!694 Refactor vm module for multigraph sink.

Merge pull request !694 from rick_sanchez/master

mindspore/ccsrc/session/ascend_session.cc

@@ -800,45 +800,77 @@ void AscendSession::UpdateGraphOrder(GraphId to_graph_id) {
   }
 }
 
+size_t AscendSession::SetChildGraphInput(const KernelGraphPtr &graph, const AnfNodePtr &node, size_t input_index) {
+  auto output_num = AnfAlgo::GetOutputTensorNum(node);
+  if (output_num > 1 && !AnfAlgo::CheckPrimitiveType(node, prim::kPrimTupleGetItem)) {
+    return input_index + output_num;
+  }
+  auto &graph_inputs = graph->inputs();
+  auto &valid_inputs = graph->ValidInputs();
+  if (valid_inputs[input_index]) {
+    SetChildGraphParameter(node, graph_inputs[input_index]);
+  } else {
+    MS_LOG(DEBUG) << "Invalid input arg: " << node->DebugString();
+  }
+  return ++input_index;
+}
+
+size_t AscendSession::SetChildGraphInput(const KernelGraphPtr &graph, const ValuePtr &value, size_t input_index) {
+  MS_EXCEPTION_IF_NULL(value);
+  if (!value->isa<Tensor>()) {
+    MS_LOG(EXCEPTION) << "Value Node should be a tensor, unexpected value: " << value->ToString();
+  }
+  auto &graph_inputs = graph->inputs();
+  SetChildGraphParameter(value->cast<TensorPtr>(), graph_inputs[input_index]);
+  return ++input_index;
+}
+
+size_t AscendSession::SetChildGraphInput(const KernelGraphPtr &graph, const VectorRef &vec_args, size_t input_index) {
+  auto index = input_index;
+  for (auto &arg : vec_args) {
+    if (utils::isa<AnfNodePtr>(arg)) {
+      // arg is a anf node
+      auto node = utils::cast<AnfNodePtr>(arg);
+      index = SetChildGraphInput(graph, node, input_index);
+    } else if (utils::isa<ValuePtr>(arg)) {
+      // arg is a tensor
+      auto value = utils::cast<ValuePtr>(arg);
+      index = SetChildGraphInput(graph, value, input_index);
+    } else {
+      MS_LOG(EXCEPTION) << "Unexpected arg type " << arg.ToString();
+    }
+  }
+  return index;
+}
+
 void AscendSession::SetChildGraphInput(GraphId g, const VectorRef &args) {
   MS_LOG(INFO) << "Set input of graph " << g;
   auto to_graph = GetGraph(g);
   MS_EXCEPTION_IF_NULL(to_graph);
   DumpGraphInputArgs(args);
   UpdateGraphOrder(g);
-  std::vector<AnfNodePtr> graph_inputs = to_graph->inputs();
-  auto valid_inputs = to_graph->ValidInputs();
+  auto &graph_inputs = to_graph->inputs();
   auto real_args = GetRealArgs(to_graph, args);
   size_t input_index = 0;
   for (size_t i = 0; i < real_args.size(); i++) {
     if (input_index >= graph_inputs.size()) {
       MS_LOG(EXCEPTION) << "input_index " << input_index << " out of range size " << graph_inputs.size();
     }
-    if (utils::isa<AnfNodePtr>(real_args[i])) {
+    auto &real_arg = real_args[i];
+    if (utils::isa<AnfNodePtr>(real_arg)) {
       // arg is a anf node
-      auto real_arg = utils::cast<AnfNodePtr>(real_args[i]);
-      auto real_arg_output_num = AnfAlgo::GetOutputTensorNum(real_arg);
-      if (!AnfAlgo::CheckPrimitiveType(real_arg, prim::kPrimTupleGetItem) && real_arg_output_num > 1) {
-        input_index += real_arg_output_num;
-        continue;
-      }
-      if (valid_inputs[input_index]) {
-        SetChildGraphParameter(real_arg, graph_inputs[input_index]);
-      } else {
-        MS_LOG(DEBUG) << "Invalid input arg" << real_arg->DebugString();
-      }
-      input_index++;
-    } else if (utils::isa<ValuePtr>(args[i])) {
-      auto value = utils::cast<ValuePtr>(args[i]);
-      MS_EXCEPTION_IF_NULL(value);
+      auto node = utils::cast<AnfNodePtr>(real_arg);
+      input_index = SetChildGraphInput(to_graph, node, input_index);
+    } else if (utils::isa<ValuePtr>(real_arg)) {
       // arg is a tensor
-      if (!value->isa<Tensor>()) {
-        MS_LOG(EXCEPTION) << "Value Node should be a tensor, unexpected value: " << value->ToString();
-      }
-      SetChildGraphParameter(value->cast<TensorPtr>(), graph_inputs[input_index]);
-      input_index++;
+      auto value = utils::cast<ValuePtr>(real_arg);
+      input_index = SetChildGraphInput(to_graph, value, input_index);
+    } else if (utils::isa<VectorRef>(real_arg)) {
+      // arg is a VectorRef
+      auto vec_args = utils::cast<VectorRef>(real_arg);
+      input_index = SetChildGraphInput(to_graph, vec_args, input_index);
     } else {
-      MS_LOG(EXCEPTION) << "Unexpected arg type " << args[i].ToString();
+      MS_LOG(EXCEPTION) << "Unexpected arg type " << real_arg.ToString();
     }
   }
   MS_LOG(INFO) << "Finish!";

