analysis_predictor.h 5.4 KB
Newer Older
Y
Yan Chunwei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

15
#pragma once
16 17
#include <algorithm>
#include <map>
18 19
#include <string>
#include <vector>
20
#include "paddle/fluid/framework/naive_executor.h"
Y
Yan Chunwei 已提交
21 22
#include "paddle/fluid/inference/analysis/analyzer.h"
#include "paddle/fluid/inference/api/api_impl.h"
Y
Yan Chunwei 已提交
23
#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
Y
Yan Chunwei 已提交
24
#include "paddle/fluid/inference/api/paddle_inference_api.h"
25
#include "paddle/fluid/string/printf.h"
26 27 28 29
#ifdef PADDLE_WITH_TESTING
#include <gtest/gtest.h>
#include <gtest/gtest_prod.h>
#endif
Y
Yan Chunwei 已提交
30 31 32 33 34
namespace paddle {

using inference::analysis::Argument;
using inference::analysis::Analyzer;
using framework::proto::ProgramDesc;
35
using framework::NaiveExecutor;
Y
Yan Chunwei 已提交
36

37 38 39 40 41
/** \brief This predictor is based on the original native predictor with IR and
 * Analysis support.
 *
 * It will optimize IR and Parameters in the runtime.
 *
Y
Yan Chunwei 已提交
42 43
 * TODO(Superjomn) Replace the Navive predictor?
 */
44
class AnalysisPredictor : public PaddlePredictor {
Y
Yan Chunwei 已提交
45
 public:
46
  explicit AnalysisPredictor(const AnalysisConfig &config) : config_(config) {}
F
flame 已提交
47
  ~AnalysisPredictor();
Y
Yan Chunwei 已提交
48

49 50
  bool Init(const std::shared_ptr<framework::Scope> &parent_scope,
            const std::shared_ptr<framework::ProgramDesc> &program = nullptr);
Y
Yan Chunwei 已提交
51

52 53 54 55 56 57 58 59 60 61 62
  bool Run(const std::vector<PaddleTensor> &inputs,
           std::vector<PaddleTensor> *output_data,
           int batch_size = -1) override;

  std::unique_ptr<ZeroCopyTensor> GetInputTensor(
      const std::string &name) override;
  std::unique_ptr<ZeroCopyTensor> GetOutputTensor(
      const std::string &name) override;

  bool ZeroCopyRun() override;

63
  void CreateFeedFetchVar(framework::Scope *scope);
64
  void PrepareFeedFetch();
Y
Yan Chunwei 已提交
65 66 67

  void OptimizeInferenceProgram();

68 69 70 71
  Argument &analysis_argument() { return argument_; }

  std::unique_ptr<PaddlePredictor> Clone() override;

72
  framework::Scope *scope() { return scope_.get(); }
73 74
  framework::ProgramDesc &program() { return *inference_program_; }

L
luotao1 已提交
75
  void SetMkldnnThreadID(int tid);
L
luotao1 已提交
76

77
  std::string GetSerializedProgram() const override;
Y
Yan Chunwei 已提交
78

79
 protected:
Y
Yan Chunwei 已提交
80 81 82 83 84
  // For memory optimization.
  bool need_collect_var_shapes_for_memory_optim();
  void CollectVarShapes();
  void SerializeBatchVarShapes(const std::string &path);

85 86 87 88 89
  bool PrepareProgram(const std::shared_ptr<framework::ProgramDesc> &program);
  bool PrepareScope(const std::shared_ptr<framework::Scope> &parent_scope);
  bool CreateExecutor();
  bool PrepareExecutor();

90
  bool LoadProgramDesc();
91
  bool LoadParameters();
92 93 94 95 96 97 98 99

  bool SetFeed(const std::vector<PaddleTensor> &input_datas,
               framework::Scope *scope);
  bool GetFetch(std::vector<PaddleTensor> *output_data,
                framework::Scope *scope);
  template <typename T>
  void GetFetchOne(const framework::LoDTensor &fetchs,
                   PaddleTensor *output_data);
Y
Yan Chunwei 已提交
100

N
nhzlx 已提交
101 102 103 104 105 106 107 108 109 110 111 112
#if PADDLE_WITH_TENSORRT
  // When we use Paddle-TRT INT8 engine, we need to generate calibration table
  // data first,
  // the calibration table contains the range for each op's input and output,
  // this whole process can be divided into several steps:
  //
  // 1. Builds a 32-bit engine, runs it on the calibration set, and records a
  // histogram for each
  // tensor of the distribution of activation values.
  // 2. Builds a calibration table from the histograms.
  //
  // After step 2, we need to store the calibration table on disk
N
nhzlx 已提交
113
  bool SaveTrtCalibToDisk();
N
nhzlx 已提交
114
#endif
N
nhzlx 已提交
115

116 117 118 119 120 121 122 123
// Some more detailed tests, they are made the friends of the predictor, so that
// the all the details can be tested.
#if PADDLE_WITH_TESTING
  FRIEND_TEST(AnalysisPredictor, analysis_off);
  FRIEND_TEST(AnalysisPredictor, analysis_on);
  FRIEND_TEST(AnalysisPredictor, with_gpu);
#endif

Y
Yan Chunwei 已提交
124
 private:
125
  AnalysisConfig config_;
Y
Yan Chunwei 已提交
126
  Argument argument_;
127 128 129 130 131 132 133
  std::unique_ptr<NaiveExecutor> executor_;
  platform::Place place_;
  std::shared_ptr<framework::Scope> scope_;
  framework::Scope *sub_scope_{nullptr};
  std::shared_ptr<framework::ProgramDesc> inference_program_;
  std::vector<framework::OpDesc *> feeds_;
  std::map<std::string, size_t> feed_names_;
Y
Yan Chunwei 已提交
134
  std::vector<framework::OpDesc *> fetches_;
135
  // Memory buffer for feed inputs. The temporary LoDTensor will cause serious
136
  // concurrency problems, wrong results and memory leak, so cache them.
137
  std::vector<framework::LoDTensor> feed_tensors_;
Y
Yan Chunwei 已提交
138
  details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
Y
Yan Chunwei 已提交
139 140
  // A mutex help to make Clone thread safe.
  std::mutex clone_mutex_;
141

Y
Yan Chunwei 已提交
142 143 144 145 146
  // For memory optimization.
  const size_t max_shape_collect_count_{1000};
  int need_collect_var_shapes_{-1};  // -1 for default, 0 for false, 1 for true.
  std::vector<std::map<std::string, std::vector<int>>> batch_var_shapes_;

147 148 149 150 151 152
 private:
  // Some status here that help to determine the status inside the predictor.
  bool status_program_optimized_{false};
  bool status_is_cloned_{false};
  bool status_use_gpu_{false};
  bool status_ir_optim_enabled_{false};
Y
Yan Chunwei 已提交
153 154 155
};

}  // namespace paddle