Merge pull request #15716 from Yancey1989/refine_pg

Refine ParallelGraph Execution

paddle/fluid/framework/details/all_reduce_deps_pass.cc

@@ -30,8 +30,6 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-static constexpr char kAllOpDescs[] = "all_op_descs";
-
 VarHandle* GetValidInput(const OpHandleBase* a) {
   for (auto p : a->Inputs()) {
     VarHandle* b = dynamic_cast<VarHandle*>(p);

paddle/fluid/framework/details/build_strategy.cc

@@ -34,9 +34,11 @@ namespace details {
 static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) {
   // Should fix the allreduce op order if scheduling
   // them in multiple threads or processes to avoid hang.
+  // NOTE: ParallelGraph would execute this pass on each graph, so
+  // don't need to append it here.
   return (!strategy.enable_sequential_execution_ &&
-          strategy.num_trainers_ > 1) ||
-         strategy.enable_parallel_graph_;
+          strategy.num_trainers_ > 1) &&
+         !strategy.enable_parallel_graph_;
 }
 
 class ParallelExecutorPassBuilder : public ir::PassBuilder {

paddle/fluid/framework/details/memory_optimize_helper.h

@@ -29,8 +29,6 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-constexpr char kAllOpDescs[] = "all_op_descs";
-
 std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
 
 // NOTE(dzh): A ordered set for node reuse in memory optimize.

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -392,20 +392,32 @@ void MultiDevSSAGraphBuilderBase::CreateComputationalOp(ir::Graph *result,
 
 void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
     ir::Graph *result, const std::string &og) const {
+  OpHandleBase *op_handle = nullptr;
+
+  auto append_allreduce_op = [&](
+      const std::vector<Scope *> &scopes,
+      const std::vector<platform::Place> &places) -> OpHandleBase * {
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
         result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
-      local_scopes_, places_, nccl_ctxs_));
+        scopes, places, nccl_ctxs_));
 #else
     result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
         result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
-      local_scopes_, places_));
+        scopes, places));
 #endif
-  auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
+    return result->Get<GraphOps>(kGraphOps).back();
+  };
+
+  if (!strategy_.enable_parallel_graph_)
+    op_handle = append_allreduce_op(local_scopes_, places_);
 
   for (size_t i = 0; i < places_.size(); ++i) {
-    auto &p = places_[i];
-    SetCommunicationContext(op_handle, p);
+    if (strategy_.enable_parallel_graph_) {
+      op_handle = append_allreduce_op({local_scopes_[i]}, {places_[i]});
+    }
+
+    SetCommunicationContext(op_handle, places_[i]);
     auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
     PADDLE_ENFORCE(!vars.empty());
     auto &prev_grad = vars.back();
@@ -413,7 +425,7 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
 
     auto var =
         new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
-                      vars.size(), i, og, p);
+                      vars.size(), i, og, places_[i]);
     vars.emplace_back(var);
     op_handle->AddOutput(var);
   }

paddle/fluid/framework/details/multi_devices_helper.h

@@ -36,13 +36,14 @@ 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<VarHandle*>` is the version of varaibles.
-typedef std::vector<std::unordered_map<std::string, std::vector<VarHandle*>>>
+typedef std::vector<std::unordered_map<std::string, std::vector<VarHandle *>>>
     GraphVars;
 const char kGraphVars[] = "vars";
 
 // aux variables to represent dependency. Useful to resolve data hazard.
-typedef std::unordered_set<VarHandleBase*> GraphDepVars;
+typedef std::unordered_set<VarHandleBase *> GraphDepVars;
 const char kGraphDepVars[] = "dep_vars";
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/op_handle_base.h

