Add kids exists detection in Scope

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -22,8 +22,7 @@ namespace framework {
 namespace details {
 
 FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
-    const ExecutionStrategy &strategy,
-    const std::vector<std::shared_ptr<Scope>> &local_scopes,
+    const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,
     std::unique_ptr<ir::Graph> &&graph)
     : strategy_(strategy),

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -29,9 +29,8 @@ namespace details {
 class OpHandleBase;
 class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
  public:
-  FastThreadedSSAGraphExecutor(
-      const ExecutionStrategy &strategy,
-      const std::vector<std::shared_ptr<Scope>> &local_scopes,
+  FastThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
+                               const std::vector<Scope *> &local_scopes,
                                const std::vector<platform::Place> &places,
                                std::unique_ptr<ir::Graph> &&graph);
   FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
@@ -39,7 +38,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
 
  private:
   ExecutionStrategy strategy_;
-  std::vector<std::shared_ptr<Scope>> local_scopes_;
+  std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   std::unique_ptr<ir::Graph> graph_;
 

paddle/fluid/framework/details/fetch_op_handle.cc

@@ -22,7 +22,7 @@ namespace framework {
 namespace details {
 
 FetchOpHandle::FetchOpHandle(ir::Node *node, FeedFetchList *data, size_t offset,
-                             std::vector<std::shared_ptr<Scope>> *local_scopes)
+                             std::vector<Scope *> *local_scopes)
     : OpHandleBase(node),
       data_(data),
       offset_(offset),

paddle/fluid/framework/details/fetch_op_handle.h

@@ -29,7 +29,7 @@ namespace details {
 struct FetchOpHandle : public OpHandleBase {
  public:
   FetchOpHandle(ir::Node *node, FeedFetchList *data, size_t offset,
-                std::vector<std::shared_ptr<Scope>> *local_scopes);
+                std::vector<Scope *> *local_scopes);
 
   ~FetchOpHandle();
 
@@ -47,7 +47,7 @@ struct FetchOpHandle : public OpHandleBase {
  private:
   FeedFetchList *data_;
   size_t offset_;
-  std::vector<std::shared_ptr<Scope>> *local_scopes_;
+  std::vector<Scope *> *local_scopes_;
   std::vector<LoDTensor> tensors_;
 };
 

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc

@@ -23,8 +23,7 @@ namespace paddle {
 namespace framework {
 namespace details {
 ScopeBufferedSSAGraphExecutor::ScopeBufferedSSAGraphExecutor(
-    ExecutionStrategy strategy,
-    std::vector<std::shared_ptr<Scope>> local_scopes,
+    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)),

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h

@@ -37,8 +37,7 @@ struct VariableInfo {
 class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
  public:
   ScopeBufferedSSAGraphExecutor(
-      ExecutionStrategy strategy,
-      std::vector<std::shared_ptr<Scope>> local_scopes,
+      ExecutionStrategy strategy, std::vector<Scope*> local_scopes,
       std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
       std::unique_ptr<SSAGraphExecutor>&& underlying_executor);
 
@@ -53,7 +52,7 @@ class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
 
   ExecutionStrategy strategy_;
   std::unique_ptr<SSAGraphExecutor> underlying_executor_;
-  std::vector<std::shared_ptr<Scope>> local_scopes_;
+  std::vector<Scope*> local_scopes_;
   std::vector<VariableInfo> var_infos_;
   std::vector<platform::Place> places_;
 };

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -21,8 +21,7 @@ namespace paddle {
 namespace framework {
 namespace details {
 ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
-    const ExecutionStrategy &strategy,
-    const std::vector<std::shared_ptr<Scope>> &local_scopes,
+    const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,
     std::unique_ptr<ir::Graph> &&graph)
     : graph_(std::move(graph)),

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -38,9 +38,8 @@ namespace details {
 
 class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
  public:
-  ThreadedSSAGraphExecutor(
-      const ExecutionStrategy &strategy,
-      const std::vector<std::shared_ptr<Scope>> &local_scopes,
+  ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
+                           const std::vector<Scope *> &local_scopes,
                            const std::vector<platform::Place> &places,
                            std::unique_ptr<ir::Graph> &&graph);
 
