graph.cc 8.0 KB
Newer Older
X
Xin Pan 已提交
1
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
X
start  
Xin Pan 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

15
#include <memory>
X
Xin Pan 已提交
16

X
start  
Xin Pan 已提交
17
#include "paddle/fluid/framework/ir/graph.h"
18
#include "paddle/fluid/framework/operator.h"
X
start  
Xin Pan 已提交
19 20

namespace paddle {
X
Xin Pan 已提交
21
namespace framework {
X
Xin Pan 已提交
22
namespace ir {
X
Xin Pan 已提交
23

X
clean  
Xin Pan 已提交
24
Graph::Graph(const ProgramDesc &program) : program_(program) {
25 26 27
  auto var_nodes = InitFromProgram(program_);
  ResolveHazard(var_nodes);
}
28

29 30
std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram(
    const ProgramDesc &program) {
M
minqiyang 已提交
31
  VLOG(3) << "block in program:" << program_.Size();
32
  std::unordered_map<std::string, VarDesc *> all_vars;
33 34
  // var nodes for each var name, will have multiple versions in SSA
  std::map<std::string, std::vector<ir::Node *>> var_nodes;
35 36 37 38
  for (auto *var : program.Block(0).AllVars()) {
    all_vars.emplace(var->Name(), var);
  }

39 40
  auto not_visited_vars = all_vars;

41
  for (auto *op : program.Block(0).AllOps()) {
X
clean  
Xin Pan 已提交
42
    ir::Node *node = CreateOpNode(op);
X
Xin Pan 已提交
43 44
    // For input args, reuse the same var name if it was created before.
    // Otherwise, create a new one.
45
    for (auto &each_var_name : op->InputArgumentNames()) {
46
      not_visited_vars.erase(each_var_name);
47
      ir::Node *var = nullptr;
X
Xin Pan 已提交
48
      if (var_nodes.find(each_var_name) != var_nodes.end()) {
X
Xin Pan 已提交
49
        var = var_nodes.at(each_var_name).back();
X
Xin Pan 已提交
50
      } else if (all_vars.count(each_var_name) != 0) {
X
clean  
Xin Pan 已提交
51
        var = CreateVarNode(all_vars.at(each_var_name));
X
Xin Pan 已提交
52
        var_nodes[each_var_name].push_back(var);
53
      } else {
X
Xin Pan 已提交
54 55 56
        // Operation input var can be optional (dispensable). Which means
        // the operation doesn't really need the var at runtime. In this
        // case, the no-existed var is ready at the beginning.
X
polish  
Xin Pan 已提交
57
        var = CreateEmptyNode(each_var_name, ir::Node::Type::kVariable);
X
Xin Pan 已提交
58
        var_nodes[each_var_name].push_back(var);
59 60 61 62
      }
      node->inputs.push_back(var);
      var->outputs.push_back(node);
    }
X
Xin Pan 已提交
63
    // For output args, always create a new var.
64
    std::unordered_set<std::string> out_arg_set;
65
    for (auto &each_var_name : op->OutputArgumentNames()) {
66
      not_visited_vars.erase(each_var_name);
67
      if (each_var_name != kEmptyVarName) {
68 69 70 71 72
        PADDLE_ENFORCE_EQ(out_arg_set.count(each_var_name), 0,
                          platform::errors::InvalidArgument(
                              "The input Program is invalid. Variable %s occurs"
                              " in output of %s multiple times.",
                              each_var_name, op->Type()));
73 74 75
        out_arg_set.insert(each_var_name);
      }

X
Xin Pan 已提交
76 77 78 79 80 81 82 83 84
      ir::Node *var = nullptr;
      if (all_vars.count(each_var_name) != 0) {
        var = CreateVarNode(all_vars.at(each_var_name));
      } else {
        // Operation output vars can be @EMPTY@. For example, while_grad
        // can have multi @EMPTY@ outputs with no VarDesc.
        // TODO(panyx0718): Add a test.
        var = CreateEmptyNode(each_var_name, ir::Node::Type::kVariable);
      }
X
Xin Pan 已提交
85
      var_nodes[each_var_name].push_back(var);
86 87 88 89
      node->outputs.push_back(var);
      var->inputs.push_back(node);
    }
  }
90 91 92 93 94 95 96 97 98 99

  for (auto &pair : not_visited_vars) {
    const auto &var_name = pair.first;
    auto *var_desc = pair.second;
    if (var_name != kEmptyVarName) {
      VLOG(10) << "Create isolated var node " << var_name;
      var_nodes[var_name].push_back(CreateVarNode(var_desc));
    }
  }

X
polish  
Xin Pan 已提交
100
  Set<const std::vector<OpDesc *>>(
X
Xin Pan 已提交
101
      details::kStaleProgramOpDescs,
X
polish  
Xin Pan 已提交
102
      new std::vector<OpDesc *>(program.Block(0).AllOps()));
G
Gabor Buella 已提交
103
  return var_nodes;
104
}
X
Xin Pan 已提交
105