@@ -70,6 +70,9 @@ class OpHandleBase {
     auto it = dev_ctxes_.find(place);
     return it != dev_ctxes_.end() ? it->second : nullptr;
   }
+  const std::map<platform::Place, platform::DeviceContext *> &DeviceContext() {
+    return dev_ctxes_;
+  }
 
   void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) {
     dev_ctxes_[place] = ctx_;

paddle/fluid/framework/details/parallel_ssa_graph_executor.cc

@@ -13,22 +13,92 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
 
+std::vector<std::unique_ptr<ir::Graph>>
+ParallelSSAGraphExecutor::SeparateMultiDevicesGraph(
+    std::unique_ptr<ir::Graph> &&graph) {
+  std::vector<std::unique_ptr<ir::Graph>> graphs;
+  graphs.reserve(places_.size());
+  for (size_t i = 0; i < places_.size(); ++i) {
+    ProgramDesc empty;
+    graphs.emplace_back(std::unique_ptr<ir::Graph>(new ir::Graph(empty)));
+    auto &g = graphs.back();
+    g->Set(kGraphVars, new GraphVars(1UL));
+    g->Set(kGraphDepVars, new GraphDepVars);
+  }
+  auto op_handles = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
+
+  for (auto &op : op_handles) {
+    auto &dev_ctx = op->DeviceContext();
+    auto &p = dev_ctx.begin()->first;
+    int dev_id = boost::get<platform::CUDAPlace>(p).device;
+    auto &dev_dummys = graphs[dev_id]->Get<GraphDepVars>(kGraphDepVars);
+    graphs[dev_id]->AddNode(graph->RemoveNode(op->Node()).release());
+
+    for (auto &var : op->Inputs()) {
+      auto dummy_ptr = dynamic_cast<DummyVarHandle *>(var);
+      if (dummy_ptr) {
+        dev_dummys.insert(var);
+        if (graph->Nodes().count(var->Node()))
+          graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release());
+      }
+    }
+    for (auto &var : op->Outputs()) {
+      auto dummy_ptr = dynamic_cast<DummyVarHandle *>(var);
+      if (dummy_ptr) {
+        dev_dummys.insert(var);
+        if (graph->Nodes().count(var->Node()))
+          graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release());
+      }
+    }
+  }
+
+  for (size_t dev_id = 0; dev_id < places_.size(); ++dev_id) {
+    auto &dev_vars = graphs[dev_id]->Get<GraphVars>(kGraphVars)[0];
+    auto &origin_vars = graph->Get<GraphVars>(kGraphVars)[dev_id];
+    for (auto &name_pair : origin_vars) {
+      dev_vars.emplace(name_pair.first, name_pair.second);
+      for (auto &version_pair : name_pair.second) {
+        if (graph->Nodes().count(version_pair->Node())) {
+          graphs[dev_id]->AddNode(
+              graph->RemoveNode(version_pair->Node()).release());
+        }
+      }
+    }
+  }
+
+  return graphs;
+}
+
 ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
     const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,
-    std::vector<std::unique_ptr<ir::Graph>> &&graphs)
+    const framework::ProgramDesc &main_prog, std::unique_ptr<ir::Graph> &&graph)
     : strategy_(std::move(strategy)),
       local_scopes_(std::move(local_scopes)),
       pool_(places.size() >= 2 ? new ::ThreadPool(places.size()) : nullptr),
       places_(std::move(places)),
-      graphs_(std::move(graphs)) {
+      main_prog_(main_prog),
+      // TODO(Yancey1989): Copying graphs is not safely since it deleted the
+      // attrs.
+      graphs_(SeparateMultiDevicesGraph(std::move(graph))) {
   PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
 
+  auto seq_allreduce_pass =
+      ir::PassRegistry::Instance().Get("all_reduce_deps_pass");
+  seq_allreduce_pass->Erase(details::kAllOpDescs);
+  seq_allreduce_pass->Set<const std::vector<OpDesc *>>(
+      details::kAllOpDescs,
+      new std::vector<OpDesc *>(main_prog_.Block(0).AllOps()));
+  for (size_t i = 0; i < graphs_.size(); ++i) {
+    graphs_[i] = seq_allreduce_pass->Apply(std::move(graphs_[i]));
+  }
+
   // set the correct size of thread pool to each device.
   strategy_.num_threads_ = strategy_.num_threads_ < places_.size()
                                ? 1UL
@@ -37,7 +107,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
           << " to run the operators of the graph on each device.";
   for (size_t i = 0; i < places.size(); ++i) {
     executors_.emplace_back(new details::ThreadedSSAGraphExecutor(
-        strategy_, {local_scopes_[i]}, {places_[i]}, std::move(graphs_[i])));
+        strategy_, local_scopes_, {places_[i]}, std::move(graphs_.at(i))));
   }
 }
 

