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

doc/fluid/dev/api_doc_std_cn.md

@@ -1,8 +1,9 @@
 # API注释撰写标准
 
-- [API注释模块](#API注释模块)
-- [格式及示例](#格式及示例)
-- [完整示例](#完整示例)
+- [API注释撰写标准](#api)
+    - [API注释模块](#api)
+    - [格式及示例](#)
+    - [完整示例](#)
 
 
 ## API注释模块
@@ -217,4 +218,4 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接
 
 ## 完整示例
 
-fc 的完整注释见[示例](src/fc.py)。
+fc 的完整注释见[示例](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。

doc/fluid/dev/api_doc_std_en.md

@@ -1,8 +1,9 @@
 # API Doc Standard
 
-- [API Doc Structure](#API Doc Structure)
-- [Format and Examples](#Format and Examples)
-- [Complete Example](#Complete Example)
+- [API Doc Standard](#api-doc-standard)
+    - [API Doc Structure](#api-doc-structure)
+    - [Format and Examples](#format-and-examples)
+    - [Complete Example](#complete-example)
 
 
 ## API Doc Structure
@@ -223,4 +224,4 @@ Format and examples of each part of API documantation are as follows: (take fc f
 
 ## Complete Example
 
-Complete Example of fc please see [here](src/fc.py)。
+Complete Example of fc please see [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。

paddle/contrib/inference/test_paddle_inference_api_impl.cc

@@ -109,7 +109,6 @@ void MainWord2Vec(bool use_gpu) {
 
 void MainImageClassification(bool use_gpu) {
   int batch_size = 2;
-  bool use_mkldnn = false;
   bool repeat = false;
   NativeConfig config = GetConfig();
   config.use_gpu = use_gpu;
@@ -134,12 +133,8 @@ void MainImageClassification(bool use_gpu) {
   std::vector<framework::LoDTensor*> cpu_fetchs1;
   cpu_fetchs1.push_back(&output1);
 
-  TestInference<platform::CPUPlace, false, true>(config.model_dir,
-                                                 cpu_feeds,
-                                                 cpu_fetchs1,
-                                                 repeat,
-                                                 is_combined,
-                                                 use_mkldnn);
+  TestInference<platform::CPUPlace, false, true>(
+      config.model_dir, cpu_feeds, cpu_fetchs1, repeat, is_combined);
 
   auto predictor = CreatePaddlePredictor(config);
   std::vector<PaddleTensor> paddle_tensor_feeds;

paddle/fluid/framework/data_type.cc

@@ -28,6 +28,9 @@ struct DataTypeMap {
 };
 
 static DataTypeMap* InitDataTypeMap();
+// C++11 removes the need for manual locking. Concurrent execution shall wait if
+// a static local variable is already being initialized.
+// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex
 static DataTypeMap& gDataTypeMap() {
   static DataTypeMap* g_data_type_map_ = InitDataTypeMap();
   return *g_data_type_map_;

paddle/fluid/framework/details/fuse_vars_op_handle.cc

@@ -42,7 +42,7 @@ void FuseVarsOpHandle::RunImpl() {
     out_t->ShareDataWith(out_tensor->Slice(s, s + numel));
     s += numel;
   }
-  this->RunAndRecordEvent([this] {});
+  this->RunAndRecordEvent([] {});
 }
 
 std::string FuseVarsOpHandle::Name() const { return "fuse vars"; }

paddle/fluid/framework/executor.cc

@@ -24,6 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/profiler.h"
 
 DECLARE_bool(benchmark);
+DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
 
 namespace paddle {
 namespace framework {
@@ -115,6 +116,7 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
                    bool create_local_scope, bool create_vars) {
   platform::RecordBlock b(block_id);
+  if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
   auto ctx = Prepare(pdesc, block_id);
   RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
 }
@@ -214,6 +216,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
                    const std::string& feed_holder_name,
                    const std::string& fetch_holder_name) {
   platform::RecordBlock b(kProgramId);
+  if (FLAGS_use_mkldnn) EnableMKLDNN(program);
   bool has_feed_ops =
       has_feed_operators(program.Block(0), *feed_targets, feed_holder_name);
   bool has_fetch_ops =
@@ -225,7 +228,6 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
     unique_ptr_of_copy_program.reset(new ProgramDesc(program));
     copy_program = unique_ptr_of_copy_program.get();
   }
-
   auto* global_block = copy_program->MutableBlock(0);
 
   if (!has_feed_ops) {
@@ -378,5 +380,19 @@ void Executor::RunPreparedContext(
   }
 }
 
+void Executor::EnableMKLDNN(const ProgramDesc& program) {
+#ifdef PADDLE_WITH_MKLDNN
+  VLOG(3) << "use_mkldnn=True";
+  for (size_t bid = 0; bid < program.Size(); ++bid) {
+    auto* block = const_cast<ProgramDesc&>(program).MutableBlock(bid);
+    for (auto* op : block->AllOps()) {
+      if (op->HasAttr("use_mkldnn")) {
+        op->SetAttr("use_mkldnn", true);
+      }
+    }
+  }
+#endif
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/executor.h

@@ -81,6 +81,8 @@ class Executor {
                           const std::string& feed_holder_name = "feed",
                           const std::string& fetch_holder_name = "fetch");
 
+  void EnableMKLDNN(const ProgramDesc& program);
+
  private:
   const platform::Place place_;
 };

paddle/fluid/inference/tests/book/test_inference_image_classification.cc

@@ -21,7 +21,6 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model.");
 DEFINE_int32(batch_size, 1, "Batch size of input data");
 DEFINE_int32(repeat, 1, "Running the inference program repeat times");
 DEFINE_bool(skip_cpu, false, "Skip the cpu test");
-DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference");
 
 TEST(inference, image_classification) {
   if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
@@ -59,10 +58,8 @@ TEST(inference, image_classification) {
     // Run inference on CPU
     LOG(INFO) << "--- CPU Runs: ---";
     LOG(INFO) << "Batch size is " << FLAGS_batch_size;
-    LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn;
     TestInference<paddle::platform::CPUPlace, false, true>(
-        dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined,
-        FLAGS_use_mkldnn);
+        dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined);
     LOG(INFO) << output1.dims();
   }
 

paddle/fluid/inference/tests/book/test_inference_nlp.cc

@@ -27,7 +27,6 @@ limitations under the License. */
 DEFINE_string(model_path, "", "Directory of the inference model.");
 DEFINE_string(data_file, "", "File of input index data.");
 DEFINE_int32(repeat, 100, "Running the inference program repeat times");
-DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference");
 DEFINE_bool(prepare_vars, true, "Prepare variables before executor");
 DEFINE_int32(num_threads, 1, "Number of threads should be used");
 
@@ -190,9 +189,6 @@ TEST(inference, nlp) {
     std::unique_ptr<paddle::framework::ProgramDesc> inference_program;
     inference_program = InitProgram(&executor, scope.get(), FLAGS_model_path,
                                     /*model combined*/ false);
-    if (FLAGS_use_mkldnn) {
-      EnableMKLDNN(inference_program);
-    }
     // always prepare context
     std::unique_ptr<paddle::framework::ExecutorPrepareContext> ctx;
     ctx = executor.Prepare(*inference_program, 0);

paddle/fluid/inference/tests/test_helper.h

@@ -22,6 +22,8 @@ limitations under the License. */
 #include "paddle/fluid/inference/io.h"
 #include "paddle/fluid/platform/profiler.h"
 
+DECLARE_bool(use_mkldnn);
+
 template <typename T>
 void SetupTensor(paddle::framework::LoDTensor* input,
                  paddle::framework::DDim dims, T lower, T upper) {
@@ -133,24 +135,11 @@ std::vector<std::vector<int64_t>> GetFeedTargetShapes(
   return feed_target_shapes;
 }
 
