mindspore2022/mindspore/ccsrc/optimizer/irpass/branch_culling.cc

/**
 * 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.
 */

#include "optimizer/irpass/branch_culling.h"

#include <memory>
#include <utility>
#include <unordered_map>

#include "ir/func_graph.h"
#include "ir/func_graph_cloner.h"
#include "operator/ops.h"

namespace mindspore {
namespace opt {
namespace irpass {
namespace internal {
AnfNodePtr GenerateSwitchNode(const FuncGraphPtr &graph, const AnfNodePtr &cond, const AnfNodePtr &data,
                              int switch_idx) {
  auto switch_node = prim::GetPythonOps("geswitch", "mindspore.ops.functional")->cast<PrimitivePtr>();
  std::vector<AnfNodePtr> switch_nodes{NewValueNode(switch_node), data, cond};
  auto switch_apply = graph->NewCNode(switch_nodes);
  std::vector<AnfNodePtr> tuple_getitem_nodes{NewValueNode(prim::kPrimTupleGetItem), switch_apply,
                                              NewValueNode(MakeValue(switch_idx))};
  return graph->NewCNode(tuple_getitem_nodes);
}

AnfNodePtr GenerateSwitchTrueNode(const FuncGraphPtr &graph, const AnfNodePtr &cond, const AnfNodePtr &data) {
  return GenerateSwitchNode(graph, cond, data, 1);
}

AnfNodePtr GenerateSwitchFalseNode(const FuncGraphPtr &graph, const AnfNodePtr &cond, const AnfNodePtr &data) {
  return GenerateSwitchNode(graph, cond, data, 0);
}

bool InConvertWhiteList(const AnfNodePtr &node, size_t index) {
  // The CNode inputs of the following Primitive with index in std::vector<size_t> should not be guarded by geswitch
  // node because it is attribute or ge specific reason.
  // Example : when convert CNode(kPrimReduceSum, x, axis), node of index 2 in CNode->inputs is axis which should not be
  // converted to switch guarded.
  std::vector<std::pair<PrimitivePtr, std::vector<size_t>>> white_list(
    {{prim::kPrimApplyMomentum, {1, 2}}, {prim::kPrimMomentum, {2, 3}},     {prim::kPrimStateSetItem, {1}},
     {prim::kPrimEnvGetItem, {1}},       {prim::kPrimEnvSetItem, {1}},      {prim::kPrimReduceSum, {2}},
     {prim::kPrimReduceMean, {2}},       {prim::kPrimReduceAll, {2}},       {prim::kPrimCast, {2}},
     {prim::kPrimTranspose, {2}},        {prim::kPrimOneHot, {2}},          {prim::kPrimGatherV2, {3}},
     {prim::kPrimReshape, {2}},          {prim::kPrimAssign, {1}},          {prim::kPrimAssignAdd, {1}},
     {prim::kPrimAssignSub, {1}},        {prim::kPrimTensorSummary, {1}},   {prim::kPrimImageSummary, {1}},
     {prim::kPrimScalarSummary, {1}},    {prim::kPrimHistogramSummary, {1}}});
  for (auto &item : white_list) {
    auto matched = std::any_of(item.second.begin(), item.second.end(), [&item, &node, &index](size_t idx) {
      return IsPrimitiveCNode(node, item.first) && idx == index;
    });
    if (matched) {
      return true;
    }
  }

  std::vector<PrimitivePtr> adapter_convert_ops = {prim::kPrimDepend, prim::kPrimControlDepend};
  for (auto &item : adapter_convert_ops) {
    if (IsPrimitiveCNode(node, item)) {
      return true;
    }
  }
  return false;
}

using NodeInputReplMap = std::unordered_map<std::pair<AnfNodePtr, size_t>, AnfNodePtr, PairHasher>;
// replace the nodes which should be changed
void RunSwitchNodeReplace(const FuncGraphManagerPtr &manager, std::vector<std::pair<CNodePtr, CNodePtr>> nodes_changed,
                          std::unordered_map<AnfNodePtr, AnfNodePtr> repl_node, NodeInputReplMap repl_node_inputs) {
  for (auto &node_pair : nodes_changed) {
    CNodePtr old_node = node_pair.first;
    CNodePtr new_node = node_pair.second;
    MS_EXCEPTION_IF_NULL(old_node);
    MS_EXCEPTION_IF_NULL(new_node);
    for (size_t i = 0; i < old_node->size(); i++) {
      auto input = old_node->input(i);
      if (repl_node.count(input) != 0) {
        new_node->add_input(repl_node[input]);
      } else if (repl_node_inputs.count(std::pair<AnfNodePtr, size_t>(old_node, i)) != 0) {
        new_node->add_input(repl_node_inputs[std::pair<AnfNodePtr, size_t>(old_node, i)]);
      } else {
        new_node->add_input(input);
      }
    }
  }

