build_cinn_pass.cc

/* Copyright (c) 2021 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. */

#include "paddle/fluid/framework/paddle2cinn/build_cinn_pass.h"

#include <algorithm>
#include <iterator>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>

#include "cinn/frontend/op_mapper_registry.h"
#include "cinn/frontend/op_mappers/use_op_mappers.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/node.h"
#include "paddle/fluid/framework/ir/subgraph_detector.h"
#include "paddle/fluid/framework/op_proto_maker.h"
#include "paddle/fluid/framework/paddle2cinn/cinn_compiler.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/errors.h"

namespace paddle {
namespace framework {
namespace paddle2cinn {

using framework::ir::Graph;
using framework::ir::Node;

using GraphNodeVec = std::vector<Node*>;
using GraphNodeSet = std::unordered_set<Node*>;

namespace {
int ExtractOpRole(const GraphNodeSet& cluster) {
  std::unordered_set<int> op_roles;
  std::string attr_name = OpProtoAndCheckerMaker::OpRoleAttrName();
  for (auto* n : cluster) {
    if (n->Op() && n->Op()->HasAttr(attr_name)) {
      op_roles.insert(BOOST_GET_CONST(int, n->Op()->GetAttr(attr_name)));
    }
  }
  if (op_roles.size() == 1U) {
    return *(op_roles.begin());
  } else {
    return static_cast<int>(OpRole::kNotSpecified);
  }
}

// Deal with subgraph's feed input var node:
// create a new input var node and it's feed op node
void AddFeedOpAndVar(const std::unordered_set<Node*>& feed_vars,
                     const GraphNodeSet& cluster,
                     const std::unordered_map<Node*, Node*>& old_op2new_op,
                     const std::unordered_map<Node*, Node*>& old_var2new_var,
                     Graph* graph) {
  for (auto* old_var : feed_vars) {
    // create feed op
    OpDesc desc;
    desc.SetType("feed");
    desc.SetOutput("Out", {old_var->Name()});
    auto op = graph->CreateOpNode(&desc);

    // get new feed var node
    auto* var = old_var2new_var.at(old_var);

    // link feed op and feed var
    IR_NODE_LINK_TO(op, var);

    // link feed var to cluster op
    for (auto* old_op : old_var->outputs) {
      if (cluster.count(old_op)) {
        IR_NODE_LINK_TO(var, old_op2new_op.at(old_op));
      }
      // Do not need relink old op or old var here, they will be
      // fixed in RemoveSubGraphFromGraph, here we just deal with
      // new subgraph's node.
    }
  }
}

// Deal with subgraph's parameter var node:
// create a new input var node, it's data will get by scope,
// so it don't need feed op
void AddParamVar(const std::unordered_set<Node*>& param_vars,
                 const GraphNodeSet& cluster,
                 const std::unordered_map<Node*, Node*>& old_op2new_op,
                 const std::unordered_map<Node*, Node*>& old_var2new_var,
                 Graph* graph) {
  for (auto* old_var : param_vars) {
    auto* var = old_var2new_var.at(old_var);

    for (auto* old_op : old_var->outputs) {
      if (cluster.count(old_op)) {
        IR_NODE_LINK_TO(var, old_op2new_op.at(old_op));
      }
    }
  }
}

// Deal with subgraph's outputs var node:
// create a new output var node and it's fetch op
void AddOutputVar(const std::unordered_set<Node*>& output_vars,
                  const GraphNodeSet& cluster,
                  const std::unordered_map<Node*, Node*>& old_op2new_op,
                  const std::unordered_map<Node*, Node*>& old_var2new_var,
                  Graph* graph) {
  for (auto* old_var : output_vars) {
    auto* var = old_var2new_var.at(old_var);

    for (auto* old_op : old_var->inputs) {
      if (cluster.count(old_op)) {
        IR_NODE_LINK_TO(old_op2new_op.at(old_op), var);
      }
    }
  }
}

// Create new subgraph with and op nodes are cluster nodes, and all
// var node are from internal nodes
std::unique_ptr<Graph> CreateNewSubGraph(const GraphNodeSet& cluster,
                                         const GraphNodeSet& cluster_internals,
                                         const GraphNodeSet& cluster_inputs,
                                         const GraphNodeSet& cluster_outputs) {
  // Graph's constructor must has one parameter, and in our code,
  // the ProgramDesc is useless, so here we pass a temporary object.
  auto subgraph = std::make_unique<Graph>(framework::ProgramDesc());

  std::unordered_map<Node*, Node*> old_op2new_op;
  for (auto* op : cluster) {
    auto sub_node = subgraph->CreateOpNode(op->Op());
    old_op2new_op[op] = sub_node;
  }

