paddle_pass_builder.cc 7.2 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 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
const std::vector<std::string> kTRTSubgraphPasses({
  "infer_clean_graph_pass",                        //
      "conv_affine_channel_fuse_pass",             //
      "conv_eltwiseadd_affine_channel_fuse_pass",  //
      "quant_conv2d_dequant_fuse_pass",            //
      "delete_quant_dequant_op_pass",              //
      // "fc_fuse_pass",                                 //
      "tensorrt_subgraph_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                                        //
      "transpose_flatten_concat_fuse_pass",
});

91 92
// The following passes works for Anakin sub-graph engine.
const std::vector<std::string> kAnakinSubgraphPasses({
N
nhzlx 已提交
93
    "infer_clean_graph_pass",                       //
94
    "quant_conv2d_dequant_fuse_pass",               //
N
nhzlx 已提交
95 96 97 98 99
    "simplify_anakin_priorbox_detection_out_pass",  //
    "fillconstant_elementwisemul_fuse",             //
    "fc_fuse_pass",                                 //
    "conv_elementwise_add_fuse_pass",               //
    "fc_gru_fuse_pass",                             //
100 101 102
    "shuffle_channel_detect_pass",                  //
    "anakin_subgraph_pass",                         //
    "fc_gru_fuse_pass",                             //
103 104
});

105 106
GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
  passes_.assign({
107
    "infer_clean_graph_pass",  //
108
        //   "identity_scale_op_clean_pass",              //
109 110 111 112 113 114 115 116
        "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",       //
N
nhzlx 已提交
117 118
#endif                                          //
        "transpose_flatten_concat_fuse_pass",
119
        // following pass should be located in the last, since it will
120 121
        // work on all fused ops.
        "runtime_context_cache_pass"
122 123 124 125 126
  });

  use_gpu_ = true;
}

W
Wojciech Uss 已提交
127 128
void GpuPassStrategy::EnableMKLDNN() {
  LOG(ERROR) << "GPU not support MKLDNN yet";
129 130
}

W
Wojciech Uss 已提交
131 132
void GpuPassStrategy::EnableMkldnnQuantizer() {
  LOG(ERROR) << "GPU not support MKL-DNN quantization";
Y
Yan Chunwei 已提交
133 134
}

M
mozga-intel 已提交
135 136 137 138
void GpuPassStrategy::EnableNgraph() {
  LOG(ERROR) << "GPU not support Ngraph yet";
}

139 140 141
CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
  // NOTE the large fusions should be located in the front, so that they will
  // not be damaged by smaller ones.
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
  passes_.assign({"infer_clean_graph_pass",         //
                  "attention_lstm_fuse_pass",       //
                  "seqconv_eltadd_relu_fuse_pass",  //
                  // "seqpool_concat_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",                  //
158 159
                  // following pass should be located in the last, since
                  // it will work on all fused ops.
160
                  "runtime_context_cache_pass"});
Y
Yan Chunwei 已提交
161

162 163
  use_gpu_ = false;
}
W
Wojciech Uss 已提交
164 165 166 167 168 169 170

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");

171 172 173 174 175
    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",    //
176
             "conv_transpose_bias_mkldnn_fuse_pass",
177 178 179 180 181 182 183 184
             "conv3d_bias_mkldnn_fuse_pass",  //
             "conv_elementwise_add_mkldnn_fuse_pass",
             "conv_concat_relu_mkldnn_fuse_pass",
             "conv_relu_mkldnn_fuse_pass",   //
             "conv_brelu_mkldnn_fuse_pass",  //
             // Disabled due to topology-dependent speed-up
             // "fc_mkldnn_pass"
         })) {
W
Wojciech Uss 已提交
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204
      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
}

M
mozga-intel 已提交
205 206 207 208 209 210 211 212 213 214
void CpuPassStrategy::EnableNgraph() {
#ifdef PADDLE_WITH_NGRAPH
  if (!use_ngraph_) {
    passes_.insert(passes_.begin(), "ngraph_subgraph_pass");
  }
  use_ngraph_ = true;
#else
  use_ngraph_ = false;
#endif
}
215
}  // namespace paddle