mindspore/ccsrc/session/ascend_session.h

@@ -79,6 +79,10 @@ class AscendSession : public SessionBasic {
   void RunOpHardwareOptimize(const std::shared_ptr<session::KernelGraph> &kernel_graph) const;
   void RunOpExecTask(const std::shared_ptr<KernelGraph> &kernel_graph) const;
 
+  size_t SetChildGraphInput(const KernelGraphPtr &graph, const AnfNodePtr &node, size_t input_index);
+  size_t SetChildGraphInput(const KernelGraphPtr &graph, const ValuePtr &value, size_t input_index);
+  size_t SetChildGraphInput(const KernelGraphPtr &graph, const VectorRef &vec_args, size_t input_index);
+
   // merge execution order list of child graphs
   void MergeGraphExecOrder();
   // insert assion op to sync data bettween different graphs

mindspore/ccsrc/session/kernel_graph.h

@@ -88,7 +88,7 @@ class KernelGraph : public FuncGraph {
   void set_executable(bool executable) { executable_ = executable; }
   // set invalid inputs for control sink
   std::vector<bool> *MutableValidInputs() { return &valid_inputs_; }
-  std::vector<bool> ValidInputs() { return valid_inputs_; }
+  const std::vector<bool> &ValidInputs() const { return valid_inputs_; }
 
  private:
   // remove value node form graph

mindspore/ccsrc/utils/base_ref.h

@@ -228,6 +228,8 @@ T cast(const BaseRef &handle) {
 
 class VectorRef : public BaseRef {
  public:
+  using value_type = BaseRef;
+
   VectorRef() {}
   explicit VectorRef(const std::vector<BaseRef> &elements) : elements_(elements) {}
   VectorRef(const const_iterator &begin, const const_iterator &end) : elements_(begin, end) {}
@@ -251,6 +253,13 @@ class VectorRef : public BaseRef {
     return elements_[dim];
   }
 
+  BaseRef &operator[](const std::size_t &dim) {
+    if (dim >= size()) {
+      MS_LOG(EXCEPTION) << "Out of the size of the tuple.";
+    }
+    return elements_[dim];
+  }
+
   uint32_t type() const override { return tid(); }
   std::string ToString() const override;
   std::vector<BaseRef> &elements() { return elements_; }