@@ -58,7 +57,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
  private:
   std::unique_ptr<ir::Graph> graph_;
   std::unique_ptr<::ThreadPool> pool_;
-  std::vector<std::shared_ptr<Scope>> local_scopes_;
+  std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   platform::DeviceContextPool fetch_ctxs_;
   ExceptionHolder exception_holder_;

paddle/fluid/framework/parallel_executor.cc

@@ -39,8 +39,7 @@ std::unique_ptr<ir::Graph> ApplyParallelExecutorPass(
     const ProgramDesc &main_program, const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &param_names,
-    const std::vector<std::shared_ptr<Scope>> &local_scopes,
-    const bool use_cuda,
+    const std::vector<Scope *> &local_scopes, const bool use_cuda,
 #ifdef PADDLE_WITH_CUDA
     const BuildStrategy &strategy, platform::NCCLContextMap *nccl_ctxs) {
 #else
@@ -67,8 +66,8 @@ std::unique_ptr<ir::Graph> ApplyParallelExecutorPass(
                                                      &loss_var_name);
   multi_devices_pass->SetNotOwned<const std::unordered_set<std::string>>(
       "params", &param_names);
-  multi_devices_pass->SetNotOwned<const std::vector<std::shared_ptr<Scope>>>(
-      "local_scopes", &local_scopes);
+  multi_devices_pass->SetNotOwned<const std::vector<Scope *>>("local_scopes",
+                                                              &local_scopes);
   multi_devices_pass->SetNotOwned<const BuildStrategy>("strategy", &strategy);
 
 #ifdef PADDLE_WITH_CUDA
@@ -101,8 +100,8 @@ class ParallelExecutorPrivate {
       : places_(places) {}
 
   std::vector<platform::Place> places_;
-  std::vector<std::shared_ptr<Scope>> local_scopes_;
-  std::shared_ptr<Scope> global_scope_;
+  std::vector<Scope *> local_scopes_;
+  Scope *global_scope_;
   std::unique_ptr<details::SSAGraphExecutor> executor_;
 
 #ifdef PADDLE_WITH_CUDA
@@ -113,7 +112,7 @@ class ParallelExecutorPrivate {
   bool use_all_reduce_;
 };
 
-std::vector<std::shared_ptr<Scope>> &ParallelExecutor::GetLocalScopes() {
+std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
   return member_->local_scopes_;
 }
 
@@ -122,8 +121,7 @@ ParallelExecutor::ParallelExecutor(
     const std::unordered_set<std::string> &params,
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
-    const std::shared_ptr<Scope> &scope,
-    const std::vector<std::shared_ptr<Scope>> &local_scopes,
+    Scope *scope, const std::vector<Scope *> &local_scopes,
     const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy,
     size_t num_trainers, size_t trainer_id)
     : member_(new ParallelExecutorPrivate(places)) {
@@ -144,13 +142,13 @@ ParallelExecutor::ParallelExecutor(
     member_->own_local_scope_ = true;
     member_->local_scopes_.emplace_back(member_->global_scope_);
     for (size_t i = 1; i < member_->places_.size(); ++i) {
-      member_->local_scopes_.emplace_back(scope->NewSharedScope());
+      member_->local_scopes_.emplace_back(&scope->NewScope());
     }
   } else {
     member_->own_local_scope_ = false;
     PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
     for (size_t i = 0; i < member_->places_.size(); ++i) {
-      member_->local_scopes_.emplace_back(local_scopes[i]->NewSharedScope());
+      member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
     }
   }
 
@@ -323,7 +321,7 @@ void ParallelExecutor::FeedTensorsIntoLocalScopes(
 
   for (size_t i = 0; i < tensors.size(); ++i) {
     auto &map = tensors[i];
-    auto &scope = member_->local_scopes_[i];
+    auto *scope = member_->local_scopes_[i];
     for (auto &pair : map) {
       auto *trg = scope->Var(pair.first)->GetMutable<LoDTensor>();
       trg->ShareDataWith(pair.second);
@@ -353,15 +351,11 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
 
 ParallelExecutor::~ParallelExecutor() {
   if (member_->own_local_scope_) {
-    std::vector<Scope *> local_scopes_ptrs;
-    local_scopes_ptrs.reserve(member_->local_scopes_.size());
     for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
-      local_scopes_ptrs.emplace_back(member_->local_scopes_[i].get());
-      member_->local_scopes_[i].reset();
+      Scope *local_scope = member_->local_scopes_[i];
+      if (member_->global_scope_->HasKid(local_scope)) {
+        member_->global_scope_->DeleteScope(local_scope);
       }
-
-    for (size_t i = 0; i != local_scopes_ptrs.size(); ++i) {
-      member_->global_scope_->DeleteScope(local_scopes_ptrs[i]);
     }
   }
 }

paddle/fluid/framework/parallel_executor.h

@@ -39,20 +39,19 @@ class ParallelExecutor {
   DISABLE_COPY_AND_ASSIGN(ParallelExecutor);
 
  public:
-  explicit ParallelExecutor(
-      const std::vector<platform::Place> &places,
+  explicit ParallelExecutor(const std::vector<platform::Place> &places,
                             const std::unordered_set<std::string> &params,
                             const std::unordered_set<std::string> &bcast_vars,
-      const ProgramDesc &main_program, const std::string &loss_var_name,
-      const std::shared_ptr<Scope> &scope,
-      const std::vector<std::shared_ptr<Scope>> &local_scopes,
+                            const ProgramDesc &main_program,
+                            const std::string &loss_var_name, Scope *scope,
+                            const std::vector<Scope *> &local_scopes,
                             const ExecutionStrategy &exec_strategy,
-      const BuildStrategy &build_strategy, size_t num_trainers = 1,
-      size_t trainer_id = 0);
+                            const BuildStrategy &build_strategy,
+                            size_t num_trainers = 1, size_t trainer_id = 0);
 
