【静态图性能优化】图依赖信息复用 (PaddlePaddle#55389)

* add share api for DependencyBuilder

* add judge codes for sharing build results

* add ShareBuildResultsFrom

* update ShareDependencyFrom

* fix error

* add share codes

* fix memory error

* update according review

* update notes

* fix code style

* remove const_cast

* fix code style

paddle/fluid/framework/new_executor/interpreter/dependency_builder.cc

@@ -60,18 +60,26 @@ const std::string StringizeDownstreamMap(
   return oss.str();
 }
 
+DependencyBuilder::DependencyBuilder()
+    : is_build_(false), instructions_(nullptr) {
+  op_downstream_map_ = std::make_shared<std::map<size_t, std::set<size_t>>>();
+  op_happens_before_ = std::make_shared<std::vector<std::vector<bool>>>();
+}
+
 const std::map<size_t, std::set<size_t>>& DependencyBuilder::Build(
     const std::vector<Instruction>& instructions) {
   if (is_build_) {
-    return op_downstream_map_;
+    return *op_downstream_map_;
   }
 
+  std::tie(op_downstream_map_, op_happens_before_) = GetDependency();
+
   instructions_ = &instructions;
   op_num_ = instructions_->size();
 
   ops_before_.assign(op_num_, {});
   ops_behind_.assign(op_num_, {});
-  op_happens_before_.assign(op_num_, std::vector<bool>(op_num_, false));
+  op_happens_before_->assign(op_num_, std::vector<bool>(op_num_, false));
 
   BuildDownstreamMap();
   VLOG(6) << "Finish BuildDownstreamMap";
@@ -97,13 +105,24 @@ const std::map<size_t, std::set<size_t>>& DependencyBuilder::Build(
   VLOG(6) << "Finish AddDependencyForReadOp";
 
   VLOG(6) << "Finish build dependency";
-  VLOG(8) << "downstream count: " << CountDownstreamMap(op_downstream_map_);
+  VLOG(8) << "downstream count: " << CountDownstreamMap(*op_downstream_map_);
   VLOG(8) << "downstream_map: " << std::endl
-          << StringizeDownstreamMap(op_downstream_map_);
+          << StringizeDownstreamMap(*op_downstream_map_);
 
   is_build_ = true;
 
-  return op_downstream_map_;
+  return *op_downstream_map_;
+}
+
+std::tuple<std::shared_ptr<std::map<size_t, std::set<size_t>>>,
+           std::shared_ptr<std::vector<std::vector<bool>>>>
+DependencyBuilder::GetDependency() const {
+  return std::make_tuple(op_downstream_map_, op_happens_before_);
+}
+
+void DependencyBuilder::ShareDependencyFrom(const DependencyBuilder& src) {
+  std::tie(op_downstream_map_, op_happens_before_) = src.GetDependency();
+  is_build_ = true;
 }
 
 const std::map<size_t, std::set<size_t>>& DependencyBuilder::OpDownstreamMap()
@@ -113,7 +132,7 @@ const std::map<size_t, std::set<size_t>>& DependencyBuilder::OpDownstreamMap()
       true,
       phi::errors::Unavailable(
           "DependencyBuilder is not yet built, call Build() firstly."));
-  return op_downstream_map_;
+  return *op_downstream_map_;
 }
 
 void DependencyBuilder::AddDependencyForCoalesceTensorOp() {
@@ -268,8 +287,8 @@ void DependencyBuilder::AddDependencyForRandomOp() {
 void DependencyBuilder::AddDependencyForReadOp() {
   std::vector<bool> is_startup_ops(op_num_, true);
   for (size_t op_idx = 0; op_idx < op_num_; ++op_idx) {
-    auto it = op_downstream_map_.find(op_idx);
-    if (it != op_downstream_map_.end()) {
+    auto it = op_downstream_map_->find(op_idx);
+    if (it != op_downstream_map_->end()) {
       for (size_t downstream_op_idx : it->second) {
         is_startup_ops[downstream_op_idx] = false;
       }
@@ -320,8 +339,7 @@ void DependencyBuilder::AddDownstreamOp(size_t prior_op_idx,
           posterior_op_idx,
           posterior_op_idx,
           prior_op_idx));
-
-  std::set<size_t>& downstream_ops = op_downstream_map_[prior_op_idx];
+  std::set<size_t>& downstream_ops = (*op_downstream_map_)[prior_op_idx];
   // NOTE(Ruibiao): Here the downstream map shrinking is best-effort, therefore
   // ShrinkDownstreamMap after BuildDownstreamMap is still helpful. For example,
   // a->c will not be shrinked in the following case: AddDownstreamOp(a, b) ->
@@ -342,8 +360,8 @@ void DependencyBuilder::AddDownstreamOp(size_t prior_op_idx,
 