mindspore/ccsrc/vm/backend.cc

@@ -143,6 +143,66 @@ void MsBackend::SetSwitchGraph() {
   }
 }
 
+// convert node from formal parameter to actual parameter,
+// and actual parameter is graph user's formal parameter.
+// get top while graph's parameter in recall while.
+AnfNodePtr MsBackend::ConvertGraphInput(const FuncGraphPtr &func_graph, const AnfNodePtr &node) {
+  std::unordered_map<AnfNodePtr, size_t> params_index;
+  auto result = node;
+  auto graph = result->func_graph();
+  while (func_graph != graph) {
+    auto iter = graph_user_inputs_.find(graph);
+    if (iter == graph_user_inputs_.end()) {
+      break;
+    }
+
+    params_index.clear();
+    auto &params = graph->parameters();
+    for (size_t i = 0; i < params.size(); ++i) {
+      params_index[params[i]] = i;
+    }
+
+    graph = iter->second.first;
+    auto &inputs = iter->second.second;
+    result = inputs[params_index[result]];
+  }
+  return result;
+}
+
+void MsBackend::SetGraphUserInputs(const FuncGraphPtr &func_graph, const FuncGraphPtr &user,
+                                   const AnfNodePtrList &inputs) {
+  if (graph_user_inputs_.find(func_graph) != graph_user_inputs_.end()) {
+    return;
+  }
+  graph_user_inputs_[func_graph] = {user, inputs};
+}
+
+void MsBackend::RecallGraphInput(const FuncGraphPtr &func_graph, const VectorRef &args, const BaseRef &c) {
+  std::unordered_map<AnfNodePtr, size_t> params_index;
+  auto &params = func_graph->parameters();
+  for (size_t i = 0; i < params.size(); ++i) {
+    params_index[params[i]] = i;
+  }
+
+  // recall all child graphs in this while
+  auto &graph_inputs = graph_inputs_[c];
+  for (auto &iter : graph_inputs) {
+    auto &graph = iter.first;
+    auto &old_args = iter.second;
+    auto &result = graph_id_map_[graph];
+    auto &inputs = result.inputs;
+    for (size_t i = 0; i < inputs.size(); ++i) {
+      auto input = ConvertGraphInput(func_graph, inputs[i]);
+      auto it = params_index.find(input);
+      if (it != params_index.end()) {
+        old_args[i] = args[it->second];
+      }
+    }
+    sess_->SetChildGraphInput(graph, old_args);
+  }
+  graph_inputs_.erase(c);
+}
+
 // compile set input output
 VectorRef MsBackend::MsSimuRunGraph(const GraphId &g, const VectorRef &args) {
   MS_LOG(DEBUG) << "set graph input:" << g;
@@ -150,13 +210,20 @@ VectorRef MsBackend::MsSimuRunGraph(const GraphId &g, const VectorRef &args) {
   sess_->SetChildGraphInput(g, args);
 
   if (is_switch_call_) {
-    bool curr_cond = simu_cond_map_[curr_switch_].curr_cond;
-    MS_LOG(DEBUG) << "switch call MsSimuRunGraph:" << curr_cond;
-    if (0 == simu_cond_map_[curr_switch_].cond_graph_map.count(curr_cond)) {
-      MS_LOG(DEBUG) << "switch call MsSimuRunGraph:" << curr_cond << ", " << g;
-      simu_cond_map_[curr_switch_].cond_graph_map[curr_cond] = g;
-      SetSwitchGraph();
+    if (!curr_switch_.is_null()) {
+      // push this {g, args} to all user while graph_inputs for nest while,
+      // when current condition recall over delete this cond in graph_inputs.
+      for (auto &iter : graph_inputs_) {
+        iter.second.push_back({g, args});
+      }
+      if (graph_inputs_.find(curr_switch_) == graph_inputs_.end()) {
+        graph_inputs_[curr_switch_].push_back({g, args});
+      }
     }
+    bool curr_cond = simu_cond_map_[curr_switch_].curr_cond;
+    MS_LOG(DEBUG) << "switch call MsSimuRunGraph:" << curr_cond << ", " << g;
+    simu_cond_map_[curr_switch_].cond_graph_map[curr_cond] = g;
+    SetSwitchGraph();
   }
 
   std::vector<BaseRef> outputs;
@@ -205,42 +272,17 @@ VectorRef MsBackend::MsRunGraph(const GraphId &g, const VectorRef &args) {
   return outputs;
 }
 
-void MsBackend::SetSimuCondFlag(const BaseRef &c, int flag) {
-  MS_LOG(DEBUG) << "while set cond :" << c.ToString() << ", " << simu_cond_map_.size();
-
-  if (simu_cond_map_.find(c) == simu_cond_map_.end()) {
-    MS_LOG(EXCEPTION) << "error c not find";
-  }
-  simu_cond_map_[c].flag = flag;
-}
-
-int MsBackend::GetSimuCondFlag(const BaseRef &c) {
-  BaseRef cond = c;
-  if (cond.is_null()) {
-    MS_LOG(DEBUG) << "get curr_switch";
-    cond = curr_switch_;
-  }
-  if (simu_cond_map_.find(cond) == simu_cond_map_.end()) {
-    MS_LOG(ERROR) << "error c not find";
-    return -1;
-  }
-  return simu_cond_map_[cond].flag;
-}
-
 SwitchCondStatus MsBackend::SetSimuCond(const BaseRef &c, bool value) {
   MS_LOG(DEBUG) << "set cond :" << c.ToString() << ", " << simu_cond_map_.size();
 
   CondGraph cond_graph;
   cond_graph.curr_cond = value;
   if (simu_cond_map_.find(c) == simu_cond_map_.end()) {
-    cond_graph.flag = 0;
     simu_cond_map_[c] = cond_graph;
   }
 
   if (simu_cond_map_[c].cond_graph_map.count(value)) {
-    if (value == true) {
-      return kCondAlreadyRun;
-    }
+    return kCondAlreadyRun;
   }
   simu_cond_map_[c].curr_cond = value;
   MS_LOG(DEBUG) << "end set cond ";

mindspore/ccsrc/vm/backend.h

@@ -16,9 +16,11 @@
 #ifndef MINDSPORE_CCSRC_VM_BACKEND_H_
 #define MINDSPORE_CCSRC_VM_BACKEND_H_
 
-#include <string>
+#include <list>
 #include <memory>
+#include <string>
 #include <unordered_map>
+#include <utility>
 
 #include "ir/anf.h"
 #include "vm/segment_runner.h"
@@ -45,6 +47,8 @@ class Backend {
   virtual bool GetCond(const BaseRef &c, bool *value);
   virtual void SetSwitchGraph() {}
   virtual void SetSwitchActive(const BaseRef &, bool) {}
+  virtual void RecallGraphInput(const FuncGraphPtr &, const VectorRef &, const BaseRef &) {}
+  virtual void SetGraphUserInputs(const FuncGraphPtr &, const FuncGraphPtr &, const AnfNodePtrList &) {}
 
   void set_curr_switch(const BaseRef &value) {
     curr_switch_ = value;
@@ -54,8 +58,6 @@ class Backend {
   BaseRef curr_switch() { return curr_switch_; }
   virtual void Link(GraphId) {}
   virtual LinConvertResult GetMultiGraphRun(const FuncGraphPtr &) { return LinConvertResult(); }
-  virtual void SetSimuCondFlag(const BaseRef &, int) {}
-  virtual int GetSimuCondFlag(const BaseRef &) { return 0; }
 
   LinConvertResult multi_result() { return multi_result_; }
   void set_multi_result(const LinConvertResult &value) { multi_result_ = value; }
@@ -75,11 +77,11 @@ class Backend {
   bool simu_flag_;
   LinConvertResult multi_result_;
   AnfNodePtr final_output_;
+  std::unordered_map<FuncGraphPtr, std::pair<FuncGraphPtr, AnfNodePtrList>> graph_user_inputs_;
 };
 
 struct CondGraph {
   bool curr_cond;
-  int flag;
   std::unordered_map<bool, GraphId> cond_graph_map;
 };
 
@@ -97,15 +99,17 @@ class MsBackend : public Backend {
 
   void SetSwitchGraph() override;
   void SetSwitchActive(const BaseRef &c, bool cond) override;
+  void RecallGraphInput(const FuncGraphPtr &, const VectorRef &, const BaseRef &) override;
+  void SetGraphUserInputs(const FuncGraphPtr &, const FuncGraphPtr &, const AnfNodePtrList &) override;
   void Link(GraphId) override;
+  AnfNodePtr ConvertGraphInput(const FuncGraphPtr &, const AnfNodePtr &);
   LinConvertResult GetMultiGraphRun(const FuncGraphPtr &g) override;
-  void SetSimuCondFlag(const BaseRef &c, int flag) override;
-  int GetSimuCondFlag(const BaseRef &c) override;
 
  private:
   session::SessionPtr sess_;
   std::unordered_map<BaseRef, CondGraph, BaseRefHash> simu_cond_map_;
   std::unordered_map<GraphId, LinConvertResult> graph_id_map_;
+  std::unordered_map<BaseRef, std::list<std::pair<GraphId, VectorRef>>, BaseRefHash> graph_inputs_;
 };
 }  // namespace compile
 }  // namespace mindspore

mindspore/ccsrc/vm/transform.cc

@@ -390,6 +390,16 @@ void CompileGraph::AddTailCall(const AnfNodePtr &fn, size_t size) {
 void CompileGraph::AddPartial(const CNodePtr &node) {
   auto inputs = node->inputs();
   VectorRef args;
+  auto fn = inputs[1];
+  if (!IsValueNode<FuncGraph>(fn)) {
+    MS_LOG(EXCEPTION) << "The type of 1st input of node must be FuncGraph";
+  }
+  if (backend_->is_multi_graph_sink()) {
+    auto func_graph = GetValueNode<FuncGraphPtr>(fn);
+    args.emplace_back(func_graph);
+    AnfNodePtrList outs(inputs.begin() + 2, inputs.end());
+    backend_->SetGraphUserInputs(func_graph, node->func_graph(), outs);
+  }
   for (size_t i = 1; i < inputs.size(); i++) {
     args.emplace_back(Ref(inputs[i]));
   }
@@ -442,12 +452,17 @@ void CompileGraph::AddPrimitive(const CNodePtr &node, const PrimitivePtr &prim)
 }
 
