analysis_predictor_tester.cc 6.2 KB
Newer Older
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
#include "paddle/fluid/inference/api/analysis_predictor.h"
16 17
#include <glog/logging.h>
#include <gtest/gtest.h>
18
#include <thread>  // NOLINT
19
#include "paddle/fluid/inference/api/helper.h"
20 21 22 23 24 25 26
#include "paddle/fluid/inference/api/paddle_inference_api.h"

DEFINE_string(dirname, "", "dirname to tests.");

namespace paddle {
using contrib::AnalysisConfig;

27
TEST(AnalysisPredictor, analysis_off) {
28 29 30
  AnalysisConfig config;
  config.SetModel(FLAGS_dirname);
  config.SwitchIrOptim(false);
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

  auto _predictor = CreatePaddlePredictor<AnalysisConfig>(config);
  auto* predictor = static_cast<AnalysisPredictor*>(_predictor.get());

  // Without analysis, the scope_ and sub_scope_ are created by predictor
  // itself.
  ASSERT_TRUE(predictor->scope_);
  ASSERT_TRUE(predictor->sub_scope_);
  ASSERT_EQ(predictor->scope_->parent(), nullptr);
  ASSERT_EQ(predictor->sub_scope_->parent(), predictor->scope_.get());
  // ir is turned off, so program shouldn't be optimized.
  ASSERT_FALSE(predictor->status_program_optimized_);
  LOG(INFO) << "scope parameters " << predictor->scope_->LocalVarNames().size();

  // 2. Dummy Input Data
  int64_t data[4] = {1, 2, 3, 4};
  PaddleTensor tensor;
  tensor.shape = std::vector<int>({4, 1});
  tensor.data.Reset(data, sizeof(data));
  tensor.dtype = PaddleDType::INT64;

  std::vector<PaddleTensor> inputs(4, tensor);
  std::vector<PaddleTensor> outputs;
  ASSERT_TRUE(predictor->Run(inputs, &outputs));
}

TEST(AnalysisPredictor, analysis_on) {
58 59 60
  AnalysisConfig config;
  config.SetModel(FLAGS_dirname);
  config.SwitchIrOptim(true);
61
#ifdef PADDLE_WITH_CUDA
62
  config.EnableUseGpu(100, 0);
63
#else
64
  config.DisableGpu();
65
#endif
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91

  auto _predictor = CreatePaddlePredictor<AnalysisConfig>(config);
  auto* predictor = static_cast<AnalysisPredictor*>(_predictor.get());

  ASSERT_TRUE(predictor->scope_);
  ASSERT_TRUE(predictor->sub_scope_);
  ASSERT_EQ(predictor->scope_->parent(), nullptr);
  ASSERT_EQ(predictor->sub_scope_->parent(), predictor->scope_.get());
  // ir is turned on, so program should be optimized.
  ASSERT_TRUE(predictor->status_program_optimized_);
  // 2. Dummy Input Data
  int64_t data[4] = {1, 2, 3, 4};
  PaddleTensor tensor;
  tensor.shape = std::vector<int>({4, 1});
  tensor.data.Reset(data, sizeof(data));
  tensor.dtype = PaddleDType::INT64;

  std::vector<PaddleTensor> inputs(4, tensor);
  std::vector<PaddleTensor> outputs;
  ASSERT_TRUE(predictor->Run(inputs, &outputs));

  for (auto& output : outputs) {
    LOG(INFO) << inference::DescribeTensor(output);
  }

  // compare with NativePredictor
92 93
  auto naive_predictor =
      CreatePaddlePredictor<NativeConfig>(config.ToNativeConfig());
94 95 96 97 98 99
  std::vector<PaddleTensor> naive_outputs;
  ASSERT_TRUE(naive_predictor->Run(inputs, &naive_outputs));
  ASSERT_EQ(naive_outputs.size(), 1UL);
  inference::CompareTensor(outputs.front(), naive_outputs.front());
}

100 101
TEST(AnalysisPredictor, ZeroCopy) {
  AnalysisConfig config;
102 103
  config.SetModel(FLAGS_dirname);
  config.SwitchUseFeedFetchOps(false);
S
superjomn 已提交
104
  auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137

  auto w0 = predictor->GetInputTensor("firstw");
  auto w1 = predictor->GetInputTensor("secondw");
  auto w2 = predictor->GetInputTensor("thirdw");
  auto w3 = predictor->GetInputTensor("forthw");

  w0->Reshape({4, 1});
  w1->Reshape({4, 1});
  w2->Reshape({4, 1});
  w3->Reshape({4, 1});

  auto* w0_data = w0->mutable_data<int64_t>(PaddlePlace::kCPU);
  auto* w1_data = w1->mutable_data<int64_t>(PaddlePlace::kCPU);
  auto* w2_data = w2->mutable_data<int64_t>(PaddlePlace::kCPU);
  auto* w3_data = w3->mutable_data<int64_t>(PaddlePlace::kCPU);

  for (int i = 0; i < 4; i++) {
    w0_data[i] = i;
    w1_data[i] = i;
    w2_data[i] = i;
    w3_data[i] = i;
  }

  predictor->ZeroCopyRun();

  auto out = predictor->GetOutputTensor("fc_1.tmp_2");
  PaddlePlace place;
  int size = 0;
  auto* out_data = out->data<float>(&place, &size);
  LOG(INFO) << "output size: " << size / sizeof(float);
  LOG(INFO) << "output_data: " << out_data;
}

138 139
TEST(AnalysisPredictor, Clone) {
  AnalysisConfig config;
140 141 142
  config.SetModel(FLAGS_dirname);
  config.SwitchUseFeedFetchOps(true);
  config.SwitchIrOptim(true);
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192

  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
  predictors.emplace_back(CreatePaddlePredictor(config));

  LOG(INFO) << "************** to clone ************************";
  const int num_threads = 3;
  for (int i = 1; i < num_threads; i++) {
    predictors.emplace_back(predictors.front()->Clone());
  }

  auto* root_scope =
      static_cast<AnalysisPredictor*>(predictors[0].get())->scope();
  ASSERT_FALSE(root_scope->kids().empty());
  LOG(INFO) << "***** scope ******\n"
            << framework::GenScopeTreeDebugInfo(root_scope);

  // 2. Dummy Input Data
  int64_t data[4] = {1, 2, 3, 4};
  PaddleTensor tensor;
  tensor.shape = std::vector<int>({4, 1});
  tensor.data.Reset(data, sizeof(data));
  tensor.dtype = PaddleDType::INT64;

  std::vector<PaddleTensor> inputs(4, tensor);
  std::vector<PaddleTensor> outputs;
  predictors[0]->Run(inputs, &outputs);

  LOG(INFO) << "Run with single thread";
  for (int i = 0; i < num_threads; i++) {
    LOG(INFO) << "run predictor " << i;
    ASSERT_TRUE(predictors[i]->Run(inputs, &outputs));
  }

  LOG(INFO) << "Run with multiple threads";
  std::vector<std::thread> threads;
  for (int i = 0; i < num_threads; i++) {
    threads.emplace_back([&predictors, &inputs, i] {
      LOG(INFO) << "thread #" << i << " running";
      std::vector<PaddleTensor> outputs;
      for (int j = 0; j < 10; j++) {
        ASSERT_TRUE(predictors[i]->Run(inputs, &outputs));
      }
    });
  }

  for (auto& t : threads) {
    t.join();
  }
}

193
}  // namespace paddle