seqpool_concat_fuse_pass.cc 8.1 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
/* Copyright (c) 2018 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/ir/seqpool_concat_fuse_pass.h"
#include <string>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"

#define MAX_CONCAT_INPUTS 200

namespace paddle {
namespace framework {
namespace ir {

PDNode* BuildSeqPoolConcatPattern(PDPattern* pattern,
                                  const std::string& name_scope,
                                  int num_inputs) {
  auto is_concat_op_with_inputs = [](Node* x, int num) -> bool {
    return x && x->IsOp() && x->Op()->Type() == "concat" &&
           x->Op()->Input("X").size() == static_cast<size_t>(num);
  };

  auto is_nth_input_var_of_concat = [=](Node* x, int idx) -> bool {
    return x && x->IsVar() && VarLinksToOp(x, "concat") &&
           x->outputs.size() == 1 && IsNthInput(x, x->outputs[0], "X", idx) &&
           is_concat_op_with_inputs(x->outputs[0], num_inputs);
  };

  auto is_seqpool_op_with_pootype_of_nth_input_of_concat = [=](
      Node* x, const std::string& type, int idx) -> bool {
T
tensor-tang 已提交
42 43 44 45 46 47 48 49 50
    bool this_is_seqpool_op =
        x && x->IsOp() && x->Op()->Type() == "sequence_pool" &&
        x->Op()->HasAttr("pooltype") &&
        boost::get<std::string>(x->Op()->GetAttr("pooltype")) == type &&
        x->outputs.size() == 2;  // seqpool should only have 2 outputs
    bool satisfied_all = this_is_seqpool_op;
    if (this_is_seqpool_op) {
      // Only one output of seqpool_op is nth_input_var of concat,
      // the other one should be unused empty var.
T
tensor-tang 已提交
51
      if (is_nth_input_var_of_concat(x->outputs[0], idx)) {
T
tensor-tang 已提交
52
        satisfied_all = satisfied_all && x->outputs[1]->IsVar() &&
T
tensor-tang 已提交
53
                        x->outputs[1]->outputs.empty();
T
tensor-tang 已提交
54
      } else {
T
tensor-tang 已提交
55 56 57
        satisfied_all =
            satisfied_all && is_nth_input_var_of_concat(x->outputs[1], idx) &&
            x->outputs[0]->IsVar() && x->outputs[0]->outputs.size() == 0;
T
tensor-tang 已提交
58 59
      }
    }
T
tensor-tang 已提交
60
    return satisfied_all;
T
tensor-tang 已提交
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
  };

  auto* concat_op = pattern->NewNode(
      [=](Node* x) { return is_concat_op_with_inputs(x, num_inputs); },
      name_scope + "/concat_op");
  concat_op->assert_op_attr<int>("axis", 1);

  auto* concat_out_var = pattern->NewNode(
      [=](Node* x) {
        return x && x->IsVar() && VarLinksFromOp(x, "concat") &&
               x->inputs.size() == 1 &&
               is_concat_op_with_inputs(x->inputs[0], num_inputs);
      },
      name_scope + "/concat_out_var");
  concat_out_var->assert_is_only_output_of_op("concat");

  std::vector<PDNode*> seqpool_ops_input_var(num_inputs);
  std::vector<PDNode*> seqpool_ops_output_var(num_inputs);
79
  std::vector<PDNode*> seqpool_ops_output_unused_var(num_inputs);
T
tensor-tang 已提交
80 81 82 83 84 85 86 87 88 89 90 91
  std::vector<PDNode*> seqpool_ops(num_inputs);

  for (int i = 0; i < num_inputs; ++i) {
    seqpool_ops_output_var[i] = pattern->NewNode(
        [=](Node* x) {
          return x && x->IsVar() && is_nth_input_var_of_concat(x, i) &&
                 x->inputs.size() == 1 &&
                 is_seqpool_op_with_pootype_of_nth_input_of_concat(x->inputs[0],
                                                                   "SUM", i);
        },
        name_scope + "/sequence_pool_out_" + std::to_string(i));

92 93 94 95 96 97 98 99 100
    seqpool_ops_output_unused_var[i] = pattern->NewNode(
        [=](Node* x) {
          return x && x->IsVar() && x->inputs.size() == 1 &&
                 x->outputs.size() == 0 &&
                 is_seqpool_op_with_pootype_of_nth_input_of_concat(x->inputs[0],
                                                                   "SUM", i);
        },
        name_scope + "/sequence_pool_unused_out_" + std::to_string(i));