  for (auto &item : repl_node) {
    if (!manager->Replace(item.first, item.second)) {
      MS_LOG(EXCEPTION) << "TransformGraphDependNode replace node failed original:" << item.first->DebugString()
                        << " to new: " << item.second->DebugString();
    }
  }
}

// trace the node that should add switch and replace them with new nodes in the graph
FuncGraphPtr TransformGraphCondBranchNodes(
  const FuncGraphPtr &graph, const AnfNodePtr &cond,
  const std::function<AnfNodePtr(FuncGraphPtr graph, AnfNodePtr cond, AnfNodePtr data)> &generate_func) {
  auto manager = graph->manager();
  MS_EXCEPTION_IF_NULL(manager);

  // record the node that has been changed
  std::vector<std::pair<CNodePtr, CNodePtr>> nodes_changed;
  // record the node to be replaced
  std::unordered_map<AnfNodePtr, AnfNodePtr> repl_node;
  // record the node input to be replaced
  NodeInputReplMap repl_node_inputs;
  const AnfNodeSet &nodes = manager->nodes()[graph];
  for (auto &node : nodes) {
    MS_EXCEPTION_IF_NULL(node);
    if (!node->isa<CNode>()) {
      continue;
    }
    auto inputs = node->cast<CNodePtr>()->inputs();
    bool should_replace = false;
    // if the apply input does not belong to graph, insert a switch node
    for (size_t index = 0; index < inputs.size(); index++) {
      auto input_node = inputs[index];
      MS_EXCEPTION_IF_NULL(input_node);
      // for some ops input should not guard it with switch
      if (InConvertWhiteList(node, index)) {
        continue;
      }

      // If the input for node is not the graph belonged, or it is an ValueNode.
      // Bypass the Primitive node which is inputs[0].
      if ((index >= 1 && inputs[index]->func_graph() != nullptr && inputs[index]->func_graph() != graph) ||
          ((index >= 1 && inputs[index]->isa<ValueNode>()))) {
        input_node = generate_func(graph, cond, inputs[index]);
        repl_node_inputs[std::pair<AnfNodePtr, size_t>(node, index)] = input_node;
        should_replace = true;
      }
      if (input_node == nullptr) {
        MS_LOG(EXCEPTION) << "generate switch node failed";
      }
    }
    if (should_replace) {
      auto new_node = graph->NewCNode();
      repl_node[node] = new_node;
      nodes_changed.emplace_back(node->cast<CNodePtr>(), new_node);
    }
  }
  RunSwitchNodeReplace(manager, nodes_changed, repl_node, repl_node_inputs);
  return graph;
}

struct SharedOp {
  tensor::TensorPtr const_data;
  CNodePtr square_ops[2];
  CNodePtr merge_ops[2];
} MergeNetOutput;

inline tensor::TensorPtr GetConstData() { return MergeNetOutput.const_data; }
inline void SetConstData(const tensor::TensorPtr &const_value) { MergeNetOutput.const_data = const_value; }

inline CNodePtr GetSquareOp(int switch_idx) { return MergeNetOutput.square_ops[switch_idx]; }
inline void SetSquareOp(int switch_idx, const CNodePtr &op) { MergeNetOutput.square_ops[switch_idx] = op; }

inline CNodePtr GetMergeOp(int switch_idx) { return MergeNetOutput.merge_ops[switch_idx]; }
inline void SetMergeOp(int switch_idx, const CNodePtr &op) { MergeNetOutput.merge_ops[switch_idx] = op; }

inline void ResetSharedOp() {
  SetConstData(nullptr);
  SetSquareOp(0, nullptr);
  SetSquareOp(1, nullptr);
  SetMergeOp(0, nullptr);
  SetMergeOp(1, nullptr);
}

tensor::TensorPtr ConstData() {
  std::vector<int> shp = {1};
  tensor::TensorPtr const_data = std::make_shared<tensor::Tensor>(kInt32->type_id(), shp);
  auto *val = static_cast<int32_t *>(const_data->data_c(true));
  *val = 0;
  return const_data;
}