106 107
void Graph::ResolveHazard(
    const std::map<std::string, std::vector<ir::Node *>> &var_nodes) {
X
polish  
Xin Pan 已提交
108
  /**
109 110 111 112 113
   * We should handle write after read(WAR) and write after write(WAW) here.
   * Because some of the operators of the program can be executed parallelly.
   * So, to make the program running in the right order, we should add the
   * dependence of WAR and WAW.
   *
X
polish  
Xin Pan 已提交
114 115 116
   *
   * https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR)
   */
X
Xin Pan 已提交
117

X
Xin Pan 已提交
118 119 120 121 122 123 124 125
  for (auto &var : var_nodes) {
    auto &versions = var.second;
    if (versions.size() <= 1) continue;

    auto it_new = versions.rbegin();
    auto it_old = versions.rbegin();
    ++it_old;
    for (; it_old != versions.rend(); it_new = it_old, ++it_old) {
M
minqiyang 已提交
126
      VLOG(3) << "deal with var: " << (*it_new)->Name();
X
Xin Pan 已提交
127 128 129 130
      ir::Node *write_op =
          (*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
      const auto &read_ops = (*it_old)->outputs;

131 132 133 134
      PADDLE_ENFORCE_NOT_NULL(
          write_op, platform::errors::NotFound(
                        "The generate operator of variable %s is null.",
                        (*it_new)->Name()));
135 136 137 138

      // Add write after write dependence
      ir::Node *upstream_op =
          (*it_old)->inputs.empty() ? nullptr : (*it_old)->inputs[0];
X
Xin Pan 已提交
139 140
      // TODO(zcd): Add a test.
      if (upstream_op && upstream_op != write_op) {
141 142 143
        ir::Node *dep_var = CreateControlDepVar();
        write_op->inputs.push_back(dep_var);
        upstream_op->outputs.push_back(dep_var);
144
        VLOG(10) << "add dep_var:" << dep_var->Name();
145 146 147 148
        dep_var->outputs.push_back(write_op);
        dep_var->inputs.push_back(upstream_op);
      }

X
Xin Pan 已提交
149 150 151 152 153 154
      for (auto *read_op : read_ops) {
        // Manually add a dependency var from read_op to write_op;
        if (read_op == write_op) {
          // Read Write is the same op.
          continue;
        }
X
Xin Pan 已提交
155 156 157 158 159 160 161 162 163 164 165
        // 2 ops might have been connected via other vars.
        bool has_dep = false;
        for (ir::Node *r_out : read_op->outputs) {
          for (ir::Node *w_in : write_op->inputs) {
            if (r_out == w_in) {
              has_dep = true;
              break;
            }
          }
        }
        if (has_dep) continue;
X
Xin Pan 已提交
166

X
Xin Pan 已提交
167
        ir::Node *dep_var = CreateControlDepVar();
168
        VLOG(10) << "add dep_var:" << dep_var->Name();
X
Xin Pan 已提交
169 170 171 172 173 174 175
        read_op->outputs.push_back(dep_var);
        dep_var->inputs.push_back(read_op);
        write_op->inputs.push_back(dep_var);
        dep_var->outputs.push_back(write_op);
      }
    }
  }
X
better  
Xin Pan 已提交
176
}
X
Xin Pan 已提交
177

178 179 180 181 182 183
std::shared_ptr<Graph> Graph::Clone() {
  auto cloned_graph = std::make_shared<Graph>(this->program_);
  cloned_graph->ReleaseNodes();
  cloned_graph->num_node_created_ = 0;
  std::unordered_map<ir::Node *, ir::Node *> origin_to_cloned;
  for (auto *n : this->node_set_) {
184 185
    PADDLE_ENFORCE_NOT_NULL(n, platform::errors::InvalidArgument(
                                   "The node to be cloned is nullptr."));
186 187 188 189 190 191 192 193 194 195
    ir::Node *cloned_node = nullptr;
    if (n->IsCtrlVar()) {
      cloned_node = cloned_graph->CreateControlDepVar();
    } else if (!n->var_desc_ && !n->op_desc_) {  // empty node
      cloned_node = cloned_graph->CreateEmptyNode(n->Name(), n->NodeType());
    } else if (n->IsVar()) {
      cloned_node = cloned_graph->CreateVarNode(n->Var());
    } else if (n->IsOp()) {
      cloned_node = cloned_graph->CreateOpNode(n->Op());
    }
196 197 198 199 200
    PADDLE_ENFORCE_NOT_NULL(
        cloned_node,
        platform::errors::InvalidArgument(
            "Failed to clone new node from original node in graph."));
    origin_to_cloned[n] = cloned_node;
201 202 203 204 205 206 207 208 209 210 211 212
  }
  for (auto *n : this->node_set_) {
    for (auto it = n->inputs.begin(); it != n->inputs.end(); it++) {
      origin_to_cloned[n]->inputs.push_back(origin_to_cloned[*it]);
    }
    for (auto it = n->outputs.begin(); it != n->outputs.end(); it++) {
      origin_to_cloned[n]->outputs.push_back(origin_to_cloned[*it]);
    }
  }
  return cloned_graph;
}

X
Xin Pan 已提交
213 214 215
bool IsControlDepVar(const ir::Node &var) {
  return var.Name().find(ir::Node::kControlDepVarName) != std::string::npos;
}
X
Xin Pan 已提交
216
}  // namespace ir
X
Xin Pan 已提交
217
}  // namespace framework
X
start  
Xin Pan 已提交
218
}  // namespace paddle