paddle/fluid/framework/details/parallel_ssa_graph_executor.h

@@ -18,7 +18,9 @@
 #include <vector>
 
 #include "ThreadPool.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
+#include "paddle/fluid/framework/ir/graph.h"
 
 namespace paddle {
 namespace framework {
@@ -29,17 +31,23 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor {
   ParallelSSAGraphExecutor(const ExecutionStrategy &strategy,
                            const std::vector<Scope *> &local_scopes,
                            const std::vector<platform::Place> &places,
-                           std::vector<std::unique_ptr<ir::Graph>> &&graphs);
+                           const framework::ProgramDesc &main_prog,
+                           std::unique_ptr<ir::Graph> &&graph);
   ~ParallelSSAGraphExecutor() final = default;
+
   const ir::Graph &Graph() const override { return *graphs_[0]; }
 
   FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
 
  private:
+  std::vector<std::unique_ptr<ir::Graph>> SeparateMultiDevicesGraph(
+      std::unique_ptr<ir::Graph> &&graph);
+
   ExecutionStrategy strategy_;
   std::vector<Scope *> local_scopes_;
   std::unique_ptr<::ThreadPool> pool_{nullptr};
   std::vector<platform::Place> places_;
+  framework::ProgramDesc main_prog_;
   std::vector<std::unique_ptr<ir::Graph>> graphs_;
 
   std::vector<std::unique_ptr<details::ThreadedSSAGraphExecutor>> executors_;

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -219,7 +219,7 @@ void ThreadedSSAGraphExecutor::RunOp(
       VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;
       ready_var_q->Extend(op->Outputs());
-      VLOG(10) << op << " " << op->Name() << "Signal posted";
+      VLOG(10) << op << " " << op->Name() << " Signal posted";
     } catch (...) {
       exception_holder_.Catch(std::current_exception());
     }

paddle/fluid/framework/ir/graph.h

@@ -26,6 +26,14 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
+
+namespace details {
+
+// This attr is not recommended, because the graph should not dependence
+// the program once it is built.
+constexpr char kAllOpDescs[] = "all_op_descs";
+}  //  namespace details
+
 namespace ir {
 
 /*
@@ -168,10 +176,13 @@ class Graph {
     return ret;
   }
 
-  void RemoveNode(ir::Node *node) {
+  std::unique_ptr<ir::Node> RemoveNode(ir::Node *node) {
     PADDLE_ENFORCE(node_set_.find(node) != node_set_.end());
-    node_set_.erase(node);
+    std::unique_ptr<ir::Node> ret;
+    ret.reset(nodes_.at(node).release());
     nodes_.erase(node);
+    node_set_.erase(node);
+    return ret;
   }
 
   // NOTE low performance, but simple and secure.
@@ -184,13 +195,6 @@ class Graph {
     return nullptr;
   }
 
-  void ResolveHazard(
-      const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
-
- private:
-  std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
-      const ProgramDesc &program);
-
   // This method takes ownership of `node`.
   ir::Node *AddNode(ir::Node *node) {
     PADDLE_ENFORCE(node_set_.find(node) == node_set_.end());
@@ -199,6 +203,13 @@ class Graph {
     return node;
   }
 
+  void ResolveHazard(
+      const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
+
+ private:
+  std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
+      const ProgramDesc &program);
+
   // NOTE: program_ shouldn't be exposed to user.
   const ProgramDesc program_;
   std::map<std::string, boost::any> attrs_;

paddle/fluid/framework/parallel_executor.cc

@@ -21,6 +21,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/ir/graph.h"
 
+#include "paddle/fluid/framework/details/all_reduce_deps_pass.h"
 #include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
@@ -193,7 +194,6 @@ ParallelExecutor::ParallelExecutor(
   member_->use_all_reduce_ =
       build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce;
   member_->nranks_ = build_strategy.num_trainers_ * places.size();
-
   if (!member_->use_all_reduce_) {
     PADDLE_ENFORCE(places.size() > 1,
                    "If you set build_strategy.reduce with 'Reduce',"
@@ -221,9 +221,10 @@ ParallelExecutor::ParallelExecutor(
   // choice the execution strategy.
   build_strategy.enable_parallel_graph_ =
       EnableParallelGraphExecution(main_program, exec_strategy, build_strategy);
-
-  VLOG(1) << "Enable ParallelGraph Execution: "
-          << build_strategy.enable_parallel_graph_;
+  if (build_strategy.enable_parallel_graph_)
+    VLOG(0) << "The Executor would execute the graph by ParallelGraph "
+               "Execution which can get better performance,"
+            << "you can force it off by env FLAGS_enable_parallel_graph=0";
 
   if (member_->use_cuda_) {
 // Bcast Parameters to all GPUs
@@ -257,42 +258,27 @@ ParallelExecutor::ParallelExecutor(
 
   // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
   // ncclOp
-  std::vector<std::unique_ptr<ir::Graph>> graphs;
+  std::unique_ptr<ir::Graph> graph;
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-  if (build_strategy.enable_parallel_graph_) {
-    for (size_t i = 0; i < member_->places_.size(); ++i) {
-      std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
-          main_program, {member_->places_[i]}, loss_var_name,
-          {member_->local_scopes_[i]}, member_->nranks_, member_->use_cuda_,
-          member_->nccl_ctxs_.get());
-      graphs.push_back(std::move(graph));
-    }
-  } else {
-    std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
-        main_program, member_->places_, loss_var_name, member_->local_scopes_,
-        member_->nranks_, member_->use_cuda_, member_->nccl_ctxs_.get());
-    graphs.push_back(std::move(graph));
-  }
+  graph = build_strategy.Apply(main_program, member_->places_, loss_var_name,
+                               member_->local_scopes_, member_->nranks_,
+                               member_->use_cuda_, member_->nccl_ctxs_.get());
 #else
-  std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
-      main_program, member_->places_, loss_var_name, member_->local_scopes_,
-      member_->nranks_, member_->use_cuda_);
-  graphs.push_back(std::move(graph));
+  graph = build_strategy.Apply(main_program, member_->places_, loss_var_name,
+                               member_->local_scopes_, member_->nranks_,
+                               member_->use_cuda_);
 #endif
   auto max_memory_size = GetEagerDeletionThreshold();
   VLOG(10) << "Eager Deletion Threshold "
            << static_cast<float>(max_memory_size) / (1 << 30);
   if (max_memory_size >= 0) {
-    for (size_t i = 0; i < graphs.size(); ++i) {
-      graphs[i] = member_->PrepareGCAndRefCnts(
-          std::move(graphs[i]), static_cast<size_t>(max_memory_size));
-    }
+    graph = member_->PrepareGCAndRefCnts(std::move(graph),
+                                         static_cast<size_t>(max_memory_size));
   }
 
   // Step 3. Create vars in each scope. Passes may also create new vars.
   //         skip control vars and empty vars
   std::vector<details::VariableInfo> var_infos;
-  for (auto &graph : graphs) {
   for (auto &node : graph->Nodes()) {
     if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
       var_infos.emplace_back();
@@ -301,16 +287,15 @@ ParallelExecutor::ParallelExecutor(
       var_infos.back().persistable_ = node->Var()->Persistable();
     }
   }
-  }
 
   // If the loss_var_name is given, the number of graph should be only one.
   if (loss_var_name.size()) {
-    size_t graph_num = ir::GraphNum(*graphs[0]);
+    size_t graph_num = ir::GraphNum(*graph);
     if (graph_num > 1) {
       LOG(WARNING)
           << "The number of graph should be only one, "
              "but the current graph has "
-          << ir::GraphNum(*graphs[0])
+          << ir::GraphNum(*graph)
           << " sub_graphs. If you want to see the nodes of the "
              "sub_graphs, you should use 'FLAGS_print_sub_graph_dir' "
              "to specify the output dir. NOTES: if you not do training, "
@@ -319,18 +304,25 @@ ParallelExecutor::ParallelExecutor(
   }
 
