Feature/global context (#6537)

* "add DeviceContextPool"

* "add devicecontextpool in pybind"

* "add comments in python side "

* "fix static link error"

* "fix CI error"

* "add executor.py"

* "fix CI error"

* "add with gpu macro"

* "remove comment out codes"

* "add TODO items"

* "update init devices"

paddle/framework/CMakeLists.txt

@@ -58,3 +58,6 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
         proto_desc)
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
+
+cc_library(init SRCS init.cc DEPS gflags executor place stringpiece)
+cc_test(init_test SRCS init_test.cc DEPS init)

paddle/framework/ddim_test.cc

@@ -1,3 +1,16 @@
+/* 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 <sstream>
 #include <vector>
 

paddle/framework/executor.cc

@@ -33,32 +33,12 @@ namespace framework {
 const std::string kFeedOpType = "feed";
 const std::string kFetchOpType = "fetch";
 
-Executor::Executor(const std::vector<platform::Place>& places) : own_(true) {
+DeviceContextPool* DeviceContextPool::pool = nullptr;
-  PADDLE_ENFORCE_GT(places.size(), 0);
-  device_contexts_.resize(places.size());
-  for (size_t i = 0; i < places.size(); i++) {
-    if (platform::is_cpu_place(places[i])) {
-      device_contexts_[i] = new platform::CPUDeviceContext(
-          boost::get<platform::CPUPlace>(places[i]));
-    } else if (platform::is_gpu_place(places[i])) {
-#ifdef PADDLE_WITH_CUDA
-      device_contexts_[i] = new platform::CUDADeviceContext(
-          boost::get<platform::GPUPlace>(places[i]));
-#else
-      PADDLE_THROW(
-          "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
-          "option");
-#endif
-    }
-  }
-}
 
-Executor::~Executor() {
+Executor::Executor(const std::vector<platform::Place>& places) {
-  if (own_) {
+  DeviceContextPool& pool = DeviceContextPool::Get();
-    for (auto& device_context : device_contexts_) {
+  auto borrowed_contexts = pool.Borrow(places);
-      delete device_context;
+  device_contexts_.swap(borrowed_contexts);
-    }
-  }
 }
 
 static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
@@ -132,8 +112,5 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
   }
 }
 
-Executor::Executor(const platform::DeviceContext& device)
-    : device_contexts_({&device}), own_(false) {}
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/executor.h

@@ -14,19 +14,98 @@ limitations under the License. */
 
 #pragma once
 
+#include <map>
+#include <unordered_map>
+
 #include "paddle/framework/op_info.h"
 #include "paddle/framework/program_desc.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
 