-void EnableMKLDNN(
-    const std::unique_ptr<paddle::framework::ProgramDesc>& program) {
-  for (size_t bid = 0; bid < program->Size(); ++bid) {
-    auto* block = program->MutableBlock(bid);
-    for (auto* op : block->AllOps()) {
-      if (op->HasAttr("use_mkldnn")) {
-        op->SetAttr("use_mkldnn", true);
-      }
-    }
-  }
-}
-
 template <typename Place, bool CreateVars = true, bool PrepareContext = false>
 void TestInference(const std::string& dirname,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_fetchs,
-                   const int repeat = 1, const bool is_combined = false,
-                   const bool use_mkldnn = false) {
+                   const int repeat = 1, const bool is_combined = false) {
   // 1. Define place, executor, scope
   auto place = Place();
   auto executor = paddle::framework::Executor(place);
@@ -182,9 +171,6 @@ void TestInference(const std::string& dirname,
         "init_program",
         paddle::platform::DeviceContextPool::Instance().Get(place));
     inference_program = InitProgram(&executor, scope, dirname, is_combined);
-    if (use_mkldnn) {
-      EnableMKLDNN(inference_program);
-    }
   }
   // Disable the profiler and print the timing information
   paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
@@ -210,7 +196,10 @@ void TestInference(const std::string& dirname,
     fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
   }
 
-  // 6. Run the inference program
+  // 6. If export Flags_use_mkldnn=True, use mkldnn related ops.
+  if (FLAGS_use_mkldnn) executor.EnableMKLDNN(*inference_program);
+
+  // 7. Run the inference program
   {
     if (!CreateVars) {
       // If users don't want to create and destroy variables every time they

paddle/fluid/operators/arg_max_op.cc

@@ -12,24 +12,22 @@ 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/fluid/operators/arg_max_op.h"
+#include "paddle/fluid/operators/arg_min_max_op_base.h"
 
-REGISTER_OPERATOR(arg_max, paddle::operators::ArgMaxOp,
+REGISTER_OPERATOR(arg_max, paddle::operators::ArgMinMaxOp,
                   paddle::operators::ArgMaxOpMaker,
                   paddle::framework::EmptyGradOpMaker);
 
 REGISTER_OP_CPU_KERNEL(
-    arg_max, paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
-                                             float, int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, double,
+    arg_max,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, float>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, double>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
                                     int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, int64_t,
-                                    int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, int32_t,
-                                    int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, int16_t,
-                                    int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, size_t,
-                                    int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, uint8_t,
-                                    int64_t>);
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
+                                    int32_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
+                                    int16_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, size_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
+                                    uint8_t>);

paddle/fluid/operators/arg_max_op.cu

@@ -12,21 +12,20 @@ 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/fluid/operators/arg_max_op.h"
+#include "paddle/fluid/operators/arg_min_max_op_base.h"
 
 REGISTER_OP_CUDA_KERNEL(
     arg_max,
-    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext, float,
-                                    int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext, double,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext, float>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
+                                    double>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
                                     int64_t>,
     paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
-                                    int64_t, int64_t>,
+                                    int32_t>,
     paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
-                                    int32_t, int64_t>,
+                                    int16_t>,
     paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
-                                    int16_t, int64_t>,
-    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext, size_t,
-                                    int64_t>,
+                                    size_t>,
     paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
-                                    uint8_t, int64_t>);
+                                    uint8_t>);

paddle/fluid/operators/arg_max_op.h

@@ -1,16 +0,0 @@
-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
-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/fluid/operators/arg_min_max_op_base.h"

