paddle_pass_builder.h 4.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
// 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.

#pragma once

#include <sstream>
#include <string>
#include <vector>

21 22 23
/*! \file */

/*! \namespace paddle */
24
namespace paddle {
25 26

/** This is a pass builder based on string. It is part of inference API.
27 28 29 30 31 32
 */
class PaddlePassBuilder {
 public:
  explicit PaddlePassBuilder(const std::vector<std::string> &passes)
      : passes_(passes) {}

33
  /** Append a pass to the end of the passes. */
34 35
  void AppendPass(const std::string &pass_type);

36 37 38 39
  /** Insert a pass to a specific position.
   * @param idx the position to insert.
   * @param pass_type the pass key.
   */
40 41
  void InsertPass(size_t idx, const std::string &pass_type);

42
  /** Delete the `idx`-th pass. */
43 44
  void DeletePass(size_t idx);

45
  /** Delete all the passes that has type `pass_type`. */
46 47
  void DeletePass(const std::string &pass_type);

48 49 50
  /** Visualize the computation graph after each pass by generating a DOT
   * language file, one can draw them with the Graphviz toolkit.
   */
51 52
  void TurnOnDebug();

53
  /** Human-readible information. */
54 55 56 57 58 59 60 61
  std::string DebugString();

  const std::vector<std::string> &AllPasses() const { return passes_; }

 protected:
  std::vector<std::string> passes_;
};

62
/**Pass strategy to help control the IR passes.
63 64 65 66 67 68
 */
class PassStrategy : public PaddlePassBuilder {
 public:
  explicit PassStrategy(const std::vector<std::string> &passes)
      : PaddlePassBuilder(passes) {}

69 70 71
  /** The MKLDNN control exists in both CPU and GPU mode, because there can be
   * still some CPU kernels running in CPU mode.
   */
72 73
  virtual void EnableMKLDNN() = 0;

74 75
  bool use_gpu() const { return use_gpu_; }

76
  virtual ~PassStrategy() = default;
77 78 79

 protected:
  bool use_gpu_{false};
80 81
};

82
/** The CPU passes controller, it is used in AnalysisPredictor with CPU mode.
83 84 85 86 87 88 89 90 91
 */
class CpuPassStrategy : public PassStrategy {
 public:
  CpuPassStrategy() : PassStrategy({}) {
    // NOTE the large fusions should be located in the front, so that they will
    // not be damaged by smaller ones.
    passes_.assign({
        "infer_clean_graph_pass",         //
        "attention_lstm_fuse_pass",       //
T
tensor-tang 已提交
92
        "seqpool_concat_fuse_pass",       //
93 94 95 96 97 98 99 100 101 102
        "seqconv_eltadd_relu_fuse_pass",  //
        // "embedding_fc_lstm_fuse_pass", //
        "fc_lstm_fuse_pass",             //
        "mul_lstm_fuse_pass",            //
        "fc_gru_fuse_pass",              //
        "mul_gru_fuse_pass",             //
        "seq_concat_fc_fuse_pass",       //
        "fc_fuse_pass",                  //
        "conv_bn_fuse_pass",             //
        "conv_eltwiseadd_bn_fuse_pass",  //
103
        "is_test_pass",                  //
104
    });
105
    use_gpu_ = false;
106 107 108 109
  }

  virtual ~CpuPassStrategy() = default;

110
  void EnableMKLDNN() override {
111 112 113 114 115
// TODO(Superjomn) Consider the way to mix CPU with GPU.
#ifdef PADDLE_WITH_MKLDNN
    passes_.insert(passes_.begin(), "mkldnn_placement_pass");

    for (auto &pass :
116 117 118 119
         std::vector<std::string>({"depthwise_conv_mkldnn_pass",    //
                                   "conv_bias_mkldnn_fuse_pass",    //
                                   "conv3d_bias_mkldnn_fuse_pass",  //
                                   "conv_relu_mkldnn_fuse_pass",    //
120 121 122 123 124 125 126 127 128
                                   "conv_elementwise_add_mkldnn_fuse_pass"})) {
      passes_.push_back(pass);
    }
#endif
  }

  CpuPassStrategy(const CpuPassStrategy &other) : PassStrategy(other.passes_) {}
};

129
/** The GPU passes strategy, it is used in AnalysisPredictor with GPU mode.
130 131 132 133 134
 */
class GpuPassStrategy : public PassStrategy {
 public:
  GpuPassStrategy() : PassStrategy({}) {
    passes_.assign({
N
nhzlx 已提交
135 136 137 138 139 140 141
        "infer_clean_graph_pass",                    //
        "conv_affine_channel_fuse_pass",             //
        "conv_eltwiseadd_affine_channel_fuse_pass",  //
        "conv_bn_fuse_pass",                         //
        "conv_elementwise_add_act_fuse_pass",        //
        "conv_elementwise_add2_act_fuse_pass",       //
        "conv_elementwise_add_fuse_pass",            //
142
    });
143

144 145 146 147
    for (int i = 6; i >= 3; i--) {
      passes_.push_back("transpose_flatten" + std::to_string(i) +
                        "_concat_fuse_pass");
    }
148
    use_gpu_ = true;
149 150 151
  }

  GpuPassStrategy(const GpuPassStrategy &other)
152 153 154
      : PassStrategy(other.AllPasses()) {
    use_gpu_ = true;
  }
155

156
  void EnableMKLDNN() override;
157 158 159 160 161

  virtual ~GpuPassStrategy() = default;
};

}  // namespace paddle