+class DeviceContextPool {
+ public:
+  static DeviceContextPool& Get() {
+    PADDLE_ENFORCE_NOT_NULL(pool, "Need to Create DeviceContextPool first!");
+    return *pool;
+  }
+
+  static DeviceContextPool& Create(const std::vector<platform::Place>& places) {
+    if (pool == nullptr) {
+      pool = new DeviceContextPool(places);
+    }
+    return *pool;
+  }
+
+  std::vector<const platform::DeviceContext*> Borrow(
+      const std::vector<platform::Place>& places) {
+    PADDLE_ENFORCE_GT(places.size(), 0);
+    PADDLE_ENFORCE_LE(places.size(), device_contexts_.size());
+    std::vector<const platform::DeviceContext*> borrowed_contexts;
+    for (auto& place : places) {
+      auto range = device_contexts_.equal_range(place);
+      if (range.first == range.second) {
+        PADDLE_THROW(
+            "'Place' is not supported, Please re-compile with WITH_GPU "
+            "option");
+      }
+      // TODO(dzhwinter) : assign the first found device. Will enhanced later.
+      // device load balancer maybe useful here.
+      borrowed_contexts.emplace_back(range.first->second);
+    }
+    return borrowed_contexts;
+  }
+
+  explicit DeviceContextPool(const std::vector<platform::Place>& places) {
+    PADDLE_ENFORCE_GT(places.size(), 0);
+    for (size_t i = 0; i < places.size(); i++) {
+      if (platform::is_cpu_place(places[i])) {
+        device_contexts_.emplace(
+            places[i], new platform::CPUDeviceContext(
+                           boost::get<platform::CPUPlace>(places[i])));
+      } else if (platform::is_gpu_place(places[i])) {
+#ifdef PADDLE_WITH_CUDA
+        device_contexts_.emplace(
+            places[i], new platform::CUDADeviceContext(
+                           boost::get<platform::GPUPlace>(places[i])));
+#else
+        PADDLE_THROW(
+            "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
+            "option");
+#endif
+      }
+    }
+  }
+
+  ~DeviceContextPool() {}
+
+ private:
+  static DeviceContextPool* pool;
+  struct Hash {
+    std::hash<int> hash_;
+    size_t operator()(const platform::Place& place) const {
+      return hash_(place.which());
+    }
+  };
+  std::unordered_multimap<const platform::Place, const platform::DeviceContext*,
+                          Hash>
+      device_contexts_;
+  DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
+};
+
 class Executor {
  public:
+  // TODO(dzhwinter) : Do not rely on this function, it will be removed
+  explicit Executor(const platform::DeviceContext& device)
+      : Executor(std::vector<platform::Place>({device.GetPlace()})) {}
+
+  explicit Executor(const platform::Place& place)
+      : Executor(std::vector<platform::Place>({place})) {}
+
   explicit Executor(const std::vector<platform::Place>& places);
-  explicit Executor(const platform::DeviceContext& devices);
-  ~Executor();
 
   /* @Brief
    * Runtime evaluation of the given ProgramDesc under certain Scope
@@ -39,7 +118,6 @@ class Executor {
 
  private:
   std::vector<const platform::DeviceContext*> device_contexts_;
-  bool own_;
 };
 
 }  // namespace framework

paddle/framework/init.cc

@@ -0,0 +1,80 @@
+/* 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 <algorithm>
+#include <string>
+
+#include "paddle/framework/executor.h"
+#include "paddle/framework/init.h"
+#include "paddle/platform/place.h"
+#include "paddle/string/piece.h"
+
+namespace paddle {
+namespace framework {
+
+std::once_flag gflags_init_flag;
+
+// TODO(qijun) move init gflags to init.cc
+void InitGflags(std::vector<std::string> &argv) {
+  std::call_once(gflags_init_flag, [&]() {
+    int argc = argv.size();
+    char **arr = new char *[argv.size()];
+    std::string line;
+    for (size_t i = 0; i < argv.size(); i++) {
+      arr[i] = &argv[i][0];
+      line += argv[i];
+      line += ' ';
+    }
+    google::ParseCommandLineFlags(&argc, &arr, true);
+    VLOG(1) << "Init commandline: " << line;
+  });
+}
+
+bool InitDevices(const std::vector<std::string> &devices) {
+  // device format
+  // CPU
+  // GPU:1
+  // TODO(dzhwinter) : add device format annotation for users.
+  std::vector<platform::Place> places;
+  for (auto &device : devices) {
+    auto p = string::Piece(device);
+    if (string::Find(p, ':', 0) == string::Piece::npos) {
+      places.emplace_back(platform::CPUPlace());
+    } else if (string::HasPrefix(p, "GPU")) {
+#ifdef PADDLE_WITH_CUDA
+      auto pos = string::RFind(p, ':', string::Piece::npos);
+      auto number = device.substr(pos + 1);
+      places.emplace_back(platform::GPUPlace(std::stoi(number)));
+#else
+      LOG(WARNING)
+          << "'GPU' is not supported, Please re-compile with WITH_GPU option";
+#endif
+    } else {
+      return false;
+    }
+  }
+
+  if (std::find_if(places.begin(), places.end(),
+                   [&](const platform::Place &place) {
+                     return platform::is_cpu_place(place);
+                   }) == places.end()) {
+    places.emplace_back(platform::CPUPlace());
+    LOG(WARNING) << "Not specified any device, use CPU by Default.";
+  }
+  DeviceContextPool::Create(places);
+  return true;
+  return true;
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/init.h