CNodePtr SquareOp(const FuncGraphPtr &graph, const AnfNodePtr &cond, int switch_idx,
                  const tensor::TensorPtr &const_data) {
  auto PrimSquare = prim::GetPythonOps("square", "mindspore.ops.functional")->cast<PrimitivePtr>();
  // for the depended node , add two const data to merge the flow ,one for depended node with same switch,
  // the other use the opposite
  auto ctrl_data = NewValueNode(const_data);
  auto ctrl_node = GenerateSwitchNode(graph, cond, ctrl_data, switch_idx);

  std::vector<AnfNodePtr> square_nodes{NewValueNode(PrimSquare), ctrl_node};
  auto square_op = graph->NewCNode(square_nodes);

  return square_op;
}

CNodePtr MergeNode(const FuncGraphPtr &graph, const AnfNodePtr &cond, int switch_idx,
                   const tensor::TensorPtr &const_data, const CNodePtr &square_op) {
  // for the depended node , add two const data to merge the flow ,one for depended node with same switch,
  // the other use the opposite
  auto oppsite_ctrl_data = NewValueNode(const_data);
  auto opposite_ctrl_node = GenerateSwitchNode(graph, cond, oppsite_ctrl_data, 1 - switch_idx);

  std::vector<AnfNodePtr> merge_nodes;
  auto PrimMerge = prim::GetPythonOps("merge", "mindspore.ops.functional")->cast<PrimitivePtr>();
  merge_nodes.push_back(NewValueNode(PrimMerge));
  std::vector<AnfNodePtr> make_tuple_nodes{NewValueNode(prim::kPrimMakeTuple), square_op, opposite_ctrl_node};
  merge_nodes.push_back(graph->NewCNode(make_tuple_nodes));
  auto merge_op = graph->NewCNode(merge_nodes);

  return merge_op;
}

// construct a depend node with merge output node, merge(square_op(switch(ctrl_data)), switch(opposite_ctrl_data))
// control_depend(output_node, square_op)
AnfNodePtr GenerateSwitchDependNode(const FuncGraphPtr &graph, const AnfNodePtr &cond, const AnfNodePtr &output_node,
                                    int switch_idx) {
  tensor::TensorPtr const_data = GetConstData();
  if (const_data == nullptr) {
    const_data = ConstData();
    SetConstData(const_data);
  }

  CNodePtr square_op = GetSquareOp(switch_idx);
  if (square_op == nullptr) {
    square_op = SquareOp(graph, cond, switch_idx, const_data);
    SetSquareOp(switch_idx, square_op);
  }

  CNodePtr merge_op = GetMergeOp(switch_idx);
  if (merge_op == nullptr) {
    merge_op = MergeNode(graph, cond, switch_idx, const_data, square_op);
    SetMergeOp(switch_idx, merge_op);
  }

  std::vector<AnfNodePtr> control_depend_nodes{NewValueNode(prim::kPrimControlDepend), output_node, square_op};
  auto control_depend_op = graph->NewCNode(control_depend_nodes);

  std::vector<AnfNodePtr> depend_nodes{NewValueNode(prim::kPrimDepend), merge_op, control_depend_op};
  auto depend_op = graph->NewCNode(depend_nodes);

  return depend_op;
}

// construct a merge output and add dependency with the netoutput node from control_depend
// we need to reserve the control_depend node, besides the generated merge node and control_depend node
CNodePtr GenerateSwitchControlDependNode(const FuncGraphPtr &graph, const AnfNodePtr &cond,
                                         const AnfNodePtr &ctrl_dep_node, const AnfNodePtr &ctrl_depend_dst,
                                         int switch_idx) {
  auto PrimMerge = prim::GetPythonOps("merge", "mindspore.ops.functional")->cast<PrimitivePtr>();
  auto PrimSquare = prim::GetPythonOps("square", "mindspore.ops.functional")->cast<PrimitivePtr>();
  std::vector<int> shp = {1};
  tensor::TensorPtr const_data = std::make_shared<tensor::Tensor>(kInt32->type_id(), shp);
  auto *val = static_cast<int32_t *>(const_data->data_c(true));
  *val = 0;
  // for the control_depend netoutput node , add two const data to merge the flow ,one for depended node with same
  // switch the other use the opposite
  auto ctrl_data = NewValueNode(const_data);
  auto oppsite_ctrl_data = NewValueNode(const_data);
  auto ctrl_node = GenerateSwitchNode(graph, cond, ctrl_data, switch_idx);
  auto opposite_ctrl_node = GenerateSwitchNode(graph, cond, oppsite_ctrl_data, 1 - switch_idx);

  std::vector<AnfNodePtr> square_nodes{NewValueNode(PrimSquare), ctrl_node};
  auto square_op = graph->NewCNode(square_nodes);

  std::vector<AnfNodePtr> merge_nodes;
  merge_nodes.push_back(NewValueNode(PrimMerge));
  std::vector<AnfNodePtr> make_tuple_nodes{NewValueNode(prim::kPrimMakeTuple), square_op, opposite_ctrl_node};
  merge_nodes.push_back(graph->NewCNode(make_tuple_nodes));
  auto merge_output = graph->NewCNode(merge_nodes);

  std::vector<AnfNodePtr> control_depend_nodes{NewValueNode(prim::kPrimControlDepend), ctrl_depend_dst, square_op};
  auto cond_dep_output = graph->NewCNode(control_depend_nodes);

  std::vector<AnfNodePtr> depended_make_tuple_nodes{NewValueNode(prim::kPrimMakeTuple), ctrl_dep_node, merge_output,
                                                    cond_dep_output};
  return graph->NewCNode(depended_make_tuple_nodes);
}

// generate switch nodes for true graph node inputs
AnfNodePtr GenerateSwitchDependTrueNode(const FuncGraphPtr &graph, const AnfNodePtr &cond, const AnfNodePtr &data) {
  // for switch op ,the output is a tuple ,0-th is false_branch, 1-th is true branch
  return GenerateSwitchDependNode(graph, cond, data, 1);
}

// generate switch nodes for false graph node inputs
AnfNodePtr GenerateSwitchDependFalseNode(const FuncGraphPtr &graph, const AnfNodePtr &cond, const AnfNodePtr &data) {
  // for switch op ,the output is a tuple ,0-th is false_branch, 1-th is true branch
  return GenerateSwitchDependNode(graph, cond, data, 0);
}

// generate switch nodes for true graph node inputs
CNodePtr GenerateSwitchControlDependTrueNode(const FuncGraphPtr &graph, const AnfNodePtr &cond,
                                             const AnfNodePtr &con_input, const AnfNodePtr &output) {
  // for switch op ,the output is a tuple ,0-th is false_branch, 1-th is true branch
  return GenerateSwitchControlDependNode(graph, cond, con_input, output, 1);
}

// generate switch nodes for false graph node inputs
CNodePtr GenerateSwitchControlDependFalseNode(const FuncGraphPtr &graph, const AnfNodePtr &cond,
                                              const AnfNodePtr &con_input, const AnfNodePtr &output) {
  // for switch op ,the output is a tuple ,0-th is false_branch, 1-th is true branch
  return GenerateSwitchControlDependNode(graph, cond, con_input, output, 0);
}

// to judge if the node used in ControlDepend is a net output node
bool IsNetOutputNode(const FuncGraphManagerPtr &manager, const AnfNodePtr &node) {
  auto uses = manager->node_users()[node];
  bool is_output_node = true;
  for (auto &item : uses) {
    if (IsPrimitiveCNode(item.first, prim::kPrimControlDepend) || IsPrimitiveCNode(item.first, prim::kPrimDepend)) {
      continue;
    }
    is_output_node = false;
    break;
  }
  return is_output_node;
}

// generate node for Depended MakeTuple
void GenerateReplNodeForDependMakeTuple(
  const AnfNodePtr &depended_node, const FuncGraphPtr &graph, const AnfNodePtr &cond,
  const std::shared_ptr<std::unordered_map<AnfNodePtr, AnfNodePtr>> &repl_node,
  const std::function<AnfNodePtr(FuncGraphPtr graph, AnfNodePtr cond, AnfNodePtr data)> &generate_func,
  const std::function<CNodePtr(FuncGraphPtr, AnfNodePtr, AnfNodePtr, AnfNodePtr)> &gen_ctl_depd_func) {
  MS_EXCEPTION_IF_NULL(graph->manager());