  std::unordered_map<Node*, Node*> old_var2new_var;
  for (auto* var : cluster_internals) {
    PADDLE_ENFORCE_NOT_NULL(var->Var(),
                            platform::errors::PreconditionNotMet(
                                "The var desc of the node in cluster_internals "
                                "shouldn't be null."));
    auto* sub_node = subgraph->CreateVarNode(var->Var());
    old_var2new_var[var] = sub_node;
  }
  for (auto* var : cluster_inputs) {
    if (var->Var()) {
      auto* sub_node = subgraph->CreateVarNode(var->Var());
      old_var2new_var[var] = sub_node;
    }
  }
  for (auto* var : cluster_outputs) {
    if (var->Var()) {
      auto* sub_node = subgraph->CreateVarNode(var->Var());
      old_var2new_var[var] = sub_node;
    }
  }

  std::unordered_set<Node*> need_feed_vars;
  std::unordered_set<Node *> param_vars, output_vars;
  // the subgraph is independently, so here we only need link
  // to the node in new subgraph, and discard the link to
  // out-graph.
  for (auto* op : cluster) {
    for (auto* var : op->inputs) {
      // one output var maybe an input of the cluster
      if (cluster_internals.count(var) ||
          (cluster_outputs.count(var) && old_var2new_var.count(var))) {
        IR_NODE_LINK_TO(old_var2new_var.at(var), old_op2new_op.at(op));
      } else if (cluster_inputs.count(var) && var->Var() != nullptr) {
        if (var->Var()->IsParameter()) {
          // Parameters have been preserved in scope, compared to feed var,
          // param just need add new var and don't need add feed op.
          // The var is used for check whether we need preserve the tensor
          // when transform paddle scope to CINN scope.
          param_vars.insert(var);
        } else {
          // When the var is subgraph input and the var is not parameter,
          // we need add a new feed op to feed the var.
          need_feed_vars.insert(var);
        }
      }
    }
    for (auto* var : op->outputs) {
      if (cluster_internals.count(var)) {
        IR_NODE_LINK_TO(old_op2new_op.at(op), old_var2new_var.at(var));
      } else if (cluster_outputs.count(var) && var->Var() != nullptr) {
        // Create new output var node to guarantee the independency of
        // subgraph. In other words, the subgraph has no connection with
        // other graph, even the input graph.
        output_vars.insert(var);
      }
    }
  }

  AddFeedOpAndVar(need_feed_vars, cluster, old_op2new_op, old_var2new_var,
                  subgraph.get());
  AddParamVar(param_vars, cluster, old_op2new_op, old_var2new_var,
              subgraph.get());
  AddOutputVar(output_vars, cluster, old_op2new_op, old_var2new_var,
               subgraph.get());

  return subgraph;
}

// This interface is used to classify all variables involved in a cluster into
// three types: inputs, outputs, and internals.
// The input node is some subgraph op's input but not any subgraph op's output.
// The output node is some subgraph op's output and some out-graph op's input.
// Specially, the internal node is a node that only used by subgraph, and
// out-graph should not using this node at all.
// cluster_inputs & cluster_outputs & cluster_internals == NULL
// cluster_outputs | cluster_internals == all graph op's outputs node
void AnalyseClusterVariables(const GraphNodeSet& cluster,
                             GraphNodeSet* cluster_inputs,
                             GraphNodeSet* cluster_outputs,
                             GraphNodeSet* cluster_internals) {
  // collecting all input and output of op
  for (auto* op_node : cluster) {
    for (auto* input_var_node : op_node->inputs) {
      cluster_inputs->insert(input_var_node);
    }
    for (auto* output_var_node : op_node->outputs) {
      cluster_outputs->insert(output_var_node);
    }
  }
  // remove output node from cluster_inputs,
  // and add cluster_internals node
  for (auto* var_node : *cluster_outputs) {
    if (cluster_inputs->count(var_node) > 0) {
      // if a input node also exists in output list, remove
      cluster_inputs->erase(var_node);

      // the internal node is must an output node of sub-graph,
      // but not any input node of out-graph.
      bool is_only_used_internal = true;
      for (auto* next_op_node : var_node->outputs) {
        is_only_used_internal &= (cluster.count(next_op_node) > 0);
      }
      if (is_only_used_internal) {
        cluster_internals->insert(var_node);
      }
    }
  }

  // if a output node also exists in internal list, remove.
  for (auto* var_node : *cluster_internals) {
    cluster_outputs->erase(var_node);
  }
}

void AddLinkToCinnOp(const GraphNodeSet& cluster_inputs,
                     const GraphNodeSet& cluster_outputs, Node* cinn_op_node) {
  // add new link from cluster_inputs to cinn_op_node
  for (auto* var_node : cluster_inputs) {
    IR_NODE_LINK_TO(var_node, cinn_op_node);
  }

