paddle_pass_builder.cc 7.7 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"
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
const std::vector<std::string> kTRTSubgraphPasses({
74
  "conv_affine_channel_fuse_pass",                 //
75
      "conv_eltwiseadd_affine_channel_fuse_pass",  //
76
      "shuffle_channel_detect_pass",               //
77 78 79 80 81 82 83 84 85 86 87 88 89 90
      "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({
93
    "quant_conv2d_dequant_fuse_pass",               //
N
nhzlx 已提交
94 95 96 97 98
    "simplify_anakin_priorbox_detection_out_pass",  //
    "fillconstant_elementwisemul_fuse",             //
    "fc_fuse_pass",                                 //
    "conv_elementwise_add_fuse_pass",               //
    "fc_gru_fuse_pass",                             //
99 100 101
    "shuffle_channel_detect_pass",                  //
    "anakin_subgraph_pass",                         //
    "fc_gru_fuse_pass",                             //
102 103
});

104 105
GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
  passes_.assign({
106
    //   "identity_scale_op_clean_pass",             //
107 108
    "is_test_pass",                                  //
        "simplify_with_basic_ops_pass",              //
109
        "fc_fuse_pass",                              //
110
        "fc_elementwise_layernorm_fuse_pass",        //
111
        "conv_affine_channel_fuse_pass",             //
112 113
        "conv_eltwiseadd_affine_channel_fuse_pass",  //
        "conv_bn_fuse_pass",                         //
114
        "conv_eltwiseadd_bn_fuse_pass",              //
115 116 117 118 119
#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
120 121
#endif                                          //
        "transpose_flatten_concat_fuse_pass",
122
        // following pass should be located in the last, since it will
123 124
        // work on all fused ops.
        "runtime_context_cache_pass"
125 126 127 128 129
  });

  use_gpu_ = true;
}

130 131 132 133 134 135 136
void GpuPassStrategy::EnableCUDNN() {
  if (!use_cudnn_) {
    passes_.insert(passes_.begin(), "cudnn_placement_pass");
  }
  use_cudnn_ = true;
}

W
Wojciech Uss 已提交
137 138
void GpuPassStrategy::EnableMKLDNN() {
  LOG(ERROR) << "GPU not support MKLDNN yet";
139 140
}

W
Wojciech Uss 已提交
141 142
void GpuPassStrategy::EnableMkldnnQuantizer() {
  LOG(ERROR) << "GPU not support MKL-DNN quantization";
Y
Yan Chunwei 已提交
143 144
}

M
mozga-intel 已提交
145 146 147 148
void GpuPassStrategy::EnableNgraph() {
  LOG(ERROR) << "GPU not support Ngraph yet";
}

149 150 151
CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
  // NOTE the large fusions should be located in the front, so that they will
  // not be damaged by smaller ones.
152 153
  passes_.assign({"simplify_with_basic_ops_pass",   //
                  "attention_lstm_fuse_pass",       //
154 155
                  "seqconv_eltadd_relu_fuse_pass",  //
                  // "seqpool_concat_fuse_pass",    //
156
                  "seqpool_cvm_concat_fuse_pass",  //
157 158 159 160 161 162 163 164 165 166 167 168
                  // "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",                  //
169 170
                  // following pass should be located in the last, since
                  // it will work on all fused ops.
171
                  "runtime_context_cache_pass"});
Y
Yan Chunwei 已提交
172

173 174
  use_gpu_ = false;
}
W
Wojciech Uss 已提交
175

176 177
void CpuPassStrategy::EnableCUDNN() { LOG(ERROR) << "CPU not support cuDNN"; }

W
Wojciech Uss 已提交
178 179 180 181 182 183
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");

184 185 186 187 188
    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",    //
189
             "conv_transpose_bias_mkldnn_fuse_pass",
190 191 192
             "conv3d_bias_mkldnn_fuse_pass",  //
             "conv_elementwise_add_mkldnn_fuse_pass",
             "conv_concat_relu_mkldnn_fuse_pass",
193 194 195
             "conv_relu_mkldnn_fuse_pass",        //
             "conv_leaky_relu_mkldnn_fuse_pass",  //
             "conv_relu6_mkldnn_fuse_pass",       //
196 197 198
             // Disabled due to topology-dependent speed-up
             // "fc_mkldnn_pass"
         })) {
W
Wojciech Uss 已提交
199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
      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
219 220 221 222 223 224 225 226 227 228
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
}
229
}  // namespace paddle