paddle_pass_builder.cc 3.7 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"
Y
Yan Chunwei 已提交
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 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 67 68
#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);
}

void GpuPassStrategy::EnableMKLDNN() {
  LOG(ERROR) << "GPU not support MKLDNN yet";
}

69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
  passes_.assign({
    "infer_clean_graph_pass",                        //
        "identity_scale_op_clean_pass",              //
        "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",       //
#endif
  });

  for (int i = 6; i >= 3; i--) {
    passes_.push_back("transpose_flatten" + std::to_string(i) +
                      "_concat_fuse_pass");
  }
  use_gpu_ = true;
}

Y
Yan Chunwei 已提交
91 92 93 94
void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) {
  analysis_passes_.push_back(pass);
}

95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118
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",  //
  });
  use_gpu_ = false;
}
119
}  // namespace paddle