paddle_pass_builder.cc 8.8 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 75
  "conv_affine_channel_fuse_pass",  //
      "adaptive_pool2d_convert_global_pass",
76
      "conv_eltwiseadd_affine_channel_fuse_pass",  //
77
      "shuffle_channel_detect_pass",               //
78 79
      "quant_conv2d_dequant_fuse_pass",            //
      "delete_quant_dequant_op_pass",              //
P
Pei Yang 已提交
80
      // "fc_fuse_pass",                                 //
81 82 83 84
      "simplify_with_basic_ops_pass",           //
      "embedding_eltwise_layernorm_fuse_pass",  //
      "multihead_matmul_fuse_pass_v2",          //
      "skip_layernorm_fuse_pass",               //
85 86 87 88 89
      "unsqueeze2_eltwise_fuse_pass",
      "conv_bn_fuse_pass",       //
      "fc_fuse_pass",            //
      "tensorrt_subgraph_pass",  //
      "conv_bn_fuse_pass",       //
90 91 92 93 94 95 96 97 98
#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",
});

石晓伟 已提交
99 100 101 102 103 104
const std::vector<std::string> kLiteSubgraphPasses({
#ifdef PADDLE_WITH_LITE
    "lite_subgraph_pass",
#endif
});

105 106
GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
  passes_.assign({
107
    //   "identity_scale_op_clean_pass",             //
108 109 110
    "is_test_pass",                                  //
        "simplify_with_basic_ops_pass",              //
        "conv_affine_channel_fuse_pass",             //
111 112
        "conv_eltwiseadd_affine_channel_fuse_pass",  //
        "conv_bn_fuse_pass",                         //
113
        "conv_eltwiseadd_bn_fuse_pass",              //
114 115 116 117
        "embedding_eltwise_layernorm_fuse_pass",     //
        "multihead_matmul_fuse_pass_v2",             //
        "fc_fuse_pass",                              //
        "fc_elementwise_layernorm_fuse_pass",        //
118 119 120 121 122
#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
123
#endif                                          //
石晓伟 已提交
124
        "transpose_flatten_concat_fuse_pass",   //
125
        // following pass should be located in the last, since it will
126 127
        // work on all fused ops.
        "runtime_context_cache_pass"
128 129 130 131 132
  });

  use_gpu_ = true;
}

133 134 135 136 137 138 139
void GpuPassStrategy::EnableCUDNN() {
  if (!use_cudnn_) {
    passes_.insert(passes_.begin(), "cudnn_placement_pass");
  }
  use_cudnn_ = true;
}

W
Wojciech Uss 已提交
140 141
void GpuPassStrategy::EnableMKLDNN() {
  LOG(ERROR) << "GPU not support MKLDNN yet";
142 143
}

W
Wojciech Uss 已提交
144 145
void GpuPassStrategy::EnableMkldnnQuantizer() {
  LOG(ERROR) << "GPU not support MKL-DNN quantization";
Y
Yan Chunwei 已提交
146 147
}

148 149 150 151
void GpuPassStrategy::EnableMkldnnBfloat16() {
  LOG(ERROR) << "GPU not support MKL-DNN bfloat16";
}

152 153 154
CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
  // NOTE the large fusions should be located in the front, so that they will
  // not be damaged by smaller ones.
155 156
  passes_.assign({"simplify_with_basic_ops_pass",   //
                  "attention_lstm_fuse_pass",       //
157 158
                  "seqconv_eltadd_relu_fuse_pass",  //
                  // "seqpool_concat_fuse_pass",    //
159
                  "seqpool_cvm_concat_fuse_pass",  //
160
                  // "embedding_fc_lstm_fuse_pass", //
161 162
                  // TODO(wilber): fix correctness problem.
                  // "fc_lstm_fuse_pass",                       //
163 164 165 166 167 168 169 170 171 172 173 174
                  "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",            //
                  "conv_transpose_bn_fuse_pass",             //
                  "conv_transpose_eltwiseadd_bn_fuse_pass",  //
                  "is_test_pass",                            //
175 176
                  // following pass should be located in the last, since
                  // it will work on all fused ops.
177
                  "runtime_context_cache_pass"});
Y
Yan Chunwei 已提交
178

179 180
  use_gpu_ = false;
}
W
Wojciech Uss 已提交
181

182 183
void CpuPassStrategy::EnableCUDNN() { LOG(ERROR) << "CPU not support cuDNN"; }

W
Wojciech Uss 已提交
184 185 186 187 188 189
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");

190
    for (auto &pass : std::vector<std::string>({
191 192 193 194 195 196 197 198
             "depthwise_conv_mkldnn_pass",     //
             "conv_bn_fuse_pass",              // Execute BN passes again to
             "conv_eltwiseadd_bn_fuse_pass",   // preserve correct pass order
             "conv_affine_channel_fuse_pass",  //
             "conv_eltwiseadd_affine_channel_fuse_pass",  //
             "conv_transpose_bn_fuse_pass",               //
             "conv_transpose_eltwiseadd_bn_fuse_pass",    //
             "conv_bias_mkldnn_fuse_pass",                //
199
             "conv_transpose_bias_mkldnn_fuse_pass",
200 201 202
             "conv3d_bias_mkldnn_fuse_pass",  //
             "conv_elementwise_add_mkldnn_fuse_pass",
             "conv_concat_relu_mkldnn_fuse_pass",
203 204 205 206 207 208 209
             "conv_relu_mkldnn_fuse_pass",                 //
             "conv_leaky_relu_mkldnn_fuse_pass",           //
             "conv_relu6_mkldnn_fuse_pass",                //
             "conv_swish_mkldnn_fuse_pass",                //
             "scale_matmul_fuse_pass",                     //
             "reshape_transpose_matmul_mkldnn_fuse_pass",  //
             "matmul_transpose_reshape_fuse_pass",         //
210
             // Disabled due to topology-dependent speed-up
211 212
             // "fc_mkldnn_pass",
             // "fc_act_mkldnn_fuse_pass",
213
             "batch_norm_act_fuse_pass",
214 215
             "mkldnn_inplace_pass",  // This pass should be activated after
                                     // fuses
216
         })) {
W
Wojciech Uss 已提交
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236
      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
}

237 238
void CpuPassStrategy::EnableMkldnnBfloat16() {
#ifdef PADDLE_WITH_MKLDNN
239 240 241 242
  if (!use_mkldnn_bfloat16_) {
    passes_.push_back("cpu_bfloat16_placement_pass");
    passes_.push_back("cpu_bfloat16_pass");
  }
243 244 245 246 247 248
  use_mkldnn_bfloat16_ = true;
#else
  use_mkldnn_bfloat16_ = false;
#endif
}

249
}  // namespace paddle