   auto update_op_happen_before = [this](size_t prior_op_idx,
                                         size_t posterior_op_idx) {
-    if (!op_happens_before_[prior_op_idx][posterior_op_idx]) {
-      op_happens_before_[prior_op_idx][posterior_op_idx] = true;
+    if (!(*op_happens_before_)[prior_op_idx][posterior_op_idx]) {
+      (*op_happens_before_)[prior_op_idx][posterior_op_idx] = true;
       ops_before_[posterior_op_idx].push_back(prior_op_idx);
       ops_behind_[prior_op_idx].push_back(posterior_op_idx);
     }
@@ -377,8 +395,8 @@ void DependencyBuilder::BuildDownstreamMap() {
       std::map<size_t, size_t>();  // # map from variable to recent write op.
   auto op2dependences =
       std::map<size_t,
-               std::set<size_t>>();  //# map from op to the dependence list,
-                                     // op must run after the dependence.
+               std::set<size_t>>();  // # map from op to the dependence list,
+                                     //  op must run after the dependence.
   std::set<size_t>
       remove_duplicate;  // remove the duplicate between inputs and outputs
 
@@ -497,15 +515,15 @@ void DependencyBuilder::ShrinkDownstreamMap() {
   // shrink, find the downstream op that has no other op in the
   // downstream list happens before it
   for (size_t i = 0; i < op_num_; ++i) {
-    if (op_downstream_map_.find(i) == op_downstream_map_.end()) {
+    if (op_downstream_map_->find(i) == op_downstream_map_->end()) {
       continue;
     }
 
     std::set<size_t> minumum_nexts;
-    for (size_t item : op_downstream_map_.at(i)) {
+    for (size_t item : op_downstream_map_->at(i)) {
       bool not_after_any = true;
       // find the op that is not executed after any
-      for (size_t other_item : op_downstream_map_.at(i)) {
+      for (size_t other_item : op_downstream_map_->at(i)) {
         if (OpHappensBefore(other_item, item)) {
           VLOG(8) << "happens_before: " << other_item << "->" << item
                   << ", so skip " << item;
@@ -520,12 +538,12 @@ void DependencyBuilder::ShrinkDownstreamMap() {
     }
     // NOTE(Ruibiao): op_happens_before will not be changed when shrink
     // dowstream map
-    op_downstream_map_.at(i) = minumum_nexts;
+    (*op_downstream_map_)[i] = minumum_nexts;
   }
   VLOG(8) << "Finish shrink downstream map";
-  VLOG(8) << "downstream count: " << CountDownstreamMap(op_downstream_map_);
+  VLOG(8) << "downstream count: " << CountDownstreamMap(*op_downstream_map_);
   VLOG(8) << "downstream_map: " << std::endl
-          << StringizeDownstreamMap(op_downstream_map_);
+          << StringizeDownstreamMap(*op_downstream_map_);
 }
 
 /// ======================== ///

paddle/fluid/framework/new_executor/interpreter/dependency_builder.h

@@ -34,23 +34,29 @@ namespace interpreter {
 
 class DependencyBuilder {
  public:
-  DependencyBuilder() : is_build_(false), instructions_(nullptr) {}
+  DependencyBuilder();
 
   // build op dependencies and return the mapping from op to its downstream-op
   // set
   const std::map<size_t, std::set<size_t>>& Build(
       const std::vector<Instruction>& instructions);
 
+  std::tuple<std::shared_ptr<std::map<size_t, std::set<size_t>>>,
+             std::shared_ptr<std::vector<std::vector<bool>>>>
+  GetDependency() const;
+
   const std::map<size_t, std::set<size_t>>& OpDownstreamMap() const;
 
   bool OpHappensBefore(size_t prior_op_idx, size_t posterior_op_idx) const {
     PADDLE_ENFORCE_GE(
-        op_happens_before_.size(),
+        op_happens_before_->size(),
         0,
         phi::errors::Unavailable("op_happen_before is not yet built"));
-    return op_happens_before_.at(prior_op_idx).at(posterior_op_idx);
+    return op_happens_before_->at(prior_op_idx).at(posterior_op_idx);
   }
 
+  void ShareDependencyFrom(const DependencyBuilder& src);
+
  private:
   void AddDependencyForCoalesceTensorOp();
   void AddDependencyForCommunicationOp();
@@ -76,13 +82,13 @@ class DependencyBuilder {
   std::vector<std::vector<size_t>> ops_behind_;
 
   // op_downstream_map_ is the mapping from op to its downstream-op set, that is
-  // to say, op_downstream_map_[i] == {a, b, c} means op[a], op[b] and op[c]
+  // to say, (*op_downstream_map_)[i] == {a, b, c} means op[a], op[b] and op[c]
   // depend on op[i] directly.
-  std::map<size_t, std::set<size_t>> op_downstream_map_;
+  std::shared_ptr<std::map<size_t, std::set<size_t>>> op_downstream_map_;
 
