Fix the interface of Pass::Apply (#16484)

* modify the interface of Pass::Allay
test=develop

* Polish code
test=develop

* Fix Travis CI
test=develop

* fix Pass::Apply interface
test=develop

* Fix Travis CI
test=develop

Author: chengduo

paddle/fluid/framework/details/all_reduce_deps_pass.cc

@@ -42,8 +42,7 @@ VarHandle* GetValidInput(const OpHandleBase* a) {
   return nullptr;
 }
 
-std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
-    std::unique_ptr<ir::Graph> graph) const {
+void AllReduceDepsPass::ApplyImpl(ir::Graph* graph) const {
   auto graph_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
 
   // get vars order
@@ -131,8 +130,6 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
     VLOG(10) << "pre_op:" << pre_op->DebugString()
              << ", op:" << op->DebugString();
   }
-
-  return graph;
 }
 
 }  // namespace details

paddle/fluid/framework/details/all_reduce_deps_pass.h

@@ -24,8 +24,7 @@ namespace details {
 // TODO(gongwb): overlap allreduce with backward computation.
 class AllReduceDepsPass : public ir::Pass {
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(
-      std::unique_ptr<ir::Graph> graph) const override;
+  void ApplyImpl(ir::Graph* graph) const override;
 };
 
 }  // namespace details

paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.cc

@@ -46,8 +46,7 @@ static framework::proto::VarType::Type kDefaultDtype =
 
 class AllocContinuousSpaceForGradPass : public ir::Pass {
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(
-      std::unique_ptr<ir::Graph> graph) const override {
+  void ApplyImpl(ir::Graph *graph) const override {
     ir::Graph &result = *graph;
 
     auto &places = Get<const std::vector<platform::Place>>(kPlaces);
@@ -65,7 +64,7 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
 
     if (params_grads.size() == 0) {
       VLOG(10) << "Doesn't find gradients";
-      return std::move(graph);
+      return;
     }
 
     std::unordered_map<std::string, ir::Node *> vars;
@@ -124,8 +123,6 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
 
     InitFusedVarsAndAllocSpaceForVars(places, local_scopes, vars,
                                       fused_var_name, params_grads);
-
-    return std::move(graph);
   }
 
   template <typename AttrType>

paddle/fluid/framework/details/build_strategy.cc

@@ -204,15 +204,16 @@ bool BuildStrategy::IsMultiDevPass(const std::string &pass_name) const {
   return framework::details::MultiDevSSAGraphBuilder().count(pass_name) > 0;
 }
 
-std::unique_ptr<ir::Graph> BuildStrategy::Apply(
-    std::unique_ptr<ir::Graph> graph,
-    const std::vector<platform::Place> &places,
-    const std::string &loss_var_name, const std::vector<Scope *> &local_scopes,
-    const size_t &nranks,
+ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
+                                const std::vector<platform::Place> &places,
+                                const std::string &loss_var_name,
+                                const std::vector<Scope *> &local_scopes,
+                                const size_t &nranks,
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-    const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const {
+                                const bool use_cuda,
+                                platform::NCCLContextMap *nccl_ctxs) const {
 #else
-    const bool use_cuda) const {
+                                const bool use_cuda) const {
 #endif
   // Create a default one if not finalized by user.
   CreatePassesFromStrategy(false);
@@ -265,7 +266,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       }
     }
     VLOG(3) << "Start Apply Pass " << pass->Type();
-    graph = pass->Apply(std::move(graph));
+    graph = pass->Apply(graph);
     VLOG(3) << "Finish Apply Pass " << pass->Type();
   }
   return graph;

paddle/fluid/framework/details/build_strategy.h

@@ -120,16 +120,15 @@ struct BuildStrategy {
 
   // Apply the passes built by the pass_builder_. The passes will be
   // applied to the Program and output an ir::Graph.
-  std::unique_ptr<ir::Graph> Apply(std::unique_ptr<ir::Graph> graph,
-                                   const std::vector<platform::Place> &places,
-                                   const std::string &loss_var_name,
-                                   const std::vector<Scope *> &local_scopes,
-                                   const size_t &nranks,
+  ir::Graph *Apply(ir::Graph *graph, const std::vector<platform::Place> &places,
+                   const std::string &loss_var_name,
+                   const std::vector<Scope *> &local_scopes,
+                   const size_t &nranks,
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-                                   const bool use_cuda,
-                                   platform::NCCLContextMap *nccl_ctxs) const;
+                   const bool use_cuda,
+                   platform::NCCLContextMap *nccl_ctxs) const;
 #else
-                                   const bool use_cuda) const;
+                   const bool use_cuda) const;
 #endif
 
