paddle_pass_builder.cc 5.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// 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/inference/api/paddle_pass_builder.h"
16 17 18
#ifdef PADDLE_WITH_CUDA
#include <cudnn.h>
#endif
19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66
#include <glog/logging.h>

namespace paddle {

void PaddlePassBuilder::AppendPass(const std::string &pass_type) {
  passes_.push_back(pass_type);
}

void PaddlePassBuilder::TurnOnDebug() {
  std::vector<std::string> passes;
  auto it = std::begin(passes_);
  while (it != std::end(passes_)) {
    if (*it != "graph_viz_pass") {
      it = passes_.insert(it + 1, "graph_viz_pass");
    } else {
      ++it;
    }
  }
}

std::string PaddlePassBuilder::DebugString() {
  std::stringstream ss;
  ss << "Passes to apply:\n";
  for (auto &pass : passes_) {
    ss << "  - " << pass << '\n';
  }
  return ss.str();
}

void PaddlePassBuilder::DeletePass(const std::string &pass_type) {
  auto it = std::begin(passes_);
  while (it != std::end(passes_)) {
    if (*it == pass_type) {
      it = passes_.erase(it);
    } else {
      ++it;
    }
  }
}

void PaddlePassBuilder::InsertPass(size_t idx, const std::string &pass_type) {
  passes_.insert(std::begin(passes_) + idx, pass_type);
}

void PaddlePassBuilder::DeletePass(size_t idx) {
  passes_.erase(std::begin(passes_) + idx);
}

W
Wojciech Uss 已提交
67 68
void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) {
  analysis_passes_.push_back(pass);
69 70
}

W
Wojciech Uss 已提交
71 72
void PaddlePassBuilder::ClearPasses() { passes_.clear(); }

73 74
// The following passes works for Anakin sub-graph engine.
const std::vector<std::string> kAnakinSubgraphPasses({
N
nhzlx 已提交
75 76 77 78 79 80 81 82 83
    "infer_clean_graph_pass",                       //
    "simplify_anakin_priorbox_detection_out_pass",  //
    "fillconstant_elementwisemul_fuse",             //
    "fc_fuse_pass",                                 //
    "conv_elementwise_add_fuse_pass",               //
    "conv_bn_fuse_pass",                            //
    "conv_elementwise_add_fuse_pass",               //
    "fc_gru_fuse_pass",                             //
    "quant_conv2d_dequant_fuse_pass",               //
84 85 86
    "anakin_subgraph_pass",
});

87 88
GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
  passes_.assign({
89 90
    "infer_clean_graph_pass",  //
        //   "identity_scale_op_clean_pass",              //
91 92 93 94 95 96 97 98
        "conv_affine_channel_fuse_pass",             //
        "conv_eltwiseadd_affine_channel_fuse_pass",  //
        "conv_bn_fuse_pass",                         //
#if CUDNN_VERSION >= 7100  // To run conv_fusion, the version of cudnn must be
                           // guaranteed at least v7
        "conv_elementwise_add_act_fuse_pass",   //
        "conv_elementwise_add2_act_fuse_pass",  //
        "conv_elementwise_add_fuse_pass",       //
L
luotao1 已提交
99
        "runtime_context_cache_pass",           //
N
nhzlx 已提交
100 101
#endif                                          //
        "transpose_flatten_concat_fuse_pass",
102 103 104 105 106
  });

  use_gpu_ = true;
}

W
Wojciech Uss 已提交
107 108
void GpuPassStrategy::EnableMKLDNN() {
  LOG(ERROR) << "GPU not support MKLDNN yet";
109 110
}

W
Wojciech Uss 已提交
111 112
void GpuPassStrategy::EnableMkldnnQuantizer() {
  LOG(ERROR) << "GPU not support MKL-DNN quantization";
Y
Yan Chunwei 已提交
113 114
}

115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
CpuPassStrategy::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",       //
      "seqpool_concat_fuse_pass",       //
      "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",                  //
      "repeated_fc_relu_fuse_pass",    //
      "squared_mat_sub_fuse_pass",     //
      "conv_bn_fuse_pass",             //
      "conv_eltwiseadd_bn_fuse_pass",  //
      "is_test_pass",                  //
      "identity_scale_op_clean_pass",  //
L
luotao1 已提交
136
      "runtime_context_cache_pass",    //
137 138 139
  });
  use_gpu_ = false;
}
W
Wojciech Uss 已提交
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

void CpuPassStrategy::EnableMKLDNN() {
// TODO(Superjomn) Consider the way to mix CPU with GPU.
#ifdef PADDLE_WITH_MKLDNN
  if (!use_mkldnn_) {
    passes_.insert(passes_.begin(), "mkldnn_placement_pass");

    for (auto &pass : std::vector<std::string>(
             {"depthwise_conv_mkldnn_pass",    //
              "conv_bn_fuse_pass",             // Execute BN passes again to
              "conv_eltwiseadd_bn_fuse_pass",  // preserve correct pass order
              "conv_bias_mkldnn_fuse_pass",    //
              "conv3d_bias_mkldnn_fuse_pass",  //
              "conv_elementwise_add_mkldnn_fuse_pass",
              "conv_relu_mkldnn_fuse_pass"})) {
      passes_.push_back(pass);
    }
  }
  use_mkldnn_ = true;
#else
  use_mkldnn_ = false;
#endif
}

void CpuPassStrategy::EnableMkldnnQuantizer() {
#ifdef PADDLE_WITH_MKLDNN
  if (!use_mkldnn_quantizer_) {
    passes_.push_back("cpu_quantize_placement_pass");
  }
  use_mkldnn_quantizer_ = true;
#else
  use_mkldnn_quantizer_ = false;
#endif
}

175
}  // namespace paddle