 int CompileGraph::AddCall(const FuncGraphPtr &graph, const CNodePtr &node) {
-  auto node_inputs = node->inputs();
-  AnfNodePtr fn = node_inputs[0];
+  auto inputs = node->inputs();
+  AnfNodePtr fn = inputs[0];
+  if (backend_->is_multi_graph_sink() && IsValueNode<FuncGraph>(fn)) {
+    auto func_graph = GetValueNode<FuncGraphPtr>(fn);
+    AnfNodePtrList outs(inputs.begin() + 1, inputs.end());
+    backend_->SetGraphUserInputs(func_graph, node->func_graph(), outs);
+  }
   (void)Ref(fn);
-  size_t size = node_inputs.size();
+  size_t size = inputs.size();
   for (size_t i = size - 1; i > 0; i--) {
-    AddInput(node_inputs[i]);
+    AddInput(inputs[i]);
   }
   if (node == graph->output()) {
     AddTailCall(fn, size);

mindspore/ccsrc/vm/vm.cc

@@ -32,7 +32,8 @@ namespace compile {
 // Arguments:
 //   fn_: Callable function.
 //   args_: Sequence of function args.
-StructPartial::StructPartial(int fn, const VectorRef &args) : fn_(fn), args_(args) {}
+//   fg_: Graph of function.
+StructPartial::StructPartial(int fn, const VectorRef &args, const FuncGraphPtr &fg) : fn_(fn), args_(args), fg_(fg) {}
 
 std::ostream &operator<<(std::ostream &os, const StructPartial &other) {
   os << "partial(" << other.fn_ << ", " << other.args_.ToString() << ")";
@@ -40,7 +41,7 @@ std::ostream &operator<<(std::ostream &os, const StructPartial &other) {
 }
 
 bool operator==(const StructPartial &lhs, const StructPartial &rhs) {
-  return (lhs.fn_ == rhs.fn_ && lhs.args_ == rhs.args_);
+  return (lhs.fn_ == rhs.fn_ && lhs.args_ == rhs.args_ && lhs.fg_ == rhs.fg_);
 }
 
 StructSimuSwitch::StructSimuSwitch(const BaseRef &fn, const BaseRef &value) : fn_(fn), value_(value) {}
@@ -242,16 +243,6 @@ void FinalVM::InstTailCall(const VectorRef &args) {
   int nargs = utils::cast<int>(args[2]);
 
   auto new_jmp = Ref(jmp);
-
-  if (backend_->simu_flag()) {
-    if (backend_->GetSimuCondFlag(BaseRef()) == 2) {
-      MS_LOG(DEBUG) << "invoke while call tail first";
-      Pop(height);
-      Push(1);
-      Popp();
-      return;
-    }
-  }
   MoveStack(nargs, height);
   MS_LOG(DEBUG) << "TailCall pushp:" << pc_ << ", jmp:" << jmp;
   DoJmp(new_jmp);
@@ -291,8 +282,30 @@ void FinalVM::InstReturn(const VectorRef &args) {
   MS_LOG(DEBUG) << "End";
 }
 
-void FinalVM::InstPartial(const VectorRef &args) {
-  MS_LOG(DEBUG) << "Start";
+void FinalVM::InstSimuPartial(const VectorRef &args) {
+  const size_t args_size = 2;
+  if (args.size() < args_size) {
+    MS_LOG(ERROR) << __FUNCTION__ << " requires " << args_size << " or more parameters, while the input size is "
+                  << args.size() << ".";
+    return;
+  }
+
+  auto &node = args[0];
+  if (!utils::isa<FuncGraphPtr>(node)) {
+    MS_LOG(ERROR) << "The type of 1st input of node must be FuncGraph";
+    return;
+  }
+  auto fg = utils::cast<FuncGraphPtr>(node);
+  int fn_ = utils::cast<int>(args[1]);
+  auto fn = utils::cast<int>(Ref(fn_));
+  MS_LOG(DEBUG) << "Partial argssize:" << args.size();
+  std::vector<BaseRef> outs(args.size() - 2);
+  (void)std::transform(args.begin() + 2, args.end(), outs.begin(),
+                       [&, this](const BaseRef &a) { return Ref(utils::cast<int>(a)); });
+  Push(std::make_shared<StructPartial>(fn, VectorRef(outs), fg));
+}
+
+void FinalVM::InstRealPartial(const VectorRef &args) {
   const size_t args_size = 1;
   if (args.size() < args_size) {
     MS_LOG(ERROR) << __FUNCTION__ << " requires " << args_size << " or more parameters, while the input size is "
@@ -304,10 +317,18 @@ void FinalVM::InstPartial(const VectorRef &args) {
   auto fn = utils::cast<int>(Ref(fn_));
   MS_LOG(DEBUG) << "Partial argssize:" << args.size();
   std::vector<BaseRef> outs(args.size() - 1);
-
   (void)std::transform(args.begin() + 1, args.end(), outs.begin(),
                        [&, this](const BaseRef &a) { return Ref(utils::cast<int>(a)); });
   Push(std::make_shared<StructPartial>(fn, VectorRef(outs)));
+}
+
+void FinalVM::InstPartial(const VectorRef &args) {
+  MS_LOG(DEBUG) << "Start";
+  if (backend_->is_multi_graph_sink()) {
+    InstSimuPartial(args);
+  } else {
+    InstRealPartial(args);
+  }
   MS_LOG(DEBUG) << "End";
 }
 