   if (build_strategy.enable_parallel_graph_) {
+#ifdef PADDLE_WITH_CUDA
+    // TODO(Yancey1989): Remove passing in the main_program when
+    // allreduce_seq_pass doesn't need it as the attr.
     member_->executor_.reset(new details::ParallelSSAGraphExecutor(
-        exec_strategy, member_->local_scopes_, member_->places_,
-        std::move(graphs)));
+        exec_strategy, member_->local_scopes_, member_->places_, main_program,
+        std::move(graph)));
+#else
+    PADDLE_THROW(
+        "Paddle should be compiled with CUDA for ParallelGraph Execution.");
+#endif
   } else {
     if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
       member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
           exec_strategy, member_->local_scopes_, member_->places_,
-          std::move(graphs[0])));
+          std::move(graph)));
     } else {
       member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
           exec_strategy, member_->local_scopes_, member_->places_,
-          std::move(graphs[0])));
+          std::move(graph)));
     }
   }
 
@@ -482,7 +474,6 @@ bool ParallelExecutor::EnableParallelGraphExecution(
   }
 
   if (!member_->use_all_reduce_ || !member_->use_cuda_)
-    enable_parallel_graph = false;
 
     if (build_strategy.enable_sequential_execution_ ||
         exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental)

python/paddle/fluid/tests/unittests/test_parallel_executor_pg.py

+# 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.
+
+from __future__ import print_function
+
+import unittest
+
+import numpy as np
+import os
+os.environ['FLAGS_enable_parallel_graph'] = str(1)
+import paddle.fluid.core as core
+import os
+import paddle.fluid as fluid
+from parallel_executor_test_base import TestParallelExecutorBase
+
+
+def simple_fc_net(use_feed):
+    img = fluid.layers.data(name='image', shape=[784], dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+    hidden = img
+    for _ in range(4):
+        hidden = fluid.layers.fc(
+            hidden,
+            size=200,
+            act='tanh',
+            bias_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=1.0)))
+    prediction = fluid.layers.fc(hidden, size=10, act='softmax')
+    loss = fluid.layers.cross_entropy(input=prediction, label=label)
+    loss = fluid.layers.mean(loss)
+    return loss
+
+
+class TestMNIST(TestParallelExecutorBase):
+    @classmethod
+    def setUpClass(cls):
+        os.environ['CPU_NUM'] = str(4)
+
+    def _init_data(self):
+        np.random.seed(5)
+        img = np.random.random(size=[32, 784]).astype(np.float32)
+        label = np.ones(shape=[32, 1], dtype='int64')
+        return img, label
+
+    # simple_fc
+    def check_simple_fc_convergence(self, use_cuda, use_reduce=False):
+        if use_cuda and not core.is_compiled_with_cuda():
+            return
+
+        img, label = self._init_data()
+
+        self.check_network_convergence(
+            simple_fc_net,
+            feed_dict={"image": img,
+                       "label": label},
+            use_cuda=use_cuda,
+            use_reduce=use_reduce)
+
+    def test_simple_fc(self):
+        # use_cuda
+        self.check_simple_fc_convergence(True)
+
+    def check_simple_fc_parallel_accuracy(self, use_cuda):
+        if use_cuda and not core.is_compiled_with_cuda():
+            return
+
+        img, label = self._init_data()
+
+        single_first_loss, single_last_loss = self.check_network_convergence(
+            method=simple_fc_net,
+            seed=1,
+            feed_dict={"image": img,
+                       "label": label},
+            use_cuda=use_cuda,
+            use_parallel_executor=False)
+        parallel_first_loss, parallel_last_loss = self.check_network_convergence(
+            method=simple_fc_net,
+            seed=1,
+            feed_dict={"image": img,
+                       "label": label},
+            use_cuda=use_cuda,
+            use_parallel_executor=True)
+
+        self.assertAlmostEquals(
+            np.mean(parallel_first_loss),
+            single_first_loss,
+            delta=1e-6, )
+        self.assertAlmostEquals(
+            np.mean(parallel_last_loss), single_last_loss, delta=1e-6)
+
+    def test_simple_fc_parallel_accuracy(self):
+        self.check_simple_fc_parallel_accuracy(True)
+
+
+if __name__ == '__main__':
+    unittest.main()