   // If set true, ParallelExecutor would build the main_program into multiple

paddle/fluid/framework/details/eager_deletion_pass.cc

@@ -170,12 +170,10 @@ static OpToVarNameSetMap ShrinkGCVars(
 
 class EagerDeletionPass : public ir::Pass {
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(
-      std::unique_ptr<ir::Graph> graph) const override;
+  void ApplyImpl(ir::Graph *graph) const override;
 };
 
-std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
-    std::unique_ptr<ir::Graph> graph) const {
+void EagerDeletionPass::ApplyImpl(ir::Graph *graph) const {
   auto &ref_cnts =
       Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount);
   PADDLE_ENFORCE(ref_cnts.empty(),
@@ -240,7 +238,7 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
 
   auto while_op_eager_deletion_pass =
       ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass");
-  return while_op_eager_deletion_pass->Apply(std::move(graph));
+  while_op_eager_deletion_pass->Apply(graph);
 }
 
 }  // namespace details

paddle/fluid/framework/details/fuse_all_reduce_op_pass.cc

@@ -28,8 +28,7 @@ namespace details {
 
 class FuseAllReduceOpPass : public ir::Pass {
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(
-      std::unique_ptr<ir::Graph> graph) const override {
+  void ApplyImpl(ir::Graph *graph) const override {
     ir::Graph &result = *graph;
 
     auto &places = Get<const std::vector<platform::Place>>(kPlaces);
@@ -71,7 +70,7 @@ class FuseAllReduceOpPass : public ir::Pass {
 
     VLOG(10) << "Find all_reduce_ops: " << all_reduce_ops.size();
     if (all_reduce_ops.size() == 0) {
-      return std::move(graph);
+      return;
     }
 
     PADDLE_ENFORCE_EQ(all_reduce_ops.size(), grads.size(),
@@ -99,7 +98,6 @@ class FuseAllReduceOpPass : public ir::Pass {
                            group_all_reduce_ops, &result);
 #endif
     }
-    return std::move(graph);
   }
 
   void InsertFusedAllReduce(const std::vector<platform::Place> &places,

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -144,22 +144,19 @@ void InplacePass::InitSSAGraphNodes() const {
   }
 }
 
-std::unique_ptr<ir::Graph> InplacePass::ApplyImpl(
-    std::unique_ptr<ir::Graph> graph) const {
+void InplacePass::ApplyImpl(ir::Graph* graph) const {
   var_nodes_.clear();
-  view_.Build(graph.get());
+  view_.Build(graph);
   InitSSAGraphNodes();
 
   auto cnt = 0;
   for (auto* op : view_.AllOps()) {
     VLOG(4) << "Handle op " << cnt++ << ": " << op->Name();
     if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name()))
       continue;
-    TryInplaceOpInputOutput(op, graph.get());
+    TryInplaceOpInputOutput(op, graph);
   }
   // graph->ResolveHazard(var_nodes_);
-
-  return graph;
 }
 
 void InplacePass::InplaceModifyDesc(const std::string& var,

paddle/fluid/framework/details/inplace_op_pass.h

@@ -69,8 +69,7 @@ class InplacePass : public ir::Pass {
   InplacePass();
 
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(
-      std::unique_ptr<ir::Graph> graph) const override;
+  void ApplyImpl(ir::Graph* graph) const override;
 
   void InitSSAGraphNodes() const;
 

paddle/fluid/framework/details/memory_optimize_pass.cc

@@ -44,8 +44,7 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
-    std::unique_ptr<ir::Graph> graph) const {
+void MemoryOptimizePass::ApplyImpl(ir::Graph* graph) const {
   auto nodes = graph->Nodes();
   CollectSkipVarsSet(nodes);
 
@@ -113,7 +112,7 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
 
           cfg_->RenameVarInCFGGraph(var_name, cache_name, idx);
           RenameVarInGraphDesc(var_name, cache_name, idx);
-          RenameVarInGraphNode(var_name, cache_name, idx, graph.get());
+          RenameVarInGraphNode(var_name, cache_name, idx, graph);
           pool_.Erase(cache_name);
         }
       }
@@ -128,8 +127,6 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
     }
   }
   graph->ResolveHazard(var_nodes_);
-
-  return graph;
 }
 
 void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {