From 2e1499994294caf936f0182536a725044aed6518 Mon Sep 17 00:00:00 2001 From: Xin Pan Date: Mon, 5 Nov 2018 22:16:23 +0800 Subject: [PATCH] clean1 test=develop --- .../fast_threaded_ssa_graph_executor.cc | 8 +++--- .../details/gather_op_handle_test.cc | 12 ++++----- .../details/multi_devices_graph_check_pass.cc | 8 +++--- .../details/multi_devices_graph_pass.cc | 26 +++++++++---------- .../framework/details/multi_devices_helper.h | 7 +++-- .../fluid/framework/details/op_handle_base.h | 4 ++- .../framework/details/reference_count_pass.cc | 22 +++++++--------- .../framework/details/ssa_graph_executor.cc | 3 +-- .../framework/details/ssa_graph_executor.h | 3 +-- .../details/threaded_ssa_graph_executor.cc | 18 ++++++------- .../details/threaded_ssa_graph_executor.h | 14 +++++----- paddle/fluid/framework/details/var_handle.h | 4 ++- paddle/fluid/framework/ir/node.h | 20 ++++++++++++++ 13 files changed, 84 insertions(+), 65 deletions(-) diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc index 98fc390e72..42849853c8 100644 --- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc @@ -36,9 +36,9 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor( for (auto &op : ops) { int dep = static_cast(op->NotReadyInputSize()); - op_deps_.emplace(op.get(), dep); + op_deps_.emplace(op, dep); if (dep == 0) { - bootstrap_ops_.emplace_back(op.get()); + bootstrap_ops_.emplace_back(op); } } @@ -54,13 +54,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( paddle::framework::FeedFetchList fetches; fetches.resize(fetch_tensors.size()); std::unordered_map> fetched_vars; - std::vector> fetch_ops; + std::vector fetch_ops; for (auto &fetch_var_name : fetch_tensors) { for (auto &var_map : graph_->Get("vars")) { auto it = var_map.find(fetch_var_name); if (it != var_map.end()) { - fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); + fetched_vars[fetch_var_name].push_back(*it->second.rbegin()); } } } diff --git a/paddle/fluid/framework/details/gather_op_handle_test.cc b/paddle/fluid/framework/details/gather_op_handle_test.cc index ed67e88ff6..c83804e262 100644 --- a/paddle/fluid/framework/details/gather_op_handle_test.cc +++ b/paddle/fluid/framework/details/gather_op_handle_test.cc @@ -31,8 +31,8 @@ struct TestGatherOpHandle { std::vector local_scopes_; std::vector param_scopes_; Scope g_scope_; - std::unique_ptr op_handle_; - std::vector> vars_; + OpHandleBase* op_handle_; + std::vector vars_; std::vector gpu_list_; void WaitAll() { @@ -84,8 +84,8 @@ struct TestGatherOpHandle { nodes.emplace_back( ir::CreateNodeForTest("node", ir::Node::Type::kOperation).release()); - op_handle_.reset( - new GatherOpHandle(nodes.back().get(), local_scopes_, gpu_list_)); + op_handle_ = + new GatherOpHandle(nodes.back().get(), local_scopes_, gpu_list_); // add input for (size_t j = 0; j < gpu_list_.size(); ++j) { op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); @@ -102,7 +102,7 @@ struct TestGatherOpHandle { ir::CreateNodeForTest("node2", ir::Node::Type::kVariable).release()); vars_.emplace_back(new DummyVarHandle(nodes.back().get())); DummyVarHandle* in_dummy_var_handle = - static_cast(vars_.back().get()); + static_cast(vars_.back()); in_dummy_var_handle->ClearGeneratedOp(); op_handle_->AddInput(in_dummy_var_handle); @@ -119,7 +119,7 @@ struct TestGatherOpHandle { ir::CreateNodeForTest("node4", ir::Node::Type::kVariable).release()); vars_.emplace_back(new DummyVarHandle(nodes.back().get())); DummyVarHandle* dummy_var_handle = - static_cast(vars_.back().get()); + static_cast(vars_.back()); op_handle_->AddOutput(dummy_var_handle); } diff --git a/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc index fc99552112..5bfafa8291 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc @@ -36,20 +36,20 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const { for (auto &var_map : graph->Get(kGraphVars)) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { - insert_pending_var(version_pair.get()); + insert_pending_var(version_pair); } } } for (auto &var : graph->Get(kGraphDepVars)) { - insert_pending_var(var.get()); + insert_pending_var(var); } for (auto &op : graph->Get(kGraphOps)) { if (op->Inputs().empty()) { - ready_ops.insert(op.get()); + ready_ops.insert(op); } else { - pending_ops.insert({op.get(), op.get()->NoDupInputSize()}); + pending_ops.insert({op, op->NoDupInputSize()}); } } diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc index 2b75f46039..e072e09ece 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -93,7 +93,7 @@ VarHandle *CreateOrGetLatestVarHandle(ir::Graph *graph, ir::Node *node, } var_holder.emplace_back(var); } else { - var = var_holder.rbegin()->get(); + var = *var_holder.rbegin(); } return var; } @@ -155,7 +155,7 @@ void MultiDevSSAGraphBuilder::CreateOpHandleIOs(ir::Graph *result, ir::Node *node, size_t place_id) const { auto p = places_[place_id]; - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); op_handle->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p)); @@ -498,7 +498,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result, result->Get(kGraphOps).emplace_back(op_handle); auto *in = - result->Get(kGraphVars).at(src_dev_id).at(p_name).back().get(); + result->Get(kGraphVars).at(src_dev_id).at(p_name).back(); op_handle->AddInput(in); for (size_t i = 0; i < places_.size(); ++i) { @@ -535,7 +535,7 @@ void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp( for (size_t dev_id = 0; dev_id < bcast_varnames.size(); ++dev_id) { for (auto &p_name : bcast_varnames[dev_id]) { auto *in = - result->Get(kGraphVars).at(dev_id).at(p_name).back().get(); + result->Get(kGraphVars).at(dev_id).at(p_name).back(); op_handle->AddInput(in); for (size_t out_dev_id = 0; out_dev_id < places_.size(); ++out_dev_id) { auto &p = places_[out_dev_id]; @@ -571,7 +571,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result, result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), local_scopes_, places_)); #endif - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (size_t i = 0; i < places_.size(); ++i) { auto &p = places_[i]; @@ -579,7 +579,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result, auto &vars = result->Get(kGraphVars)[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); - op_handle->AddInput(prev_grad.get()); + op_handle->AddInput(prev_grad); auto var = new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable), @@ -600,14 +600,14 @@ void MultiDevSSAGraphBuilder::InsertDataBalanceOp( result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation), local_scopes_, places_)); #endif - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (size_t i = 0; i < places_.size(); ++i) { auto &p = places_[i]; SetCommunicationContext(op_handle, p); for (const std::string &d_name : datas) { auto &vars = result->Get(kGraphVars)[i][d_name]; PADDLE_ENFORCE(!vars.empty()); - op_handle->AddInput(vars.back().get()); + op_handle->AddInput(vars.back()); auto var = new VarHandle( result->CreateEmptyNode(d_name, ir::Node::Type::kVariable), vars.size(), i, d_name, p); @@ -691,7 +691,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result, result->CreateEmptyNode("reduce", ir::Node::Type::kOperation), local_scopes_, places_)); #endif - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (size_t i = 0; i < places_.size(); ++i) { auto &p = places_[i]; @@ -699,7 +699,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result, auto &vars = result->Get(kGraphVars)[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); - op_handle->AddInput(prev_grad.get()); + op_handle->AddInput(prev_grad); } auto &vars = result->Get(kGraphVars)[dst_dev_id][og]; auto var = @@ -760,14 +760,14 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp( } void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) { - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (ir::Node *input : node->inputs) { VarHandle *var = nullptr; for (int place_offset = 0; place_offset < num_places; ++place_offset) { auto &var_holders = result->Get(kGraphVars)[place_offset]; auto &var_holder = var_holders[input->Name()]; if (!var_holder.empty()) { - var = var_holder.rbegin()->get(); + var = *var_holder.rbegin(); op_handle->AddInput(var); } } @@ -840,7 +840,7 @@ int MultiDevSSAGraphBuilder::CreateRPCOp( // send_barrier, recv, fetch_barrier's inputs are deps var, get them from // all places auto p = places_[op_dev_id]; - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); op_handle->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p)); diff --git a/paddle/fluid/framework/details/multi_devices_helper.h b/paddle/fluid/framework/details/multi_devices_helper.h index 0a31735dd6..bed2fdb864 100644 --- a/paddle/fluid/framework/details/multi_devices_helper.h +++ b/paddle/fluid/framework/details/multi_devices_helper.h @@ -36,18 +36,17 @@ namespace details { // map from variable name to variables. The variables, who have the same name, // will have a differsent version. The offset in the // `std::vector>` is the version of varaibles. -typedef std::vector< - std::unordered_map>>> +typedef std::vector>> GraphVars; const char kGraphVars[] = "vars"; // aux variables to represent dependency. Useful to resolve data hazard. -typedef std::unordered_set> GraphDepVars; +typedef std::unordered_set GraphDepVars; const char kGraphDepVars[] = "dep_vars"; // all operators. NOTE that even we use a vector here, the operators is // unordered. -typedef std::vector> GraphOps; +typedef std::vector GraphOps; const char kGraphOps[] = "ops"; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index d09b94a3fd..0c608e276e 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -31,7 +31,9 @@ constexpr char kLocalExecScopeName[] = "@LCOAL_SCOPE@"; // It's responsible for populating necessary fields of ir::Node. class OpHandleBase { public: - explicit OpHandleBase(ir::Node *node) : node_(node) {} + explicit OpHandleBase(ir::Node *node) : node_(node) { + node_->WrappedBy(this); + } virtual ~OpHandleBase(); diff --git a/paddle/fluid/framework/details/reference_count_pass.cc b/paddle/fluid/framework/details/reference_count_pass.cc index 0b994ced7f..19b943cb9c 100644 --- a/paddle/fluid/framework/details/reference_count_pass.cc +++ b/paddle/fluid/framework/details/reference_count_pass.cc @@ -71,14 +71,13 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( // Step 2: Find all variables in non-computation ops which refers to variables // in computation ops std::unordered_set names; - std::unordered_map> + std::unordered_map compute_ref_cnt_map; auto get_ref_cnts_from_compute_op = [&]( - const std::unique_ptr &op, - const std::vector &vars) { + OpHandleBase *op, const std::vector &vars) { std::vector var_names_in_op; - auto *compute_op = dynamic_cast(op.get()); + auto *compute_op = dynamic_cast(op); if (compute_op == nullptr || !platform::is_gpu_place(compute_op->GetPlace())) return var_names_in_op; @@ -121,9 +120,8 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( }; auto update_ref_cnts_from_non_compute_op = [&]( - const std::unique_ptr &op, - const std::vector &vars) { - if (dynamic_cast(op.get()) != nullptr) return; + OpHandleBase *op, const std::vector &vars) { + if (dynamic_cast(op) != nullptr) return; for (VarHandleBase *var_handle_base : vars) { auto *var_handle = dynamic_cast(var_handle_base); if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue; @@ -151,7 +149,7 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( ref_cnt_node, next_compute_op->GetScope(), place, {var_name}, gcs[place.device].get(), cur_ref_cnts[place.device].get()); AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get()); - compute_ref_cnt_map[next_compute_op].reset(ref_cnt_handle); + compute_ref_cnt_map[next_compute_op] = ref_cnt_handle; } } } @@ -165,7 +163,7 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( if (in_var_names.empty() && out_var_names.empty()) continue; in_var_names.insert(in_var_names.end(), out_var_names.begin(), out_var_names.end()); - auto *compute_op = dynamic_cast(op.get()); + auto *compute_op = dynamic_cast(op); auto place = boost::get(compute_op->GetPlace()); ir::Node *ref_cnt_node = graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation); @@ -173,7 +171,7 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( ref_cnt_node, compute_op->GetScope(), place, in_var_names, gcs[place.device].get(), cur_ref_cnts[place.device].get()); AddDependencyBetween(compute_op, ref_cnt_handle, graph.get()); - compute_ref_cnt_map[compute_op].reset(ref_cnt_handle); + compute_ref_cnt_map[compute_op] = ref_cnt_handle; } for (auto &op : all_ops) { @@ -181,11 +179,11 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( update_ref_cnts_from_non_compute_op(op, op->Outputs()); } - std::vector> new_all_ops; + std::vector new_all_ops; new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size()); for (auto &op : all_ops) { new_all_ops.emplace_back(std::move(op)); - auto it = compute_ref_cnt_map.find(new_all_ops.back().get()); + auto it = compute_ref_cnt_map.find(new_all_ops.back()); if (it != compute_ref_cnt_map.end()) { // Add LeafNode to ReferenceCountOpHandle auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar()); diff --git a/paddle/fluid/framework/details/ssa_graph_executor.cc b/paddle/fluid/framework/details/ssa_graph_executor.cc index 780da5478f..d283a34ba9 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/ssa_graph_executor.cc @@ -19,8 +19,7 @@ namespace framework { namespace details { SSAGraphExecutor::~SSAGraphExecutor() {} -void ClearFetchOp(ir::Graph* graph, - std::vector>* fetch_ops) { +void ClearFetchOp(ir::Graph* graph, std::vector* fetch_ops) { if (fetch_ops->empty()) return; for (auto& op : *fetch_ops) { diff --git a/paddle/fluid/framework/details/ssa_graph_executor.h b/paddle/fluid/framework/details/ssa_graph_executor.h index d5cf7737d5..860eaa25b5 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.h +++ b/paddle/fluid/framework/details/ssa_graph_executor.h @@ -38,8 +38,7 @@ class SSAGraphExecutor { virtual FeedFetchList Run(const std::vector& fetch_tensors) = 0; }; -void ClearFetchOp(ir::Graph* graph, - std::vector>* fetch_ops); +void ClearFetchOp(ir::Graph* graph, std::vector* fetch_ops); } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index dc63effd1b..781116ccba 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -51,25 +51,25 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( for (auto &var_map : graph_->Get(details::kGraphVars)) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { - InsertPendingVar(&pending_vars, ready_vars.get(), version_pair.