@@ -0,0 +1,28 @@
+/* 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 <mutex>
+
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+
+namespace paddle {
+namespace framework {
+
+void InitGflags(std::vector<std::string> &argv);
+
+bool InitDevices(const std::vector<std::string> &devices);
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/init_test.cc

@@ -0,0 +1,27 @@
+/* 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 "gtest/gtest.h"
+
+#include "paddle/framework/init.h"
+
+TEST(Init, InitDevices) {
+  using paddle::framework::InitDevices;
+  std::vector<std::string> ds1 = {"CPU"};
+  ASSERT_EQ(InitDevices(ds1), true);
+
+#ifdef PADDLE_WITH_CUDA
+  std::vector<std::string> ds2 = {"CPU", "GPU:0", "GPU:1"};
+  ASSERT_EQ(InitDevices(ds2), true);
+#endif
+}

paddle/pybind/CMakeLists.txt

@@ -1,7 +1,7 @@
 if(WITH_PYTHON)
   cc_library(paddle_pybind SHARED
     SRCS pybind.cc exception.cc protobuf.cc
-    DEPS pybind python backward proto_desc paddle_memory executor prune
+    DEPS pybind python backward proto_desc paddle_memory executor prune init
     ${GLOB_OP_LIB})
 endif(WITH_PYTHON)
 

paddle/pybind/pybind.cc

@@ -16,11 +16,11 @@ limitations under the License. */
 
 #include <mutex>  // for call_once
 #include <unordered_map>
-#include "gflags/gflags.h"
 #include "paddle/framework/backward.h"
 #include "paddle/framework/executor.h"
 #include "paddle/framework/feed_fetch_method.h"
 #include "paddle/framework/framework.pb.h"
+#include "paddle/framework/init.h"
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/lod_tensor_array.h"
@@ -51,24 +51,6 @@ static size_t UniqueIntegerGenerator(const std::string &prefix) {
   return generators[prefix].fetch_add(1);
 }
 
-std::once_flag gflags_init_flag;
-
-// TODO(qijun) move init gflags to init.cc
-void InitGflags(std::vector<std::string> &argv) {
-  std::call_once(gflags_init_flag, [&]() {
-    int argc = argv.size();
-    char **arr = new char *[argv.size()];
-    std::string line;
-    for (size_t i = 0; i < argv.size(); i++) {
-      arr[i] = &argv[i][0];
-      line += argv[i];
-      line += ' ';
-    }
-    google::ParseCommandLineFlags(&argc, &arr, true);
-    VLOG(1) << "Init commandline: " << line;
-  });
-}
-
 bool IsCompileGPU() {
 #ifndef PADDLE_WITH_CUDA
   return false;
@@ -438,7 +420,8 @@ All parameter, weight, gradient are variables in Paddle.
       .def("run", &Executor::Run);
 
   m.def("unique_integer", UniqueIntegerGenerator);
-  m.def("init_gflags", InitGflags);
+  m.def("init_gflags", framework::InitGflags);
+  m.def("init_devices", &framework::InitDevices);
 
   m.def("is_compile_gpu", IsCompileGPU);
   m.def("set_feed_variable", framework::SetFeedVariable);

python/paddle/v2/fluid/executor.py

@@ -46,6 +46,13 @@ class Executor(object):
             p.set_place(each)
             act_places.append(p)
 
+        # TODO(dzhwinter) : consider that our fluid tests all written in 
+        # GPUPlace(gpu_id), this will be changed in next PR.
+        if core.is_compile_gpu():
+            core.init_devices(["CPU", "GPU:0"])
+        else:
+            core.init_devices(["CPU"])
+
         self.executor = core.Executor(act_places)
         self.places = places