  auto make_tuple_inputs = depended_node->cast<CNodePtr>()->inputs();
  const size_t make_tuple_begin_idx = 1;
  std::vector<AnfNodePtr> new_make_tuple_nodes;
  bool replace_make_tuple = false;
  new_make_tuple_nodes.push_back(NewValueNode(prim::kPrimMakeTuple));
  for (size_t idx = make_tuple_begin_idx; idx < make_tuple_inputs.size(); idx++) {
    auto depended_tuple_input_node = make_tuple_inputs[idx];
    if (IsPrimitiveCNode(depended_tuple_input_node->cast<CNodePtr>(), prim::kPrimDepend)) {
      new_make_tuple_nodes.push_back(depended_tuple_input_node);
      continue;
    }
    if (IsPrimitiveCNode(depended_tuple_input_node->cast<CNodePtr>(), prim::kPrimControlDepend)) {
      // only when the control depend input is not square op (the op to use as merge output)
      auto control_inputs = depended_tuple_input_node->cast<CNodePtr>()->inputs();
      if (control_inputs.size() != 3) {
        MS_LOG(EXCEPTION) << "controldepend input size != 3, got " << control_inputs.size();
      }
      // control inputs: primitive, src, dst
      auto dst_node = control_inputs[2];
      if (!IsPrimitiveCNode(dst_node, prim::kPrimSquare) && IsNetOutputNode(graph->manager(), dst_node)) {
        auto gen_node = gen_ctl_depd_func(graph, cond, make_tuple_inputs[idx], dst_node);
        MS_EXCEPTION_IF_NULL(gen_node);
        auto tuple_inputs = gen_node->inputs();
        // add depended tuple inputs to new_make_tuple directly
        for (size_t i = 1; i < tuple_inputs.size(); i++) {
          new_make_tuple_nodes.push_back(tuple_inputs[i]);
        }
      }
      replace_make_tuple = true;
      continue;
    }

    if (graph->manager()->node_users()[depended_tuple_input_node].size() == 1) {
      auto gen_node = generate_func(graph, cond, depended_tuple_input_node);
      new_make_tuple_nodes.push_back(gen_node);
      replace_make_tuple = true;
      continue;
    }

    MS_LOG(WARNING) << "depended node being used by others, ";
  }
  if (replace_make_tuple) {
    auto make_tuple_op = graph->NewCNode(new_make_tuple_nodes);
    (*repl_node)[depended_node] = make_tuple_op;
  }
}

// generate a replace depend node for a single network output node
void GenerateRepDepend(
  const CNodePtr &node, const FuncGraphPtr &graph, const AnfNodePtr &cond,
  const std::shared_ptr<std::unordered_map<AnfNodePtr, AnfNodePtr>> &repl_node,
  const std::function<AnfNodePtr(FuncGraphPtr graph, AnfNodePtr cond, AnfNodePtr data)> &generate_func,
  const std::function<CNodePtr(FuncGraphPtr, AnfNodePtr, AnfNodePtr, AnfNodePtr)> &gen_ctl_depd_func) {
  auto inputs = node->inputs();
  if (inputs.size() != 3) {
    MS_LOG(EXCEPTION) << "Inputs should be [depend, actual_value, depended_node].";
  }

  std::vector<AnfNodePtr> new_depened_inputs;
  // Inputs should be [depend, actual_value, depended_node]
  auto depended_node = inputs[2];
  new_depened_inputs.push_back(inputs[0]);
  new_depened_inputs.push_back(inputs[1]);
  // depended node should be make_tuple or a single depended node
  if (IsPrimitiveCNode(depended_node, prim::kPrimMakeTuple)) {
    GenerateReplNodeForDependMakeTuple(depended_node, graph, cond, repl_node, generate_func, gen_ctl_depd_func);
  } else if (IsPrimitiveCNode(depended_node, prim::kPrimControlDepend)) {
    // only when the control depend input is not square op (the op to use as merge output)
    auto control_inputs = depended_node->cast<CNodePtr>()->inputs();
    // control inputs: primitive, src, dst
    if (control_inputs.size() != 3) {
      MS_LOG(EXCEPTION) << "controldepend input size != 3, got " << control_inputs.size();
    }
    auto dst_node = control_inputs[2];
    if (!IsPrimitiveCNode(dst_node, prim::kPrimSquare) && IsNetOutputNode(graph->manager(), dst_node)) {
      auto gen_node = gen_ctl_depd_func(graph, cond, depended_node, dst_node);
      (*repl_node)[depended_node] = gen_node;
    }
  } else {
    // Check if there is only single user for depend_node.
    if (graph->manager()->node_users()[depended_node].size() == 1) {
      auto gen_node = generate_func(graph, cond, depended_node);
      (*repl_node)[depended_node] = gen_node;
    } else {
      MS_LOG(WARNING) << "depended node being used by others";
    }
  }
}