   // op_happens_before_ is a matrix form of ops_before_ and ops_behind_, it is
   // used to speed up the query.
-  std::vector<std::vector<bool>> op_happens_before_;
+  std::shared_ptr<std::vector<std::vector<bool>>> op_happens_before_;
 };
 
 // /// ======================== ///

paddle/fluid/framework/new_executor/interpreter_base_impl.h

@@ -80,6 +80,8 @@ class InterpreterBaseImpl {
 
   virtual void ShareWorkQueueFrom(InterpreterBaseImpl* src) = 0;
 
+  virtual void ShareBuildResultsFrom(const InterpreterBaseImpl& src) = 0;
+
   virtual void SetCopyProgram(std::shared_ptr<ProgramDesc> prog) = 0;
 
   virtual void SetSkipGcVars(const std::set<std::string>& skip_gc_vars) = 0;
@@ -97,6 +99,11 @@ class InterpreterBaseImpl {
   virtual const platform::Place& GetPlace() const = 0;
 
   virtual void SetOutputHooks(const std::vector<HookFunc>& hookfuncs) = 0;
+
+  virtual const interpreter::DependencyBuilder& GetDependencyBuilder()
+      const = 0;
+
+  virtual std::shared_ptr<std::vector<size_t>> GetDependencyCount() const = 0;
 };
 
 inline void SetDeviceId(const platform::Place& place) {

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -81,6 +81,12 @@ void InterpreterCore::ShareWorkQueueFrom(std::shared_ptr<InterpreterCore> src) {
   impl_->ShareWorkQueueFrom(const_cast<InterpreterBaseImpl*>(src->Impl()));
 }
 
+void InterpreterCore::ShareBuildResultsFrom(
+    std::shared_ptr<InterpreterCore> src) {
+  // ShareBuildResultsFrom required const InterpreterBaseImpl& src as input
+  impl_->ShareBuildResultsFrom(*src->Impl());
+}
+
 void InterpreterCore::SetCopyProgram(std::shared_ptr<ProgramDesc> prog) {
   impl_->SetCopyProgram(prog);
 }

paddle/fluid/framework/new_executor/interpretercore.h

@@ -56,6 +56,8 @@ class InterpreterCore {
 
   void ShareWorkQueueFrom(std::shared_ptr<InterpreterCore> src);
 
+  void ShareBuildResultsFrom(std::shared_ptr<InterpreterCore> src);
+
   void SetCopyProgram(std::shared_ptr<ProgramDesc> prog);
 
   void SetSkipGcVars(const std::set<std::string>& skip_gc_vars);

paddle/fluid/framework/new_executor/new_ir_interpreter.cc

@@ -351,6 +351,24 @@ void NewIRInterpreter::ShareWorkQueueFrom(InterpreterBaseImpl* src) {
           << ") to InterpreterCore(" << this << ")";
 }
 
+void NewIRInterpreter::ShareBuildResultsFrom(const InterpreterBaseImpl& src) {
+  PADDLE_THROW(platform::errors::Unimplemented(
+      "ShareBuildResultsFrom is not implemented in NewIRInterpreter."));
+}
+
+// op dependences
+const interpreter::DependencyBuilder& NewIRInterpreter::GetDependencyBuilder()
+    const {
+  PADDLE_THROW(platform::errors::Unimplemented(
+      "GetDependencyBuilder is not implemented in NewIRInterpreter."));
+}
+
+std::shared_ptr<std::vector<size_t>> NewIRInterpreter::GetDependencyCount()
+    const {
+  PADDLE_THROW(platform::errors::Unimplemented(
+      "GetDependencyCount is not implemented in NewIRInterpreter."));
+}
+
 bool NewIRInterpreter::BuildInplaceCheckVarIsOnlyInput(
     const std::vector<std::vector<size_t>>& input_var2op, size_t var_index) {
   if (!var_scope_.VarDesc(var_index)) {

paddle/fluid/framework/new_executor/new_ir_interpreter.h

@@ -53,6 +53,13 @@ class NewIRInterpreter : public InterpreterBaseImpl {
 
   void ShareWorkQueueFrom(InterpreterBaseImpl* src) override;
 
+  void ShareBuildResultsFrom(const InterpreterBaseImpl& src) override;
+
+  // op dependences
+  const interpreter::DependencyBuilder& GetDependencyBuilder() const override;
+
+  std::shared_ptr<std::vector<size_t>> GetDependencyCount() const override;
+
   void SetCopyProgram(std::shared_ptr<ProgramDesc> prog) override;
 
   void SetSkipGcVars(const std::set<std::string>& skip_gc_vars) override;