Initialize the benchmark tester for operator. (#15772)

* Initialize the benchmark tester for operator.
test=develop

* Rearrange the codes.
test=develop

paddle/fluid/operators/CMakeLists.txt

@@ -97,3 +97,4 @@ if (WITH_PYTHON)
 endif()
 
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
+add_subdirectory(benchmark)

paddle/fluid/operators/benchmark/CMakeLists.txt

+cc_test(op_tester SRCS op_tester.cc op_tester_config.cc
+        DEPS memory timer framework_proto proto_desc lod_tensor op_registry
+        device_context scope ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS})

paddle/fluid/operators/benchmark/op_tester.cc

+/* Copyright (c) 2016 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. */
+
+#include "paddle/fluid/operators/benchmark/op_tester.h"
+#include "gflags/gflags.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/op_info.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/platform/init.h"
+#include "paddle/fluid/platform/profiler.h"
+#include "paddle/fluid/platform/timer.h"
+#include "paddle/fluid/pybind/pybind.h"
+
+namespace paddle {
+namespace operators {
+namespace benchmark {
+
+DEFINE_string(op_config_list, "", "Path of op config file.");
+
+void OpTester::Init(const std::string &filename) {
+  Init(OpTesterConfig(filename));
+}
+
+void OpTester::Init(const OpTesterConfig &config) {
+  config_ = config;
+
+  auto &op_desc_info = framework::OpInfoMap::Instance();
+  // Initialize the OpDesc
+  if (op_desc_info.Has(config_.op_type)) {
+    type_ = config_.op_type;
+    op_desc_.SetType(config_.op_type);
+
+    CreateInputVarDesc();
+    CreateOutputVarDesc();
+  } else {
+    LOG(FATAL) << "Op \"" << config_.op_type << "\" is not registered.";
+  }
+
+  if (config_.device_id >= 0) {
+    place_ = paddle::platform::CUDAPlace(config_.device_id);
+  } else {
+    place_ = paddle::platform::CPUPlace();
+  }
+
+  framework::InitDevices(false);
+  scope_.reset(new paddle::framework::Scope());
+
+  op_ = framework::OpRegistry::CreateOp(op_desc_);
+  CreateVariables(scope_.get());
+}
+
+void OpTester::Run() {
+  if (config_.print_debug_string) {
+    LOG(INFO) << DebugString();
+  }
+
+  // Warm up
+  RunImpl();
+
+  platform::Timer timer;
+  if (config_.profile) {
+    if (platform::is_cpu_place(place_)) {
+      platform::EnableProfiler(platform::ProfilerState::kCPU);
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      platform::EnableProfiler(platform::ProfilerState::kAll);
+      platform::SetDeviceId(config_.device_id);
+#else
+      PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
+#endif
+    }
+
+    timer.Start();
+    for (int i = config_.repeat; i > 0; --i) {
+      RunImpl();
+    }
+    timer.Pause();
+    platform::DisableProfiler(platform::EventSortingKey::kDefault,
+                              "op_tester_profiler");
+  } else {
+    timer.Start();
+    for (int i = config_.repeat; i > 0; --i) {
+      RunImpl();
+    }
+    timer.Pause();
+  }
+  config_.runtime = timer.ElapsedMS() / config_.repeat;
+  LOG(INFO) << "=== Run " << config_.repeat
+            << " times, latency: " << config_.runtime << " ms ===";
+}
+
+void OpTester::RunImpl() {
+  op_->Run(*scope_, place_);
+  platform::DeviceContextPool::Instance().Get(place_)->Wait();
+  scope_->DropKids();
+}
+
+std::vector<std::string> OpTester::GetOpProtoInputNames() {
+  std::vector<std::string> input_names;
+  const framework::proto::OpProto &proto =
+      framework::OpInfoMap::Instance().Get(type_).Proto();
+  for (int i = 0; i != proto.inputs_size(); ++i) {
+    const auto &input = proto.inputs(i);
+    input_names.push_back(input.name());
+  }
+  return input_names;
+}
+
+std::vector<std::string> OpTester::GetOpProtoOutputNames() {
+  std::vector<std::string> output_names;
+  const framework::proto::OpProto &proto =
+      framework::OpInfoMap::Instance().Get(type_).Proto();
+  for (int i = 0; i != proto.outputs_size(); ++i) {
+    const auto &output = proto.outputs(i);
+    output_names.push_back(output.name());
+  }
+  return output_names;
+}
+
+void OpTester::CreateInputVarDesc() {
+  std::vector<std::string> input_names = GetOpProtoInputNames();
+  for (auto &name : input_names) {
+    const OpInputConfig *input = config_.GetInput(name);
+    if (input == nullptr) {
+      LOG(FATAL) << "The input " << name << " of op " << config_.op_type
+                 << " is not correctlly provided.";
+    }
+
+    std::string var_name = config_.op_type + "." + name;
+    framework::VarDesc *var = Var(var_name);
+    // Need to support more type
+    var->SetType(framework::proto::VarType::LOD_TENSOR);
+    var->SetPersistable(false);
+    var->SetDataType(framework::proto::VarType::FP32);
+    var->SetShape(input->dims);
+
+    op_desc_.SetInput(name, {var_name});
+    inputs_.push_back(var_name);
+  }
+}
+
+void OpTester::CreateOutputVarDesc() {
+  std::vector<std::string> output_names = GetOpProtoOutputNames();
+  for (auto &name : output_names) {
+    std::string var_name = config_.op_type + "." + name;
+    framework::VarDesc *var = Var(var_name);
+    // Need to support more type
+    var->SetType(framework::proto::VarType::LOD_TENSOR);
+    var->SetPersistable(false);
+    var->SetDataType(framework::proto::VarType::FP32);
+
+    op_desc_.SetOutput(name, {var_name});
+    outputs_.push_back(var_name);
+  }
+}
+
+framework::VarDesc *OpTester::Var(const std::string &name) {
+  auto it = vars_.find(name);
+  if (it != vars_.end()) {
+    return it->second.get();
+  }
+  auto *var = new framework::VarDesc(name);
+  vars_[name].reset(var);
+  return var;
+}
+
+template <typename T>
+void OpTester::SetupTensor(framework::LoDTensor *tensor,
+                           const std::vector<int64_t> &shape, T lower,
+                           T upper) {
+  static unsigned int seed = 100;
+  std::mt19937 rng(seed++);
+  std::uniform_real_distribution<double> uniform_dist(0, 1);
+
+  T *ptr = tensor->mutable_data<T>(framework::make_ddim(shape), place_);
+  if (platform::is_cpu_place(place_)) {
+    for (int i = 0; i < tensor->numel(); ++i) {
+      ptr[i] = static_cast<T>(uniform_dist(rng) * (upper - lower) + lower);
+    }
+  } else {
+    framework::LoDTensor cpu_tensor;
+    T *cpu_ptr = cpu_tensor.mutable_data<T>(framework::make_ddim(shape),
+                                            platform::CPUPlace());
+    for (int i = 0; i < cpu_tensor.numel(); ++i) {
+      cpu_ptr[i] = static_cast<T>(uniform_dist(rng) * (upper - lower) + lower);
+    }
+    TensorCopySync(cpu_tensor, place_, tensor);
+  }
+}
+
+void OpTester::CreateVariables(framework::Scope *scope) {
+  for (auto &item : vars_) {
+    auto &var = item.second;
+    if (var->Name() == framework::kEmptyVarName) {
+      continue;
+    }
+
+    auto *ptr = scope->Var(var->Name());
+    framework::InitializeVariable(ptr, var->GetType());
+    if (var->Persistable()) {
+      VLOG(3) << "Create Variable " << var->Name()
+              << " global, which pointer is " << ptr;
+    } else {
+      VLOG(3) << "Create Variable " << var->Name()
+              << " locally, which pointer is " << ptr;
+    }
+  }
+
+  // Allocate memory for input tensor
+  for (auto &name : inputs_) {
+    VLOG(3) << "Allocate memory for tensor " << name;
+    auto &var_desc = vars_[name];
+    std::vector<int64_t> shape = var_desc->GetShape();
+
+    auto *var = scope->Var(name);
+    auto *tensor = var->GetMutable<framework::LoDTensor>();
+    SetupTensor<float>(tensor, shape, static_cast<float>(0.0),
+                       static_cast<float>(1.0));
+  }
+}
+
+static std::string GenSpaces(int count) {
+  std::stringstream ss;
+  for (int i = 0; i < count; ++i) {
+    ss << "  ";
+  }
+  return ss.str();
+}
+
+std::string OpTester::DebugString() {
+  std::stringstream ss;
+  int count = 0;
+  for (auto &item : vars_) {
+    auto &var = item.second;
+    ss << GenSpaces(count++) << "vars {\n";
+    ss << GenSpaces(count) << "name: \"" << var->Name() << "\"\n";
+    ss << GenSpaces(count++) << "type: {\n";
+    ss << GenSpaces(count) << "type: LOD_TENSOR\n";
+    ss << GenSpaces(count++) << "lod_tensor {\n";
+    ss << GenSpaces(count++) << "tensor {\n";
+    ss << GenSpaces(count) << "data_type: FP32\n";
+    std::vector<int64_t> shape = var->GetShape();
+    for (auto d : shape) {
+      ss << GenSpaces(count) << "dims: " << d << "\n";
+    }
+    ss << GenSpaces(--count) << "}\n";
+    ss << GenSpaces(--count) << "}\n";
+    ss << GenSpaces(--count) << "}\n";
+    ss << GenSpaces(count) << "persistable: " << var->Persistable() << "\n";
+    ss << GenSpaces(--count) << "}\n";
+  }
+  ss << GenSpaces(count++) << "ops {\n";
+  for (auto &name : op_desc_.InputNames()) {
+    ss << GenSpaces(count++) << "inputs {\n";
+    ss << GenSpaces(count) << "parameters: \"" << name << "\"\n";
+    ss << GenSpaces(count) << "arguments: \"" << op_desc_.Input(name)[0]
+       << "\"\n";
+    ss << GenSpaces(--count) << "}\n";
+  }
+  for (auto &name : op_desc_.OutputNames()) {
+    ss << GenSpaces(count++) << "outputs {\n";
+    ss << GenSpaces(count) << "parameters: \"" << name << "\"\n";
+    ss << GenSpaces(count) << "arguments: \"" << op_desc_.Output(name)[0]
+       << "\"\n";
+    ss << GenSpaces(--count) << "}\n";
+  }
+  ss << GenSpaces(count) << "type: " << op_desc_.Type() << "\n";
+  ss << GenSpaces(--count) << "}\n";
+  return ss.str();
+}
+
+TEST(op_tester, base) {
+  OpTester tester;
+  if (!FLAGS_op_config_list.empty()) {
+    tester.Init(FLAGS_op_config_list);
+  } else {
+    OpTesterConfig config;
+    config.op_type = "elementwise_add";
+    config.inputs.resize(2);
+    config.inputs[0].name = "X";
+    config.inputs[0].dims = {64, 64};
+    config.inputs[1].name = "Y";
+    config.inputs[1].dims = {64, 1};
+    tester.Init(config);
+  }
+  tester.Run();
+}
+
+}  // namespace benchmark
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/benchmark/op_tester.h

