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 6e22fedf1c39428528c00cce4c9a4460dfb95cb3..98fc390e72fab3701538fd6f974460fa5114fdb0 100644 --- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc @@ -92,13 +92,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( size_t num_complete = 0; remaining_ = 0; - BlockingQueue complete_q; + auto complete_q = std::make_shared>(); for (auto op : bootstrap_ops_) { - RunOpAsync(op_deps.get(), op, &complete_q); + RunOpAsync(op_deps.get(), op, complete_q); } while (num_complete != op_deps->size()) { - size_t num_comp = complete_q.Pop(); + size_t num_comp = complete_q->Pop(); if (num_comp == -1UL) { int remaining = 0; while (true) { @@ -107,7 +107,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( break; } for (int i = 0; i < remaining; ++i) { - complete_q.Pop(); + complete_q->Pop(); } } exception_.ReThrow(); @@ -120,7 +120,8 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( } void FastThreadedSSAGraphExecutor::RunOpAsync( std::unordered_map> *op_deps, - OpHandleBase *op, BlockingQueue *complete_q) { + OpHandleBase *op, + const std::shared_ptr> &complete_q) { ++remaining_; this->pool_.enqueue([=] { OpHandleBase *op_to_run = op; @@ -144,7 +145,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync( if (op_to_run == nullptr) { op_to_run = pending_op; } else { - this->RunOpAsync(op_deps, pending_op, complete_q); + RunOpAsync(op_deps, pending_op, complete_q); } } } @@ -156,8 +157,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync( } void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() { atomic_op_deps_ = pool_.enqueue([&] { - std::unordered_map> *op_deps = - new std::unordered_map>; + auto *op_deps = new std::unordered_map>; for (auto &pair : op_deps_) { (*op_deps)[pair.first] = pair.second; } diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h index dad3a231cba6402f57ba654a9ac5fb520b9c8f04..8b8382447105c8caa36963214684d6ee9fa15200 100644 --- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h @@ -50,7 +50,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor { std::atomic remaining_; void RunOpAsync(std::unordered_map> *op_deps, - OpHandleBase *op, BlockingQueue *complete_q); + OpHandleBase *op, + const std::shared_ptr> &complete_q); void PrepareAtomicOpDeps(); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 31beef3ae829d72570ee7c879dac71ed600cd216..dc63effd1b7c8fe5bb3fc91058eb855e552d3926 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -39,7 +39,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr)); std::unordered_map pending_ops; std::unordered_set pending_vars; - BlockingQueue ready_vars; + auto ready_vars = std::make_shared>(); std::unordered_set ready_ops; // For ops (e.g. nccl_all_reduce) that need to coordinate multiple // streams from multiple GPUs, it's faster to buffer them and schedule @@ -51,12 +51,12 @@ 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, version_pair.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), version_pair.get()); } } } for (auto &var : graph_->Get(details::kGraphDepVars)) { - InsertPendingVar(&pending_vars, &ready_vars, var.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), var.get()); } for (auto &op : graph_->Get(details::kGraphOps)) { @@ -73,12 +73,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( FeedFetchList fetch_data(fetch_tensors.size()); InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops, - &pending_vars, &ready_vars, &fetch_data); + &pending_vars, ready_vars.get(), &fetch_data); auto run_all_ops = [&](std::unordered_set &set) { for (auto *op : set) { running_ops_++; - RunOp(&ready_vars, op); + RunOp(ready_vars, op); } set.clear(); }; @@ -87,7 +87,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( run_op_futures_.clear(); exception_holder_.Clear(); event.reset(nullptr); - // Step 3. Execution while (!pending_vars.empty()) { // 1. Run All Ready ops @@ -103,7 +102,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // 2. Find ready variable bool timeout; - auto cur_ready_vars = ready_vars.PopAll(1, &timeout); + auto cur_ready_vars = ready_vars->PopAll(1, &timeout); if (timeout) { if (exception_holder_.IsCaught()) { @@ -133,7 +132,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( } } PADDLE_ENFORCE(ready_ops.empty()); - // Wait FetchOps. ClearFetchOp(graph_.get(), &fetch_ops); @@ -206,7 +204,8 @@ void ThreadedSSAGraphExecutor::InsertPendingVar( } void ThreadedSSAGraphExecutor::RunOp( - BlockingQueue *ready_var_q, details::OpHandleBase *op) { + const std::shared_ptr> &ready_var_q, + details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { if (VLOG_IS_ON(10)) { diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index 512f8a4ca5a9b82a395dde11722b8db44ea5ec27..dbb0b498d995a897b109bd4ef98521b2193276ed 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -51,7 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { ~ThreadedSSAGraphExecutor() {} private: - void RunOp(BlockingQueue *ready_var_q, + void RunOp(const std::shared_ptr> &ready_var_q, details::OpHandleBase *op); private: