0, 0],
+                             global_pool=0):
+
+    N, C, D, H, W = x.shape
+    if global_pool == 1:
+        ksize = [D, H, W]
+    D_out = (D - ksize[0] + 2 * paddings[0]) / strides[0] + 1
+    H_out = (H - ksize[1] + 2 * paddings[1]) / strides[1] + 1
+    W_out = (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1
+    out = np.zeros((N, C, D_out, H_out, W_out))
+    mask = np.zeros((N, C, D_out, H_out, W_out))
+    for k in xrange(D_out):
+        d_start = np.max((k * strides[0] - paddings[0], 0))
+        d_end = np.min((k * strides[0] + ksize[0] - paddings[0], D))
+        for i in xrange(H_out):
+            h_start = np.max((i * strides[0] - paddings[0], 0))
+            h_end = np.min((i * strides[0] + ksize[0] - paddings[0], H))
+            for j in xrange(W_out):
+                w_start = np.max((j * strides[1] - paddings[1], 0))
+                w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W))
+                x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end]
+
+                out[:, :, k, i, j] = np.max(x_masked, axis=(2, 3, 4))
+
+                for n in xrange(N):
+                    for c in xrange(C):
+                        arr = x_masked[n, c, :, :, :]
+                        index = np.where(arr == np.max(arr))
+                        sub_deep = index[0][0]
+                        sub_row = index[1][0]
+                        sub_col = index[2][0]
+                        index = ((d_start + sub_deep) * H +
+                                 (h_start + sub_row)) * W + w_start + sub_col
+                        mask[n, c, k, i, j] = index
+
+    return out, mask
+
+
+def max_pool2D_forward_naive(x, ksize, strides, paddings=[0, 0], global_pool=0):
+
+    N, C, H, W = x.shape
+    if global_pool == 1:
+        ksize = [H, W]
+    H_out = (H - ksize[0] + 2 * paddings[0]) / strides[0] + 1
+    W_out = (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1
+    out = np.zeros((N, C, H_out, W_out))
+    mask = np.zeros((N, C, H_out, W_out))
+    for i in xrange(H_out):
+        for j in xrange(W_out):
+            r_start = np.max((i * strides[0] - paddings[0], 0))
+            r_end = np.min((i * strides[0] + ksize[0] - paddings[0], H))
+            c_start = np.max((j * strides[1] - paddings[1], 0))
+            c_end = np.min((j * strides[1] + ksize[1] - paddings[1], W))
+            x_masked = x[:, :, r_start:r_end, c_start:c_end]
+
+            out[:, :, i, j] = np.max(x_masked, axis=(2, 3))
+
+            for n in xrange(N):
+                for c in xrange(C):
+                    arr = x_masked[n, c, :, :]
+                    index = np.where(arr == np.max(arr))
+                    sub_row = index[0][0]
+                    sub_col = index[1][0]
+                    index = (r_start + sub_row) * W + c_start + sub_col
+                    mask[n, c, i, j] = index
+
+    return out, mask
+
+
+class TestMaxPoolWithIndex_Op(OpTest):
+    def setUp(self):
+        self.initTestCase()
+        input = np.random.random(self.shape).astype("float32")
+        output, mask = self.pool_forward_naive(input, self.ksize, self.strides,
+                                               self.paddings, self.global_pool)
+
+        self.attrs = {
+            'strides': self.strides,
+            'paddings': self.paddings,
+            'ksize': self.ksize,
+            'globalPooling': self.global_pool,
+        }
+
+        self.inputs = {'X': input}
+        self.outputs = {'Out': output, "Mask": mask}
+
+    def test_check_output(self):
+        self.check_output()
+
+    # def test_check_grad(self):
+    #     self.check_grad(set(['X']), ['Out'], max_relative_error=0.07)
+
+    def initTestCase(self):
+        self.global_pool = True
+        self.index = "max_pool3d_with_index"
+        self.op_type = "%s" % self.index
+        self.pool_forward_naive = max_pool3D_forward_naive
+        self.shape = [2, 3, 5, 5, 5]
+        self.ksize = [3, 3, 3]
+        self.strides = [1, 1, 1]
+        self.paddings = [1, 1, 1]
+
+
+class TestCase1(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = True
+        self.op_type = "max_pool3d_with_index"
+        self.pool_forward_naive = max_pool3D_forward_naive
+        self.shape = [2, 3, 5, 5, 5]
+        self.ksize = [3, 3, 3]
+        self.strides = [1, 1, 1]
+        self.paddings = [1, 1, 1]
+
+
+class TestCase2(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = False
+        self.op_type = "max_pool3d_with_index"
+        self.pool_forward_naive = max_pool3D_forward_naive
+        self.shape = [2, 3, 7, 7, 7]
+        self.ksize = [3, 3, 3]
+        self.strides = [1, 1, 1]
+        self.paddings = [1, 1, 1]
+
+
+class TestCase3(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = False
+        self.op_type = "max_pool3d_with_index"
+        self.pool_forward_naive = max_pool3D_forward_naive
+        self.shape = [2, 3, 7, 7, 7]
+        self.ksize = [3, 3, 3]
+        self.strides = [2, 2, 2]
+        self.paddings = [0, 0, 0]
+
+
+class TestCase4(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = True
+        self.op_type = "max_pool3d_with_index"
+        self.pool_forward_naive = max_pool3D_forward_naive
+        self.shape = [2, 3, 5, 5, 5]
+        self.ksize = [3, 3, 3]
+        self.strides = [1, 1, 1]
+        self.paddings = [1, 1, 1]
+
+
+class TestCase5(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = True
+        self.op_type = "max_pool3d_with_index"
+        self.pool_forward_naive = max_pool3D_forward_naive
+        self.shape = [2, 3, 5, 5, 5]
+        self.ksize = [3, 3, 3]
+        self.strides = [2, 2, 2]
+        self.paddings = [0, 0, 0]
+
+
+class TestCase6(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = False
+        self.op_type = "max_pool2d_with_index"
+        self.pool_forward_naive = max_pool2D_forward_naive
+        self.shape = [2, 3, 7, 7]
+        self.ksize = [3, 3]
+        self.strides = [1, 1]
+        self.paddings = [1, 1]
+
+
+class TestCase7(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = False
+        self.op_type = "max_pool2d_with_index"
+        self.pool_forward_naive = max_pool2D_forward_naive
+        self.shape = [2, 3, 7, 7]
+        self.ksize = [3, 3]
+        self.strides = [2, 2]
+        self.paddings = [0, 0]
+
+
+class TestCase8(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = True
+        self.op_type = "max_pool2d_with_index"
+        self.pool_forward_naive = max_pool2D_forward_naive
+        self.shape = [2, 3, 5, 5]
+        self.ksize = [3, 3]
+        self.strides = [1, 1]
+        self.paddings = [1, 1]
+
+
+class TestCase9(TestMaxPoolWithIndex_Op):
+    def initTestCase(self):
+        self.global_pool = True
+        self.op_type = "max_pool2d_with_index"
+        self.pool_forward_naive = max_pool2D_forward_naive
+        self.shape = [2, 3, 5, 5]
+        self.ksize = [3, 3]
+        self.strides = [2, 2]
+        self.paddings = [0, 0]
+
+
+if __name__ == '__main__':
+    unittest.main()