@@ -328,43 +349,57 @@ void FinalVM::InstSimuSwitch(const VectorRef &args) {
   bool bool_value = cond;
   SwitchCondStatus cond_stat = backend_->SetSimuCond(c, bool_value);
 
-  int cond_flag = backend_->GetSimuCondFlag(c);
-  MS_LOG(DEBUG) << "Simu switch cond:" << cond << ", " << cond_flag << ", " << c.cast<AnfNodePtr>()->DebugString();
-  if (cond_flag == 2) {
-    Popp();
-    Popp();
-    backend_->SetSimuCondFlag(c, 0);
-    return;
-  }
-
   if (cond_stat == kCondAlreadyRun) {
     MS_LOG(DEBUG) << "switch alreay run bool while true jmp";
-    if (cond_flag == 0) {
-      MS_LOG(DEBUG) << "switch second run bool while true jmp";
-      backend_->SetSwitchActive(c, true);
-      Push(std::make_shared<StructSimuSwitch>(Ref(vtrue), c));
-      Pushsp();
-      backend_->SetSimuCondFlag(c, 1);
-      return;
-    } else if (cond_flag == 1) {
-      MS_LOG(DEBUG) << "switch first run bool while if jmp";
-      Push(std::make_shared<StructSimuSwitch>(Ref(vfalse), c));
-      (void)backend_->SetSimuCond(c, false);
-      backend_->SetSimuCondFlag(c, 2);
-      return;
-    } else {
-      MS_LOG(EXCEPTION) << "error cond not find";
-      return;
+    BaseRef jmp = Ref(vtrue);
+    if (utils::isa<StructPartial>(jmp)) {
+      auto new_jmp = utils::cast<std::shared_ptr<StructPartial>>(jmp);
+      backend_->RecallGraphInput(new_jmp->fg_, new_jmp->args_, c);
     }
+    cond_jmp_[c] = Ref(vfalse);
+    Push(static_cast<int>(cond_stat));
+    Popp();
+    backend_->SetSwitchActive(c, bool_value);
+    return;
   }
   if (bool_value) {
     Push(std::make_shared<StructSimuSwitch>(Ref(vtrue), c));
     Pushsp();
   } else {
+    MergeJmpArgs(Ref(vfalse), c);
     Push(std::make_shared<StructSimuSwitch>(Ref(vfalse), c));
   }
 }
 
+void FinalVM::MergeJmpArgs(const BaseRef &jmp, const BaseRef &c) {
+  auto iter = cond_jmp_.find(c);
+  if (iter == cond_jmp_.end()) {
+    return;
+  }
+  auto old_jmp = utils::cast<std::shared_ptr<StructPartial>>(iter->second);
+  auto new_jmp = utils::cast<std::shared_ptr<StructPartial>>(jmp);
+  auto &old_args = old_jmp->args_;
+  auto &new_args = new_jmp->args_;
+  for (size_t i = 0; i < new_args.size(); ++i) {
+    auto &old_arg = old_args[i];
+    auto &new_arg = new_args[i];
+    if (utils::isa<VectorRef>(old_arg)) {
+      auto old_vec_ref = utils::cast<VectorRef>(old_arg);
+      if (utils::isa<VectorRef>(new_arg)) {
+        auto new_vec_ref = utils::cast<VectorRef>(new_arg);
+        std::copy(new_vec_ref.begin(), new_vec_ref.end(), std::back_inserter(old_vec_ref));
+      }
+      new_arg = old_vec_ref;
+    } else if (utils::isa<VectorRef>(new_arg)) {
+      auto new_vec_ref = utils::cast<VectorRef>(new_arg);
+      new_vec_ref.push_back(old_arg);
+      new_arg = new_vec_ref;
+    } else {
+      new_arg = VectorRef({new_arg, old_arg});
+    }
+  }
+}
+
 void FinalVM::InstRealSwitch(const VectorRef &args) {
   const size_t args_size = 3;
   if (args.size() != args_size) {
@@ -399,6 +434,7 @@ void FinalVM::InstSwitch(const VectorRef &args) {
   } else {
     InstRealSwitch(args);
   }
+  MS_LOG(DEBUG) << "End";
 }
 
