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

Author: reyoung

paddle/fluid/framework/CMakeLists.txt

@@ -87,7 +87,7 @@ cc_library(executor SRCS executor.cc DEPS op_registry device_context scope
 framework_proto glog lod_rank_table feed_fetch_method)
 
 
-cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor)
+cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

paddle/fluid/framework/details/CMakeLists.txt

@@ -36,5 +36,6 @@ cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_ha
         device_context broadcast_op_handle)
 cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
         device_context gather_op_handle)
+cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_executor.cc DEPS ssa_graph_executor)
 #cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
 #        device_context reduce_op_handle )

paddle/fluid/framework/details/execution_strategy.h

@@ -22,6 +22,7 @@ struct ExecutionStrategy {
   size_t num_threads_{0};
   bool use_event_{true};
   bool allow_op_delay_{false};
+  size_t num_iteration_per_drop_scope_{100};
 };
 
 }  //  namespace details

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc

@@ -0,0 +1,76 @@
+// 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.
+
+#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/executor.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+ScopeBufferedSSAGraphExecutor::ScopeBufferedSSAGraphExecutor(
+    ExecutionStrategy strategy, std::vector<Scope *> local_scopes,
+    std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
+    std::unique_ptr<SSAGraphExecutor> &&underlying_executor)
+    : strategy_(std::move(strategy)),
+      underlying_executor_(std::move(underlying_executor)),
+      local_scopes_(std::move(local_scopes)),
+      var_infos_(std::move(var_infos)),
+      places_(std::move(places)) {}
+
+FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
+    const std::vector<std::string> &fetch_tensors) {
+  if (drop_scope_counter_ == 0) {
+    // Create local scopes.
+    for (auto it = local_scopes_.rbegin(); it != local_scopes_.rend(); ++it) {
+      auto &scope = *it;
+      Scope &local_scope = scope->NewScope();
+      *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>() =
+          &local_scope;
+
+      for (auto &info : var_infos_) {
+        if (scope->FindVar(info.name_) != nullptr) {
+          continue;
+        }
+
+        if (info.persistable_) {  // Persistable
+          InitializeVariable(scope->Var(info.name_), info.type_);
+        } else {
+          InitializeVariable(local_scope.Var(info.name_), info.type_);
+        }
+      }
+    }
+  }
+
+  auto fetch_data = underlying_executor_->Run(fetch_tensors);
+  drop_scope_counter_ += 1;
+  if (!fetch_tensors.empty() ||
+      drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
+    drop_scope_counter_ = 0;
+    // Wait All computational streams
+    for (auto p : places_) {
+      platform::DeviceContextPool::Instance().Get(p)->Wait();
+    }
+    for (auto &scope : local_scopes_) {
+      auto &local_scope =
+          *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>();
+      scope->DeleteScope(local_scope);
+    }
+  }
+  return fetch_data;
+}
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h

@@ -0,0 +1,53 @@
+// 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 <memory>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/details/execution_strategy.h"
+#include "paddle/fluid/framework/details/ssa_graph_executor.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/platform/place.h"
+namespace paddle {
+namespace framework {
+namespace details {
+
+struct VariableInfo {
+  std::string name_;
+  proto::VarType::Type type_;
+  bool persistable_;
+};
+
+class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
+ public:
+  ScopeBufferedSSAGraphExecutor(
+      ExecutionStrategy strategy, std::vector<Scope*> local_scopes,
+      std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
+      std::unique_ptr<SSAGraphExecutor>&& underlying_executor);
+  FeedFetchList Run(const std::vector<std::string>& fetch_tensors) override;
+
+ private:
+  size_t drop_scope_counter_{0};
+
+  ExecutionStrategy strategy_;
+  std::unique_ptr<SSAGraphExecutor> underlying_executor_;
+  std::vector<Scope*> local_scopes_;
+  std::vector<VariableInfo> var_infos_;
+  std::vector<platform::Place> places_;
+};
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/ssa_graph_executor.cc

@@ -17,10 +17,6 @@
 namespace paddle {
 namespace framework {
 namespace details {
-
-SSAGraphExecutor::SSAGraphExecutor(std::unique_ptr<SSAGraph> &&graph)
-    : graph_(std::move(graph)) {}
-
 SSAGraphExecutor::~SSAGraphExecutor() {}
 
 }  // namespace details

paddle/fluid/framework/details/ssa_graph_executor.h

@@ -28,15 +28,11 @@ class SSAGraphExecutor {
   DISABLE_COPY_AND_ASSIGN(SSAGraphExecutor);
 
  public:
-  // Steal graph inside
-  explicit SSAGraphExecutor(std::unique_ptr<SSAGraph> &&graph);
+  SSAGraphExecutor() {}
 
   virtual ~SSAGraphExecutor();
 
   virtual FeedFetchList Run(const std::vector<std::string> &fetch_tensors) = 0;
-
- protected:
-  std::unique_ptr<SSAGraph> graph_;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -21,7 +21,7 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
     const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,
     std::unique_ptr<SSAGraph> &&graph)
-    : SSAGraphExecutor(std::move(graph)),
+    : graph_(std::move(graph)),
       pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
                                        : nullptr),
       local_scopes_(local_scopes),
@@ -189,7 +189,9 @@ void ThreadedSSAGraphExecutor::RunOp(
     BlockingQueue<VarHandleBase *> *ready_var_q, details::OpHandleBase *op) {
   auto op_run = [ready_var_q, op, this] {
     try {
-      VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
+      if (VLOG_IS_ON(10)) {
+        VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
+      }
       op->Run(strategy_.use_event_);
       VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -51,6 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
              details::OpHandleBase *op);
 
  private:
+  std::unique_ptr<SSAGraph> graph_;
   std::unique_ptr<::ThreadPool> pool_;
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;

paddle/fluid/framework/parallel_executor.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 #endif
 
 #include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
+#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -42,8 +43,6 @@ class ParallelExecutorPrivate {
 #ifdef PADDLE_WITH_CUDA
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
-
-  std::vector<std::tuple<std::string, proto::VarType::Type, bool>> var_types_;
   bool own_local_scope;
 };
 
@@ -92,9 +91,18 @@ ParallelExecutor::ParallelExecutor(
       local_scopes.empty()) {  // Is CUDA
     BCastParamsToGPUs(bcast_vars);
   }
-// Startup Program has been run. All local scopes has correct parameters.
+  // Startup Program has been run. All local scopes has correct parameters.
+
+  // Step 2. Create vars in each scope;
+  std::vector<details::VariableInfo> var_infos;
+  for (auto *var : main_program.Block(0).AllVars()) {
+    var_infos.emplace_back();
+    var_infos.back().name_ = var->Name();
+    var_infos.back().type_ = var->GetType();
+    var_infos.back().persistable_ = var->Persistable();
+  }
 
-// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
+// Step 3. Convert main_program to SSA form and dependency graph. Also, insert
 // ncclOp
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
@@ -105,16 +113,15 @@ ParallelExecutor::ParallelExecutor(
                                            params, member_->local_scopes_,
                                            build_strategy);
 #endif
+
   auto graph = builder.Build(main_program);
 
   member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
       exec_strategy, member_->local_scopes_, places, std::move(graph)));
 
-  // Step 3. Create vars in each scope;
-  for (auto *var : main_program.Block(0).AllVars()) {
-    member_->var_types_.emplace_back(var->Name(), var->GetType(),
-                                     var->Persistable());
-  }
+  member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
+      exec_strategy, member_->local_scopes_, std::move(var_infos),
+      member_->places_, std::move(member_->executor_)));
 }
 
 void ParallelExecutor::BCastParamsToGPUs(
@@ -169,42 +176,9 @@ void ParallelExecutor::BCastParamsToGPUs(
 void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                            const std::string &fetched_var_name) {
   platform::RecordBlock b(0);
-  // Create local scopes.
-  for (auto it = member_->local_scopes_.rbegin();
-       it != member_->local_scopes_.rend(); ++it) {
-    auto &scope = *it;
-    Scope &local_scope = scope->NewScope();
-    *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>() =
-        &local_scope;
-
-    for (auto &name_type_pair : member_->var_types_) {
-      if (scope->FindVar(std::get<0>(name_type_pair)) != nullptr) {
-        continue;
-      }
-
-      if (std::get<2>(name_type_pair)) {  // Persistable
-        InitializeVariable(scope->Var(std::get<0>(name_type_pair)),
-                           std::get<1>(name_type_pair));
-      } else {
-        InitializeVariable(local_scope.Var(std::get<0>(name_type_pair)),
-                           std::get<1>(name_type_pair));
-      }
-    }
-  }
-
   auto fetch_data = member_->executor_->Run(fetch_tensors);
   *member_->global_scope_->Var(fetched_var_name)->GetMutable<FeedFetchList>() =
       fetch_data;
-
-  // Wait All computational streams
-  for (auto p : member_->places_) {
-    platform::DeviceContextPool::Instance().Get(p)->Wait();
-  }
-  for (auto &scope : member_->local_scopes_) {
-    auto &local_scope =
-        *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>();
-    scope->DeleteScope(local_scope);
-  }
 }
 
 void ParallelExecutor::FeedTensorsIntoLocalScopes(