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), version_pair); } } } for (auto &var : graph_->Get(details::kGraphDepVars)) { - InsertPendingVar(&pending_vars, ready_vars.get(), var.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), var); } for (auto &op : graph_->Get(details::kGraphOps)) { if (op->Inputs().empty()) { // Special case, Op has no input. - ready_ops.insert(op.get()); + ready_ops.insert(op); } else { - InsertPendingOp(&pending_ops, op.get()); + InsertPendingOp(&pending_ops, op); } } // Step 2. Insert FetchOps - std::vector> fetch_ops; - std::unordered_set> fetch_dependencies; + std::vector fetch_ops; + std::unordered_set fetch_dependencies; FeedFetchList fetch_data(fetch_tensors.size()); InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops, @@ -140,8 +140,8 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( void ThreadedSSAGraphExecutor::InsertFetchOps( const std::vector &fetch_tensors, - std::vector> *fetch_ops, - std::unordered_set> *fetch_dependencies, + std::vector *fetch_ops, + std::unordered_set *fetch_dependencies, std::unordered_map *pending_ops, std::unordered_set *pending_vars, BlockingQueue *ready_vars, FeedFetchList *fetch_data) { @@ -151,7 +151,7 @@ void ThreadedSSAGraphExecutor::InsertFetchOps( for (auto &var_map : graph_->Get(details::kGraphVars)) { auto it = var_map.find(fetch_var_name); if (it != var_map.end()) { - fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); + fetched_vars[fetch_var_name].push_back(*it->second.rbegin()); } } } diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index dbb0b498d9..6d9dd68927 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -70,13 +70,13 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { BlockingQueue *ready_vars, VarHandleBase *var) const; - void InsertFetchOps( - const std::vector &fetch_tensors, - std::vector> *fetch_ops, - std::unordered_set> *fetch_dependencies, - std::unordered_map *pending_ops, - std::unordered_set *pending_vars, - BlockingQueue *ready_vars, FeedFetchList *fetch_data); + void InsertFetchOps(const std::vector &fetch_tensors, + std::vector *fetch_ops, + std::unordered_set *fetch_dependencies, + std::unordered_map *pending_ops, + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, + FeedFetchList *fetch_data); private: ExecutionStrategy strategy_; diff --git a/paddle/fluid/framework/details/var_handle.h b/paddle/fluid/framework/details/var_handle.h index a1f458c660..bc8d99cf73 100644 --- a/paddle/fluid/framework/details/var_handle.h +++ b/paddle/fluid/framework/details/var_handle.h @@ -35,7 +35,9 @@ class OpHandleBase; // A variable can only be generated by a single operator. i.e. // This is a single assignment graph. struct VarHandleBase { - explicit VarHandleBase(ir::Node* node) : node_(node) {} + explicit VarHandleBase(ir::Node* node) : node_(node) { + node_->WrappedBy(this); + } virtual ~VarHandleBase(); diff --git a/paddle/fluid/framework/ir/node.h b/paddle/fluid/framework/ir/node.h index d6d42f5e92..e41e032d14 100644 --- a/paddle/fluid/framework/ir/node.h +++ b/paddle/fluid/framework/ir/node.h @@ -27,6 +27,8 @@ namespace ir { // Node should normally created by Graph::CreateXXXNode(). class Node { public: + virtual ~Node() {} + enum class Type { kOperation, kVariable }; static constexpr char kControlDepVarName[] = "__control_var"; @@ -44,6 +46,20 @@ class Node { return op_desc_.get(); } + template + void WrappedBy(T* wrapper) { + if (!wrapper_.empty()) { + wrapper_deleter_(); + } + wrapper_ = wrapper; + wrapper_deleter_ = [wrapper]() { delete wrapper; }; + } + + template + T& Wrapper() { + return *boost::any_cast(wrapper_); + } + // Please don't use this API! int id() const { return id_; } @@ -95,6 +111,10 @@ class Node { static int count_; // Please don't use this API or make this public. static void ResetId() { count_ = 0; } + + boost::any wrapper_; + std::function wrapper_deleter_; + DISABLE_COPY_AND_ASSIGN(Node); }; -- GitLab