// generate depend node for netoutput node, to resolve the stream synchronize problem of ge
// traverse all nodes of depend node, find the graph output node , generaete a merge node of (square, const)
// and add control_depend of graph output node and square node.
FuncGraphPtr TransformGraphDependNode(
  const FuncGraphPtr &graph, const AnfNodePtr &cond,
  const std::function<AnfNodePtr(FuncGraphPtr graph, AnfNodePtr cond, AnfNodePtr data)> &gen_depend_func,
  const std::function<CNodePtr(FuncGraphPtr, AnfNodePtr, AnfNodePtr, AnfNodePtr)> &gen_ctl_depd_func) {
  auto manager = graph->manager();
  MS_EXCEPTION_IF_NULL(manager);

  ResetSharedOp();
  std::shared_ptr<std::unordered_map<AnfNodePtr, AnfNodePtr>> repl_node =
    std::make_shared<std::unordered_map<AnfNodePtr, AnfNodePtr>>();  // record the node to be replaced
  const AnfNodeSet &nodes = manager->nodes()[graph];
  for (auto &node : nodes) {
    MS_EXCEPTION_IF_NULL(node);
    if (!node->isa<CNode>()) {
      continue;
    }
    if (IsPrimitiveCNode(node, prim::kPrimDepend)) {
      auto cnode = node->cast<CNodePtr>();
      if (cnode->size() != 3) {
        MS_LOG(EXCEPTION) << "Dependnode input size != 3";
      }
      auto depended_node = cnode->input(2);
      MS_EXCEPTION_IF_NULL(depended_node);
      if (!depended_node->isa<CNode>()) {
        continue;
      }
      if (IsPrimitiveCNode(depended_node, prim::kPrimDepend)) {
        continue;
      }
      GenerateRepDepend(cnode, graph, cond, repl_node, gen_depend_func, gen_ctl_depd_func);
    }
  }
  ResetSharedOp();

  for (auto &item : *repl_node) {
    if (!manager->Replace(item.first, item.second)) {
      MS_LOG(EXCEPTION) << "TransformGraphDependNode replace node failed";
    }
  }

  return graph;
}

FuncGraphPtr TransformGraphCondTrueBranchNodes(const FuncGraphPtr &graph, const AnfNodePtr &cond) {
  (void)TransformGraphCondBranchNodes(graph, cond, GenerateSwitchTrueNode);
  return TransformGraphDependNode(graph, cond, GenerateSwitchDependTrueNode, GenerateSwitchControlDependTrueNode);
}

FuncGraphPtr TransformGraphCondFalseBranchNodes(const FuncGraphPtr &graph, const AnfNodePtr &cond) {
  (void)TransformGraphCondBranchNodes(graph, cond, GenerateSwitchFalseNode);
  return TransformGraphDependNode(graph, cond, GenerateSwitchDependFalseNode, GenerateSwitchControlDependFalseNode);
}

// judge if the true and false graph output is compatible(they shall have same tuple size)
bool GraphOutputCompatible(const AbstractBasePtr &true_branch_abs, const AbstractBasePtr &false_branch_abs) {
  MS_EXCEPTION_IF_NULL(true_branch_abs);
  MS_EXCEPTION_IF_NULL(false_branch_abs);

  if (true_branch_abs->isa<abstract::AbstractTuple>() && false_branch_abs->isa<abstract::AbstractTuple>()) {
    abstract::AbstractTuplePtr true_branch_tuple = true_branch_abs->cast<abstract::AbstractTuplePtr>();
    abstract::AbstractTuplePtr false_branch_tuple = false_branch_abs->cast<abstract::AbstractTuplePtr>();
    if (true_branch_tuple->elements().size() != false_branch_tuple->elements().size()) {
      MS_LOG(ERROR) << "true branch size:" << true_branch_tuple->elements().size()
                    << ", not equal to false banch size:" << false_branch_tuple->elements().size() << " ";
      return false;
    }
    bool all_compatible = true;
    for (size_t i = 0; i < true_branch_tuple->elements().size(); i++) {
      all_compatible =
        all_compatible && GraphOutputCompatible(true_branch_tuple->elements()[i], false_branch_tuple->elements()[i]);
    }
    return all_compatible;
  }
  TypePtr true_branch_type = true_branch_abs->BuildType();
  TypePtr false_branch_type = false_branch_abs->BuildType();
  MS_LOG(DEBUG) << "branch output Type equal?" << (*true_branch_type == *false_branch_type)
                << " true:" << true_branch_type->ToString() << " false:" << false_branch_type->ToString();
  return (*true_branch_type == *false_branch_type);
}