   ~ParallelExecutor();
 
-  std::vector<std::shared_ptr<Scope>> &GetLocalScopes();
+  std::vector<Scope *> &GetLocalScopes();
 
   /**
    * Feed tensors to local scopes. The size of tensors should be equal to the

paddle/fluid/framework/scope.cc

@@ -38,8 +38,8 @@ Scope::~Scope() { DropKids(); }
 
 Scope& Scope::NewScope() const {
   std::unique_lock<std::mutex> lock(mutex_);
-  kids_.push_back(std::shared_ptr<Scope>(new Scope(this)));
-  return kids_.back().get();
+  kids_.push_back(new Scope(this));
+  return *kids_.back();
 }
 
 Variable* Scope::Var(const std::string& name) {
@@ -68,9 +68,16 @@ const Scope* Scope::FindScope(const Variable* var) const {
 
 void Scope::DropKids() {
   std::unique_lock<std::mutex> lock(mutex_);
+  for (Scope* s : kids_) delete s;
   kids_.clear();
 }
 
+bool Scope::HasKid(const Scope* scope) const {
+  std::unique_lock<std::mutex> lock(mutex_);
+  auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
+  return it != this->kids_.end();
+}
+
 std::vector<std::string> Scope::LocalVarNames() const {
   std::unique_lock<std::mutex> lock(mutex_);
   std::vector<std::string> known_vars;
@@ -83,12 +90,8 @@ std::vector<std::string> Scope::LocalVarNames() const {
 
 void Scope::DeleteScope(Scope* scope) const {
   std::unique_lock<std::mutex> lock(mutex_);
-  auto it = std::find_if(this->kids_.begin(), this->kids_.end(),
-                         [&scope](const std::shared_ptr<Scope>& kid) {
-                           return kid.get() == scope;
-                         });
+  auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
-  it->reset();
   this->kids_.erase(it);
   // When making memory benchmark on Fluid, we have to delete scope sync.
   if (FLAGS_benchmark || FLAGS_eager_delete_scope) {

paddle/fluid/framework/scope.h

@@ -71,6 +71,9 @@ class Scope {
   /// Drop all kids scopes belonged to this scope.
   void DropKids();
 
+  /// Find if a scope exists in the kid scopes
+  bool HasKid(const Scope* scope) const;
+
   // enumerate all the variables current contains.
   std::vector<std::string> LocalVarNames() const;
 
@@ -105,7 +108,7 @@ class Scope {
   Variable* FindVarLocally(const std::string& name) const;
 
   // Scope in `kids_` are owned by this class.
-  mutable std::list<std::shared_ptr<Scope>> kids_;
+  mutable std::list<Scope*> kids_;
   Scope const* parent_{nullptr};
 
   DISABLE_COPY_AND_ASSIGN(Scope);

python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py

@@ -178,7 +178,4 @@ if __name__ == '__main__':
         for parallel in (False, True):
             if use_cuda and not core.is_compiled_with_cuda():
                 continue
-            # TODO(minqiyang): remove this line after fixing the deletion
-            # order problem of Scope in ParallelExecutor in manylinux
-            if six.PY2:
             main(use_cuda=use_cuda, parallel=parallel)

python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py

@@ -152,7 +152,4 @@ if __name__ == '__main__':
         for parallel in (False, True):
             if use_cuda and not core.is_compiled_with_cuda():
                 continue
-            # TODO(minqiyang): remove this line after fixing the deletion
-            # order problem of Scope in ParallelExecutor in manylinux
-            if six.PY2:
             main(use_cuda=use_cuda, parallel=parallel)

python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py

@@ -155,7 +155,4 @@ if __name__ == '__main__':
         for parallel in (False, True):
             if use_cuda and not core.is_compiled_with_cuda():
                 continue
-            # TODO(minqiyang): remove this line after fixing the deletion
-            # order problem of Scope in ParallelExecutor in manylinux
-            if six.PY2:
             main(use_cuda=use_cuda, parallel=parallel)

python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py

@@ -137,7 +137,4 @@ if __name__ == '__main__':
         for parallel in (False, True):
             if use_cuda and not core.is_compiled_with_cuda():
                 continue
-            # TODO(minqiyang): remove this line after fixing the deletion
-            # order problem of Scope in ParallelExecutor in manylinux
-            if six.PY2:
             main(use_cuda=use_cuda, parallel=parallel)