  // add new link from cinn_op_node to cluster_outputs
  for (auto* var_node : cluster_outputs) {
    IR_NODE_LINK_TO(cinn_op_node, var_node);
  }
}

void AddCinnOpToGraph(const GraphNodeSet& cluster,
                      const GraphNodeSet& cluster_inputs,
                      const GraphNodeSet& cluster_outputs,
                      const std::string& compilation_key, Graph* graph) {
  // Add the cinn launch op
  framework::OpDesc cinn_op_desc;
  cinn_op_desc.SetType(kCinnLaunchOp);
  std::vector<std::string> input_names;
  std::for_each(cluster_inputs.begin(), cluster_inputs.end(),
                [&input_names](Node* n) {
                  if (n->Var() != nullptr) {
                    input_names.emplace_back(n->Name());
                  }
                });
  cinn_op_desc.SetInput("X", input_names);
  std::vector<std::string> output_names;
  std::for_each(cluster_outputs.begin(), cluster_outputs.end(),
                [&output_names](Node* n) {
                  if (n->Var() != nullptr) {
                    output_names.emplace_back(n->Name());
                  }
                });
  cinn_op_desc.SetOutput("Out", output_names);
  cinn_op_desc.SetAttr(kCompilationKey, compilation_key);
  cinn_op_desc.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
                       ExtractOpRole(cluster));
  cinn_op_desc.Flush();
  auto* cinn_op_node = graph->CreateOpNode(&cinn_op_desc);
  // Add new links from or to the the cinn launch op node
  AddLinkToCinnOp(cluster_inputs, cluster_outputs, cinn_op_node);
}

// Removing cluster node and internals node from Graph
void RemoveSubGraphFromGraph(const GraphNodeSet& cluster,
                             const GraphNodeSet& cluster_internals,
                             Graph* graph) {
  const std::unordered_set<const Node*> const_cluster{cluster.cbegin(),
                                                      cluster.cend()};
  const std::unordered_set<const Node*> const_internals{
      cluster_internals.cbegin(), cluster_internals.cend()};
  ir::GraphSafeRemoveNodes(graph, const_cluster);
  ir::GraphSafeRemoveNodes(graph, const_internals);
}

// Replacing Cinn subgraph to a cinn op node, whose op_type is
// kCinnLaunchOp, and inputs ares cluster_inputs and outputs are
// cluster_outputs.
// Meanwhile, move all links of cluster to the cinn op.
void ReplaceSubGraphWithCinnOpNode(const GraphNodeSet& cluster,
                                   const GraphNodeSet& cluster_inputs,
                                   const GraphNodeSet& cluster_outputs,
                                   const GraphNodeSet& cluster_internals,
                                   const std::string& compilation_key,
                                   Graph* graph) {
  // Add the cinn op node whose name is "kCinnLaunchOp" into graph
  AddCinnOpToGraph(cluster, cluster_inputs, cluster_outputs, compilation_key,
                   graph);
  // Remove the cinn subgraph from graph
  RemoveSubGraphFromGraph(cluster, cluster_internals, graph);
}

// Search all subgraphs which all op node supported by CINN,
// Here we using SubgraphDetector to detecte the subgraph that
// all of op node supported by CINN. We using OpMapperRegistry
// to check whether the op node supported by CINN.
void SearchAllSubgraphs(Graph* graph) {
  auto teller = [](const Node* node) {
    return ::cinn::frontend::OpMapperRegistry::Global()->Find(node->Name()) !=
           nullptr;
  };
  std::vector<GraphNodeVec> clusters =
      framework::ir::SubgraphDetector(graph, teller)();

  auto* cinn_compiler = CinnCompiler::GetInstance();
  for (const auto& node_vec : clusters) {
    // Classify var node to inputs, outputs, and internals.
    GraphNodeSet cluster_set(node_vec.begin(), node_vec.end());

    GraphNodeSet cluster_inputs, cluster_outputs, cluster_internals;
    AnalyseClusterVariables(cluster_set, &cluster_inputs, &cluster_outputs,
                            &cluster_internals);
    // Create a new subgraph according to the found cluster and
    // save it in CinnCompiler
    std::string compilation_key = cinn_compiler->AddGraph(CreateNewSubGraph(
        cluster_set, cluster_internals, cluster_inputs, cluster_outputs));
    // Replace the found cluster to a new cinn op node
    ReplaceSubGraphWithCinnOpNode(cluster_set, cluster_inputs, cluster_outputs,
                                  cluster_internals, compilation_key, graph);
  }
}
}  // namespace

void BuildCinnPass::ApplyImpl(Graph* graph) const { SearchAllSubgraphs(graph); }

}  // namespace paddle2cinn
}  // namespace framework
}  // namespace paddle

REGISTER_PASS(build_cinn_pass, paddle::framework::paddle2cinn::BuildCinnPass);