 void FinalVM::InstTuple(const VectorRef &args) {

mindspore/ccsrc/vm/vm.h

@@ -27,6 +27,9 @@
 #include <utility>
 #include <vector>
 #include <deque>
+#include <unordered_map>
+
+#include "ir/anf.h"
 #include "utils/base_ref.h"
 
 namespace mindspore {
@@ -60,13 +63,14 @@ const std::vector<std::string> inst_str{"call",  "tail_call", "return", "partial
 class StructPartial : public Base {
  public:
   // Initialize StructPartial.
-  StructPartial(int fn, const VectorRef &args);
+  StructPartial(int fn, const VectorRef &args, const FuncGraphPtr &fg = nullptr);
 
   virtual ~StructPartial() = default;
   MS_DECLARE_PARENT(StructPartial, Base)
 
   int fn_;
   VectorRef args_;
+  FuncGraphPtr fg_;
 };
 
 std::ostream &operator<<(std::ostream &os, const StructPartial &other);
@@ -98,6 +102,8 @@ class FinalVM {
   void InstTailCall(const VectorRef &args);
   void InstReturn(const VectorRef &args);
   void InstPartial(const VectorRef &args);
+  void InstSimuPartial(const VectorRef &args);
+  void InstRealPartial(const VectorRef &args);
   void InstSwitch(const VectorRef &args);
   void InstSimuSwitch(const VectorRef &args);
   void InstRealSwitch(const VectorRef &args);
@@ -120,6 +126,7 @@ class FinalVM {
   void Pushsp();
   void Popsp();
   void DoJmp(const BaseRef &jmp);
+  void MergeJmpArgs(const BaseRef &jmp, const BaseRef &c);
 
  private:
   InstSet insts_;
@@ -128,6 +135,7 @@ class FinalVM {
   std::stack<int> retsp_;
   int pc_;
   int sp_;
+  std::unordered_map<BaseRef, BaseRef, BaseRefHash> cond_jmp_;
   BackendPtr backend_;
   const InstFunctionMap inst_function_map = {
     {Instruction::kCall, [this](const VectorRef &args) { InstCall(args); }},

tests/st/control/test_multigraph_sink.py

@@ -0,0 +1,184 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+""" test_multigraph_sink """
+import pytest
+import numpy as np
+import mindspore.nn as nn
+import mindspore.context as context
+from mindspore.common.tensor import Tensor
+from mindspore.common import dtype as mstype
+from mindspore.common import ms_function
+from mindspore.ops import operations as P
+
+
+def setup_module(module):
+    context.set_context(mode = context.PYNATIVE_MODE, save_graphs = True, device_target = "Ascend")
+    context.set_context(enable_task_sink = True, device_id = 0)
+
+
+c1 = Tensor([2], mstype.int32)
+c2 = Tensor([14], mstype.int32)
+c3 = Tensor([1], mstype.int32)
+c4 = Tensor([0], mstype.int32)
+c5 = Tensor([14], mstype.int32)
+
+
+@ms_function
+def simple_if(x, y, z):
+    if x < y:
+        x = x + 1
+    else:
+        x = x + 2
+    x = x + 3
+    return x
+
+
+@ms_function
+def if_by_if(x, y, z):
+    if x < y:
+        x = x + 1
+    if y > x:
+        x = x + 2
+    x = x + 3
+    return x
+
+
+@ms_function
+def if_in_if(x, y, z):
+    out = c4
+    if x < y:
+        z = c4 + c4
+        if z < y:
+            z = z + 2
+            out = out + z
+        x = x + 3
+    out = out + x
+    return out
+
+
+@ms_function
+def simple_while(x, y, z):
+    y = y + 4
+    while x < y:
+        x = x + 1
+    x = x + 3
+    return x
+
+
+@ms_function
+def while_by_while(x, y, z):
+    while x < y:
+        x = x + 1
+    while z < c5:
+        z = z + 1
+        x = x + 1
+    x = x + 1
+    return x
+
+
+@ms_function
+def while_in_while(x, y, z):
+    out = c4
+    while x < y:
+        z = c4 + c4
+        while z < y:
+            z = z + 1
+            out = out + z
+        x = x + 1
+    out = out + x
+    return out
+
+
+@ms_function
+def while_by_while_in_while(x, y, z):
+    out = c4
+    while x < c2:
+        y = c4 + c4
+        while y < c2:
+            y = y + 1
+        out = out + y
+        z = c4 + c4
+        while z < c2:
+            z = z + 1
+        out = out + z
+        x = x + 1
+    out = out + x
+    return out
+
+
+@ms_function
+def while_in_while_in_while(x, y, z):
+    out = c4
+    while x < c2:
+        y = c4 + c4
+        while y < c2:
+            y = y + 1
+            z = c4 + c4
+            while z < c2:
+                z = z + 1
+            out = out + z
+        out = out + y
+        x = x + 1
+    out = out + x
+    return out
+
+
+def test_simple_if():
+    output = simple_if(c1, c2, c3)
+    expect = Tensor([6], mstype.int32)
+    assert output == expect
+
+
+def test_if_by_if():
+    output = if_by_if(c1, c2, c3)
+    expect = Tensor([8], mstype.int32)
+    assert output == expect
+
+
+def test_if_in_if():
+    output = if_in_if(c1, c2, c3)
+    expect = Tensor([7], mstype.int32)
+    assert output == expect
+
+
+def test_simple_while():
+    output = simple_while(c1, c2, c3)
+    expect = Tensor([21], mstype.int32)
+    assert output == expect
+
+
+def test_while_by_while():
+    output = while_by_while(c1, c2, c3)
+    expect = Tensor([28], mstype.int32)
+    assert output == expect
+
+
+def test_while_in_while():
+    output = while_in_while(c1, c2, c3)
+    expect = Tensor([1274], mstype.int32)
+    assert output == expect
+
+
+def test_while_by_while_in_while():
+    output = while_by_while_in_while(c1, c2, c3)
+    expect = Tensor([350], mstype.int32)
+    assert output == expect
+
+
+def test_while_in_while_in_while():
+    output = while_in_while_in_while(c1, c2, c3)
+    expect = Tensor([2534], mstype.int32)
+    assert output == expect
+