paddle/fluid/operators/arg_min_max_op_base.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <string>
 #include <type_traits>
 #include <vector>
 #include "paddle/fluid/framework/ddim.h"
@@ -37,9 +38,9 @@ struct ArgMinMaxFunctor {};
   struct ArgMinMaxFunctor<DeviceContext, T, Tout, Rank,                       \
                           enum_argminmax_value> {                             \
     void operator()(const DeviceContext& ctx, const framework::LoDTensor& in, \
-                    framework::LoDTensor& out, int64_t axis) {                \
+                    framework::LoDTensor* out, int64_t axis) {                \
       auto in_eigen = framework::EigenTensor<T, Rank>::From(in);              \
-      auto out_eigen = framework::EigenTensor<Tout, Rank - 1>::From(out);     \
+      auto out_eigen = framework::EigenTensor<Tout, Rank - 1>::From(*out);    \
       out_eigen.device(*(ctx.eigen_device())) =                               \
           in_eigen.eigen_op_type(axis).template cast<Tout>();                 \
     }                                                                         \
@@ -62,7 +63,7 @@ class ArgMinMaxKernel : public framework::OpKernel<T> {
 #define CALL_ARG_MINMAX_FUNCTOR(rank)                                \
   ArgMinMaxFunctor<DeviceContext, T, Tout, rank, EnumArgMinMaxValue> \
       functor##rank;                                                 \
-  functor##rank(dev_ctx, x, out, axis)
+  functor##rank(dev_ctx, x, &out, axis)
 
     switch (x.dims().size()) {
       case 1:
@@ -89,19 +90,20 @@ class ArgMinMaxKernel : public framework::OpKernel<T> {
             "than 6.",
             (EnumArgMinMaxValue == kArgMin ? "argmin" : "argmax"));
         break;
+#undef CALL_ARG_MINMAX_FUNCTOR
     }
   }
 };
 
-template <typename DeviceContext, typename T, typename Tout>
+template <typename DeviceContext, typename T>
 using ArgMinKernel =
-    ArgMinMaxKernel<DeviceContext, T, Tout, ArgMinMaxType::kArgMin>;
+    ArgMinMaxKernel<DeviceContext, T, int64_t, ArgMinMaxType::kArgMin>;
 
-template <typename DeviceContext, typename T, typename Tout>
+template <typename DeviceContext, typename T>
 using ArgMaxKernel =
-    ArgMinMaxKernel<DeviceContext, T, Tout, ArgMinMaxType::kArgMax>;
+    ArgMinMaxKernel<DeviceContext, T, int64_t, ArgMinMaxType::kArgMax>;
 
-typedef class BaseArgMinMaxOp : public framework::OperatorWithKernel {
+class ArgMinMaxOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
@@ -121,7 +123,7 @@ typedef class BaseArgMinMaxOp : public framework::OperatorWithKernel {
     for (int64_t i = axis + 1; i < x_rank; i++) vec.push_back(x_dims[i]);
     ctx->SetOutputDim("Out", framework::make_ddim(vec));
   }
-} ArgMinOp, ArgMaxOp;
+};
 
 class BaseArgMinMaxOpMaker : public framework::OpProtoAndCheckerMaker {
  protected:
@@ -133,12 +135,13 @@ class BaseArgMinMaxOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "Input tensor.");
     AddOutput("Out", "Output tensor.");
     AddAttr<int64_t>("axis", "The axis in which to compute the arg indics.");
-    AddComment(::paddle::string::Sprintf(R"DOC(
-				%s Operator.
+    AddComment(string::Sprintf(R"DOC(
+      %s Operator.
 
-				Computes the indices of the %s elements of the input tensor's element along the provided axis.
+      Computes the indices of the %s elements of the input tensor's element
+      along the provided axis.
 )DOC",
-                                         OpName(), Name()));
+                               OpName(), Name()));
   }
 };
 

paddle/fluid/operators/arg_min_op.cc

@@ -12,24 +12,22 @@ 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/fluid/operators/arg_min_op.h"
+#include "paddle/fluid/operators/arg_min_max_op_base.h"
 
-REGISTER_OPERATOR(arg_min, paddle::operators::ArgMinOp,
+REGISTER_OPERATOR(arg_min, paddle::operators::ArgMinMaxOp,
                   paddle::operators::ArgMinOpMaker,
                   paddle::framework::EmptyGradOpMaker);
 
 REGISTER_OP_CPU_KERNEL(
-    arg_min, paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
-                                             float, int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, double,
+    arg_min,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, float>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, double>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
                                     int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, int64_t,
-                                    int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, int32_t,
-                                    int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, int16_t,
-                                    int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, size_t,
-                                    int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, uint8_t,
-                                    int64_t>);
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
+                                    int32_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
+                                    int16_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, size_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
+                                    uint8_t>);

paddle/fluid/operators/arg_min_op.cu

@@ -12,21 +12,20 @@ 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/fluid/operators/arg_min_op.h"
+#include "paddle/fluid/operators/arg_min_max_op_base.h"
 
 REGISTER_OP_CUDA_KERNEL(
     arg_min,
-    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext, float,
-                                    int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext, double,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext, float>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
+                                    double>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
                                     int64_t>,
     paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
-                                    int64_t, int64_t>,
+                                    int32_t>,
     paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
-                                    int32_t, int64_t>,
+                                    int16_t>,
     paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
-                                    int16_t, int64_t>,
-    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext, size_t,
-                                    int64_t>,
+                                    size_t>,
     paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
-                                    uint8_t, int64_t>);
+                                    uint8_t>);

paddle/fluid/operators/arg_min_op.h

@@ -1,16 +0,0 @@
-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
-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/fluid/operators/arg_min_max_op_base.h"

paddle/fluid/operators/batch_size_like.h

@@ -54,18 +54,18 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel {
 class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() final {
-    AddInput("Input",
-             "(Tensor) Tensor "
-             "whose input_dim_idx'th dimension specifies the batch_size");
+    AddInput(
+        "Input",
+        "Tensor whose input_dim_idx'th dimension specifies the batch_size");
     AddOutput("Out",
-              "(Tensor) Tensor of specified shape will be filled "
+              "Tensor of specified shape will be filled "
               "with the specified value");
-    AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
+    AddAttr<std::vector<int>>("shape", "The shape of the output");
     AddAttr<int>("input_dim_idx",
-                 "(int, default 0) The index of input's batch size dimension")
+                 "default 0. The index of input's batch size dimension")
         .SetDefault(0);
     AddAttr<int>("output_dim_idx",
-                 "(int, default 0) The index of output's batch size dimension")
+                 "default 0. The index of output's batch size dimension")
         .SetDefault(0);
     Apply();
   }

paddle/fluid/operators/bilinear_interp_op.cc

@@ -56,17 +56,16 @@ class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("X",
-             "(Tensor) The input tensor of bilinear interpolation, "
+             "The input tensor of bilinear interpolation, "
              "This is a 4-D tensor with shape of (N x C x h x w)");
     AddInput("OutSize",
-             "(Tensor) This is a 1-D tensor with two number. "
+             "This is a 1-D tensor with two number. "
              "The first number is height and the second number is width.")
         .AsDispensable();
-    AddOutput("Out",
-              "(Tensor) The dimension of output is (N x C x out_h x out_w]");
+    AddOutput("Out", "The dimension of output is (N x C x out_h x out_w)");
 
-    AddAttr<int>("out_h", "(int) output height of bilinear interpolation op.");
-    AddAttr<int>("out_w", "(int) output width of bilinear interpolation op.");
+    AddAttr<int>("out_h", "output height of bilinear interpolation op.");
+    AddAttr<int>("out_w", "output width of bilinear interpolation op.");
     AddComment(R"DOC(
           Bilinear interpolation is an extension of linear interpolation for 
           interpolating functions of two variables (e.g. H-direction and 

paddle/fluid/operators/detail/request_handler_impl.cc

@@ -64,13 +64,21 @@ bool RequestSendHandler::Handle(const std::string& varname,
       return false;
     }
     if (invar->IsType<framework::SelectedRows>()) {
-      rpc_server_->RecordSparseVar(invar);
+      std::unique_lock<std::mutex> lock(mutex_sparse_vars_);
+      sparse_vars_.push_back(invar);
     }
   }
-
   return true;
 }
 