+/* Copyright (c) 2016 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 <string>
+#include <vector>
+#include "paddle/fluid/framework/ddim.h"
+#include "paddle/fluid/framework/op_desc.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/operators/benchmark/op_tester_config.h"
+
+namespace paddle {
+namespace operators {
+namespace benchmark {
+
+class OpTester {
+ public:
+  OpTester() {}
+
+  void Init(const std::string &filename);
+  void Init(const OpTesterConfig &config);
+
+  void Run();
+
+  std::string DebugString();
+
+ private:
+  std::vector<std::string> GetOpProtoInputNames();
+  std::vector<std::string> GetOpProtoOutputNames();
+
+  void CreateInputVarDesc();
+  void CreateOutputVarDesc();
+
+  framework::VarDesc *Var(const std::string &name);
+  void CreateVariables(framework::Scope *scope);
+
+  template <typename T>
+  void SetupTensor(framework::LoDTensor *input,
+                   const std::vector<int64_t> &shape, T lower, T upper);
+
+  void RunImpl();
+
+ private:
+  OpTesterConfig config_;
+  std::string type_;
+  framework::OpDesc op_desc_;
+  std::unordered_map<std::string, std::unique_ptr<framework::VarDesc>> vars_;
+  std::vector<std::string> inputs_;
+  std::vector<std::string> outputs_;
+  std::unique_ptr<framework::OperatorBase> op_;
+  platform::Place place_;
+  std::unique_ptr<framework::Scope> scope_;
+};
+
+}  // namespace benchmark
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/benchmark/op_tester_config.cc

+/* Copyright (c) 2016 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. */
+
+#include "paddle/fluid/operators/benchmark/op_tester_config.h"
+#include <fstream>
+#include "glog/logging.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace benchmark {
+
+static const char kStartSeparator[] = "{";
+static const char kEndSeparator[] = "}";
+static const char kSepBetweenItems[] = ";";
+
+static bool StartWith(const std::string& str, const std::string& substr) {
+  return str.find(substr) == 0;
+}
+
+static bool EndWith(const std::string& str, const std::string& substr) {
+  return str.rfind(substr) == (str.length() - substr.length());
+}
+
+static void EraseEndSep(std::string* str) {
+  std::string substr = kSepBetweenItems;
+  if (EndWith(*str, substr)) {
+    str->erase(str->length() - substr.length(), str->length());
+  }
+}
+
+static std::vector<int64_t> ParseDims(std::string dims_str) {
+  std::vector<int64_t> dims;
+  std::string token;
+  std::istringstream token_stream(dims_str);
+  while (std::getline(token_stream, token, 'x')) {
+    dims.push_back(std::stoi(token));
+  }
+  return dims;
+}
+
+OpInputConfig::OpInputConfig(std::istream& is) {
+  std::string sep;
+  is >> sep;
+  if (sep == kStartSeparator) {
+    while (sep != kEndSeparator) {
+      is >> sep;
+      if (sep == "name" || sep == "name:") {
+        is >> name;
+        EraseEndSep(&name);
+      } else if (sep == "dims" || sep == "dims:") {
+        std::string dims_str;
+        is >> dims_str;
+        dims = ParseDims(dims_str);
+      }
+    }
+  }
+}
+
+OpTesterConfig::OpTesterConfig(const std::string& filename) {
+  std::ifstream fin(filename, std::ios::in | std::ios::binary);
+  PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s",
+                 filename.c_str());
+
+  Init(fin);
+}
+
+void OpTesterConfig::Init(std::istream& is) {
+  std::string sep;
+  is >> sep;
+  if (sep == kStartSeparator) {
+    while (sep != kEndSeparator) {
+      is >> sep;
+      if (sep == "op_type" || sep == "op_type:") {
+        is >> op_type;
+      } else if (sep == "device_id" || sep == "device_id:") {
+        is >> device_id;
+      } else if (sep == "repeat" || sep == "repeat:") {
+        is >> repeat;
+      } else if (sep == "profile" || sep == "profile:") {
+        is >> profile;
+      } else if (sep == "print_debug_string" || sep == "print_debug_string:") {
+        is >> print_debug_string;
+      } else if (sep == "input" || sep == "input:") {
+        OpInputConfig input_config(is);
+        inputs.push_back(input_config);
+      }
+    }
+  }
+}
+
+const OpInputConfig* OpTesterConfig::GetInput(const std::string& name) {
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    if (inputs[i].name == name) {
+      return &inputs[i];
+    }
+  }
+  return nullptr;
+}
+
+}  // namespace benchmark
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/benchmark/op_tester_config.h

+/* Copyright (c) 2016 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 <istream>
+#include <string>
+#include <vector>
+
+namespace paddle {
+namespace operators {
+namespace benchmark {
+
+struct OpInputConfig {
+  OpInputConfig() {}
+  explicit OpInputConfig(std::istream& is);
+
+  std::string name;
+  std::vector<int64_t> dims;
+};
+
+struct OpTesterConfig {
+  OpTesterConfig() {}
+  explicit OpTesterConfig(const std::string& filename);
+  void Init(std::istream& is);
+
+  const OpInputConfig* GetInput(const std::string& name);
+
+  std::string op_type;
+  std::vector<OpInputConfig> inputs;
+  int device_id{-1};  // CPU: -1
+  int repeat{1};
+  int profile{0};
+  int print_debug_string{0};
+  double runtime{0.0};
+};
+
+}  // namespace benchmark
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/test.cc

 /* 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. */
+
+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 <random>
 #include <string>
@@ -259,7 +259,7 @@ struct TestFuncWithRefer<jit::SeqPoolTuples<T>, std::vector<T>, std::vector<T>,
                   const std::vector<T>& x, const std::vector<T>& yref,
                   const typename jit::SeqPoolTuples<T>::attr_type& attr) {
     EXPECT_TRUE(tgt != nullptr);
-    EXPECT_EQ(x.size() % yref.size(), 0);
+    EXPECT_EQ(x.size() % yref.size(), static_cast<size_t>(0));
     int w = yref.size();
     std::vector<T> y(w);
     const T* x_data = x.data();