tests/ut/python/pynative_mode/test_multigraph_sink.py

@@ -0,0 +1,119 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+""" test_multigraph_sink """
+import pytest
+import numpy as np
+import mindspore.nn as nn
+import mindspore.context as context
+from mindspore.common.tensor import Tensor
+from mindspore.common import dtype as mstype
+from mindspore.common import ms_function
+from mindspore.ops import operations as P
+
+
+def setup_module(module):
+    context.set_context(mode = context.PYNATIVE_MODE, save_graphs = True, device_target = "Ascend")
+    context.set_context(enable_task_sink = True, device_id = 0)
+
+
+c1 = Tensor([2], mstype.int32)
+c2 = Tensor([14], mstype.int32)
+c3 = Tensor([1], mstype.int32)
+c4 = Tensor([0], mstype.int32)
+c5 = Tensor([14], mstype.int32)
+
+
+@ms_function
+def simple_if(x, y, z):
+    if x < y:
+        x = x + 1
+    else:
+        x = x + 2
+    x = x + 3
+    return x
+
+
+@ms_function
+def if_by_if(x, y, z):
+    if x < y:
+        x = x + 1
+    if y > x:
+        x = x + 2
+    x = x + 3
+    return x
+
+
+@ms_function
+def if_in_if(x, y, z):
+    out = c4
+    if x < y:
+        z = c4 + c4
+        if z < y:
+            z = z + 2
+            out = out + z
+        x = x + 3
+    out = out + x
+    return out
+
+
+@ms_function
+def simple_while(x, y, z):
+    y = y + 4
+    while x < y:
+        x = x + 1
+    x = x + 3
+    return x
+
+
+@ms_function
+def while_by_while(x, y, z):
+    while x < y:
+        x = x + 1
+    while z < c5:
+        z = z + 1
+        x = x + 1
+    x = x + 1
+    return x
+
+
+def test_simple_if():
+    output = simple_if(c1, c2, c3)
+    expect = Tensor([6], mstype.int32)
+    assert output == expect
+
+
+def test_if_by_if():
+    output = if_by_if(c1, c2, c3)
+    expect = Tensor([8], mstype.int32)
+    assert output == expect
+
+
+def test_if_in_if():
+    output = if_in_if(c1, c2, c3)
+    expect = Tensor([7], mstype.int32)
+    assert output == expect
+
+
+def test_simple_while():
+    output = simple_while(c1, c2, c3)
+    expect = Tensor([21], mstype.int32)
+    assert output == expect
+
+
+def test_while_by_while():
+    output = while_by_while(c1, c2, c3)
+    expect = Tensor([28], mstype.int32)
+    assert output == expect
+