AnfNodePtr GenerateMergeNodes(const AnfNodePtr &true_output_node, const AnfNodePtr &false_output_node,
                              const AbstractBasePtr &true_graph_output_abs,
                              const AbstractBasePtr &false_graph_output_abs, const AnfNodePtr &cond) {
  MS_EXCEPTION_IF_NULL(true_graph_output_abs);
  MS_EXCEPTION_IF_NULL(false_graph_output_abs);
  MS_EXCEPTION_IF_NULL(cond);
  MS_EXCEPTION_IF_NULL(cond->func_graph());
  auto PrimMerge = prim::GetPythonOps("merge", "mindspore.ops.functional")->cast<PrimitivePtr>();
  MS_EXCEPTION_IF_NULL(PrimMerge);

  if (!true_graph_output_abs->isa<abstract::AbstractTuple>()) {
    std::vector<AnfNodePtr> merge_nodes;
    merge_nodes.push_back(NewValueNode(PrimMerge));
    std::vector<AnfNodePtr> make_tuple_nodes{NewValueNode(prim::kPrimMakeTuple), true_output_node, false_output_node};
    merge_nodes.push_back(cond->func_graph()->NewCNode(make_tuple_nodes));
    std::vector<AnfNodePtr> tuple_getitem_nodes{NewValueNode(prim::kPrimTupleGetItem),
                                                cond->func_graph()->NewCNode(merge_nodes), NewValueNode(MakeValue(0))};
    return cond->func_graph()->NewCNode(tuple_getitem_nodes);
  } else {
    abstract::AbstractTuplePtr true_branch_tuple = true_graph_output_abs->cast<abstract::AbstractTuplePtr>();
    abstract::AbstractTuplePtr false_branch_tuple = false_graph_output_abs->cast<abstract::AbstractTuplePtr>();

    std::vector<AnfNodePtr> make_tuple_nodes;
    make_tuple_nodes.push_back(NewValueNode(prim::kPrimMakeTuple));
    for (size_t i = 0; i < true_branch_tuple->elements().size(); i++) {
      std::vector<AnfNodePtr> true_getitem_nodes{NewValueNode(prim::kPrimTupleGetItem), true_output_node,
                                                 NewValueNode(MakeValue(SizeToInt(i)))};
      auto true_node = cond->func_graph()->NewCNode(true_getitem_nodes);
      std::vector<AnfNodePtr> false_getitem_nodes{NewValueNode(prim::kPrimTupleGetItem), false_output_node,
                                                  NewValueNode(MakeValue(SizeToInt(i)))};
      auto false_node = cond->func_graph()->NewCNode(false_getitem_nodes);

      auto merge_node = GenerateMergeNodes(true_node, false_node, true_branch_tuple->elements()[i],
                                           false_branch_tuple->elements()[i], cond);
      make_tuple_nodes.push_back(merge_node);
    }
    return cond->func_graph()->NewCNode(make_tuple_nodes);
  }
}

AnfNodePtr TransformMergeBranches(const AnfNodePtr &true_output_node, const AnfNodePtr &false_output_node,
                                  const AbstractBasePtr &true_graph_output_abs,
                                  const AbstractBasePtr &false_graph_output_abs, const AnfNodePtr &cond) {
  if (!GraphOutputCompatible(true_graph_output_abs, false_graph_output_abs)) {
    MS_LOG(EXCEPTION) << "Switch output branch not compatible, true:" << true_graph_output_abs->ToString()
                      << "， false:" << false_graph_output_abs->ToString();
  }
  return GenerateMergeNodes(true_output_node, false_output_node, true_graph_output_abs, false_graph_output_abs, cond);
}
}  // namespace internal
}  // namespace irpass
}  // namespace opt
}  // namespace mindspore