+void RequestSendHandler::ResetSparseVarRecorder() {
+  std::unique_lock<std::mutex> lock(mutex_sparse_vars_);
+  for (auto* var : sparse_vars_) {
+    var->GetMutable<framework::SelectedRows>()->mutable_rows()->clear();
+  }
+  sparse_vars_.clear();
+}
+
 bool RequestGetHandler::Handle(const std::string& varname,
                                framework::Scope* scope,
                                framework::Variable* invar,

paddle/fluid/operators/detail/request_handler_impl.h

@@ -41,6 +41,11 @@ class RequestSendHandler final : public RequestHandler {
   virtual ~RequestSendHandler() {}
   bool Handle(const std::string& varname, framework::Scope* scope,
               framework::Variable* var, framework::Variable** outvar) override;
+  void ResetSparseVarRecorder();
+
+ private:
+  std::mutex mutex_sparse_vars_;
+  std::vector<framework::Variable*> sparse_vars_;
 };
 
 class RequestGetHandler final : public RequestHandler {

paddle/fluid/operators/detail/rpc_server.cc

@@ -73,19 +73,6 @@ void RPCServer::ResetBarrierCounter() {
     t.second = 0;
   }
 }
-void RPCServer::RecordSparseVar(framework::Variable* sparse_var) {
-  std::unique_lock<std::mutex> lock(mutex_sparse_var_recorder_);
-  sparse_vars_.push_back(sparse_var);
-}
-
-void RPCServer::ResetSparseVarsRecorder() {
-  VLOG(3) << "RPCServer reset sparse vars recorder.";
-  std::unique_lock<std::mutex> lock(mutex_sparse_var_recorder_);
-  for (auto* var : sparse_vars_) {
-    var->GetMutable<framework::SelectedRows>()->mutable_rows()->clear();
-  }
-  sparse_vars_.clear();
-}
 
 void RPCServer::RegisterRPC(const std::string& rpc_name,
                             RequestHandler* handler, int thread_num) {

paddle/fluid/operators/detail/rpc_server.h

@@ -62,8 +62,6 @@ class RPCServer {
   void IncreaseBatchBarrier(const std::string rpc_name);
 
   void ResetBarrierCounter();
-  void RecordSparseVar(framework::Variable* sparse_var);
-  void ResetSparseVarsRecorder();
 
  protected:
   virtual void ShutDownImpl() = 0;
@@ -77,9 +75,6 @@ class RPCServer {
   std::atomic<int> cur_cond_;
   std::condition_variable rpc_cond_;
 
-  std::vector<framework::Variable*> sparse_vars_;
-  std::mutex mutex_sparse_var_recorder_;
-
  protected:
   std::string bind_address_;
   std::atomic<int> exit_flag_;

paddle/fluid/operators/fill_constant_batch_size_like_op.cc

@@ -32,16 +32,16 @@ class FillConstantBatchSizeLikeOp : public BatchSizeLikeOp {
 class FillConstantBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
  protected:
   void Apply() override {
-    AddAttr<int>("dtype",
-                 "(int, default 5 (FP32)) "
-                 "Output data type")
+    AddAttr<int>(
+        "dtype",
+        "It could be numpy.dtype. Output data type. Default is float32")
         .SetDefault(framework::proto::VarType::FP32);
-    AddAttr<float>("value", "(float, default 0) The value to be filled")
+    AddAttr<float>("value", "default 0. The value to be filled")
         .SetDefault(0.0f);
     AddComment(R"DOC(
-FillConstantBatchSizeLike Operator.
-
-Fill up a variable with specified constant value.
+This function creates a tensor of specified *shape*, *dtype* and batch size,
+and initializes this with a constant supplied in *value*. The batch size is
+obtained from the `input` tensor.
 
 )DOC");
   }

paddle/fluid/operators/gather_test.cc

@@ -43,7 +43,8 @@ TEST(Gather, GatherData) {
   auto* cpu_place = new paddle::platform::CPUPlace();
   paddle::platform::CPUDeviceContext ctx(*cpu_place);
   paddle::operators::CPUGather<int>(ctx, *src, *index, output);
-
+  delete cpu_place;
+  cpu_place = NULL;
   for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
   for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);