T
tensor-tang 已提交
101 102 103 104 105 106 107 108 109
    seqpool_ops[i] = pattern->NewNode(
        [=](Node* x) {
          return x && x->IsOp() &&
                 is_seqpool_op_with_pootype_of_nth_input_of_concat(x, "SUM", i);
        },
        name_scope + "/sequence_pool_op_" + std::to_string(i));

    seqpool_ops_input_var[i] = pattern->NewNode(
        [=](Node* x) {
110 111 112 113 114 115 116 117 118 119
          bool basic = x && x->IsVar() && x->outputs.size() >= 1;
          bool next_is_fine = false;
          for (auto* o : x->outputs) {
            if (is_seqpool_op_with_pootype_of_nth_input_of_concat(o, "SUM",
                                                                  i)) {
              next_is_fine = true;
              break;
            }
          }
          return basic && next_is_fine;
T
tensor-tang 已提交
120 121 122 123 124 125
        },
        name_scope + "/sequence_pool_in_" + std::to_string(i));

    // Links
    seqpool_ops[i]
        ->LinksFrom({seqpool_ops_input_var[i]})
126
        .LinksTo({seqpool_ops_output_var[i], seqpool_ops_output_unused_var[i]});
T
tensor-tang 已提交
127 128 129 130 131
  }
  concat_op->LinksFrom(seqpool_ops_output_var).LinksTo({concat_out_var});
  return concat_out_var;
}

132
int BuildFusion(Graph* graph, const std::string& name_scope, int num_inputs) {
T
tensor-tang 已提交
133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
  GraphPatternDetector gpd;
  auto* pattern = gpd.mutable_pattern();
  BuildSeqPoolConcatPattern(pattern, name_scope, num_inputs);

  auto retrieve_node = [](const std::string& name,
                          const GraphPatternDetector::subgraph_t& subgraph,
                          const PDPattern& pat) -> Node* {
    PADDLE_ENFORCE(subgraph.count(pat.RetrieveNode(name)),
                   "pattern has no Node called %s", name.c_str());
    Node* p = subgraph.at(pat.RetrieveNode(name));
    PADDLE_ENFORCE_NOT_NULL(p, "subgraph has no node %s", name.c_str());
    return p;
  };

  int fusion_count{0};
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "handle SeqPool Concat fuse";
    std::vector<std::string> input_names(num_inputs);
    std::vector<Node*> input_vars(num_inputs);
    auto& fused_pattern = gpd.pattern();
    for (int i = 0; i < num_inputs; ++i) {
      input_vars[i] =
          retrieve_node(name_scope + "/sequence_pool_in_" + std::to_string(i),
                        subgraph, fused_pattern);
      input_names[i] = input_vars[i]->Name();
    }
    auto* concat_op =
        retrieve_node(name_scope + "/concat_op", subgraph, fused_pattern);
    auto* concat_out_var =
        retrieve_node(name_scope + "/concat_out_var", subgraph, fused_pattern);
    auto* seqpool_op0 = retrieve_node(name_scope + "/sequence_pool_op_0",
                                      subgraph, fused_pattern);

    // Create New OpDesc
    OpDesc op_desc;
    op_desc.SetType("fusion_seqpool_concat");
    op_desc.SetInput("X", input_names);
    op_desc.SetAttr("pooltype", seqpool_op0->Op()->GetAttr("pooltype"));
    op_desc.SetAttr("axis", concat_op->Op()->GetAttr("axis"));
    op_desc.SetOutput("Out", {concat_out_var->Name()});
    auto* op = graph->CreateOpNode(&op_desc);
    for (size_t i = 0; i < input_vars.size(); ++i) {
      IR_NODE_LINK_TO(input_vars[i], op);
    }
    IR_NODE_LINK_TO(op, concat_out_var);

    std::unordered_set<const Node*> marked_nodes;
    for (auto& item : subgraph) {
      marked_nodes.insert(item.second);
    }
    for (size_t i = 0; i < input_vars.size(); ++i) {
      marked_nodes.erase(input_vars[i]);
    }
    marked_nodes.erase(concat_out_var);
    GraphSafeRemoveNodes(graph, marked_nodes);
    ++fusion_count;
  };

  gpd(graph, handler);
  return fusion_count;
}

std::unique_ptr<ir::Graph> SeqPoolConcatFusePass::ApplyImpl(
    std::unique_ptr<ir::Graph> graph) const {
  FusePassBase::Init(name_scope_, graph.get());
  int fusion_count = 0;
  for (int i = MAX_CONCAT_INPUTS; i > 0; --i) {
201 202
    fusion_count +=
        BuildFusion(graph.get(), name_scope_ + "/" + std::to_string(i), i);
T
tensor-tang 已提交
203 204 205 206 207 208 209 210 211 212 213 214
  }
  AddStatis(fusion_count);

  return graph;
}

}  // namespace ir
}  // namespace framework
}  // namespace paddle

REGISTER_PASS(seqpool_concat_fuse_pass,
              paddle::framework::ir::SeqPoolConcatFusePass);