Program Listing for File paddle_analysis_config.h

↰ Return to documentation for file (docs/paddle_include_file/paddle_analysis_config.h)

// 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.


#pragma once

#include <cassert>
#include <map>
#include <memory>
#include <string>
#include <unordered_set>
#include <utility>
#include <vector>

// Here we include some header files with relative paths, for that in deploy,
// the abstract path of this header file will be changed.
#include "paddle_api.h"           // NOLINT
#include "paddle_pass_builder.h"  // NOLINT
#ifdef PADDLE_WITH_MKLDNN
#include "paddle_mkldnn_quantizer_config.h"  // NOLINT
#endif

namespace paddle {

class AnalysisPredictor;
struct MkldnnQuantizerConfig;

struct AnalysisConfig {
  AnalysisConfig() = default;
  explicit AnalysisConfig(const AnalysisConfig& other);
  explicit AnalysisConfig(const std::string& model_dir);
  explicit AnalysisConfig(const std::string& prog_file,
                          const std::string& params_file);
  enum class Precision {
    kFloat32 = 0,
    kInt8,
    kHalf,
  };

  void SetModel(const std::string& model_dir) { model_dir_ = model_dir; }

  void SetModel(const std::string& prog_file_path,
                const std::string& params_file_path);
  void SetProgFile(const std::string& x) { prog_file_ = x; }
  void SetParamsFile(const std::string& x) { params_file_ = x; }

  void SetOptimCacheDir(const std::string& opt_cache_dir) {
    opt_cache_dir_ = opt_cache_dir;
  }
  const std::string& model_dir() const { return model_dir_; }
  const std::string& prog_file() const { return prog_file_; }
  const std::string& params_file() const { return params_file_; }

  // Padding related.

  void DisableFCPadding();
  bool use_fc_padding() const { return use_fc_padding_; }

  // GPU related.

  void EnableUseGpu(uint64_t memory_pool_init_size_mb, int device_id = 0);
  void DisableGpu();
  bool use_gpu() const { return use_gpu_; }
  int gpu_device_id() const { return device_id_; }
  int memory_pool_init_size_mb() const { return memory_pool_init_size_mb_; }
  float fraction_of_gpu_memory_for_pool() const;

  // CUDNN related.
  void EnableCUDNN();
  bool cudnn_enabled() const { return use_cudnn_; }

  void SwitchIrOptim(int x = true) { enable_ir_optim_ = x; }
  bool ir_optim() const { return enable_ir_optim_; }

  void SwitchUseFeedFetchOps(int x = true) { use_feed_fetch_ops_ = x; }
  bool use_feed_fetch_ops_enabled() const { return use_feed_fetch_ops_; }

  void SwitchSpecifyInputNames(bool x = true) { specify_input_name_ = x; }
  bool specify_input_name() const { return specify_input_name_; }

  void EnableTensorRtEngine(int workspace_size = 1 << 20,
                            int max_batch_size = 1, int min_subgraph_size = 3,
                            Precision precision = Precision::kFloat32,
                            bool use_static = false,
                            bool use_calib_mode = true);
  bool tensorrt_engine_enabled() const { return use_tensorrt_; }
  void SetTRTDynamicShapeInfo(
      std::map<std::string, std::vector<int>> min_input_shape,
      std::map<std::string, std::vector<int>> max_input_shape,
      std::map<std::string, std::vector<int>> optim_input_shape,
      bool disable_trt_plugin_fp16 = false);
  void EnableLiteEngine(
      AnalysisConfig::Precision precision_mode = Precision::kFloat32,
      const std::vector<std::string>& passes_filter = {},
      const std::vector<std::string>& ops_filter = {});

  bool lite_engine_enabled() const { return use_lite_; }

  void SwitchIrDebug(int x = true);

  void EnableMKLDNN();
  void SetMkldnnCacheCapacity(int capacity);
  bool mkldnn_enabled() const { return use_mkldnn_; }

  void SetCpuMathLibraryNumThreads(int cpu_math_library_num_threads);
  int cpu_math_library_num_threads() const {
    return cpu_math_library_num_threads_;
  }

  NativeConfig ToNativeConfig() const;
  void SetMKLDNNOp(std::unordered_set<std::string> op_list) {
    mkldnn_enabled_op_types_ = op_list;
  }

  void EnableMkldnnQuantizer();

  bool mkldnn_quantizer_enabled() const { return use_mkldnn_quantizer_; }

  MkldnnQuantizerConfig* mkldnn_quantizer_config() const;

  void SetModelBuffer(const char* prog_buffer, size_t prog_buffer_size,
                      const char* params_buffer, size_t params_buffer_size);
  bool model_from_memory() const { return model_from_memory_; }

  void EnableMemoryOptim();
  bool enable_memory_optim() const;

  void EnableProfile();
  bool profile_enabled() const { return with_profile_; }

  void DisableGlogInfo();
  bool glog_info_disabled() const { return !with_glog_info_; }

  void SetInValid() const { is_valid_ = false; }
  bool is_valid() const { return is_valid_; }

  friend class ::paddle::AnalysisPredictor;

  PassStrategy* pass_builder() const;
  void PartiallyRelease();

 protected:
  // Update the config.
  void Update();

  std::string SerializeInfoCache();

 protected:
  // Model pathes.
  std::string model_dir_;
  mutable std::string prog_file_;
  mutable std::string params_file_;

  // GPU related.
  bool use_gpu_{false};
  int device_id_{0};
  uint64_t memory_pool_init_size_mb_{100};  // initial size is 100MB.

  bool use_cudnn_{false};

  // Padding related
  bool use_fc_padding_{true};

  // TensorRT related.
  bool use_tensorrt_{false};
  // For workspace_size, refer it from here:
  // https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#troubleshooting
  int tensorrt_workspace_size_{1 << 30};
  // While TensorRT allows an engine optimized for a given max batch size
  // to run at any smaller size, the performance for those smaller
  // sizes may not be as well-optimized. Therefore, Max batch is best
  // equivalent to the runtime batch size.
  int tensorrt_max_batchsize_{1};
  //  We transform the Ops that can be converted into TRT layer in the model,
  //  and aggregate these Ops into subgraphs for TRT execution.
  //  We set this variable to control the minimum number of nodes in the
  //  subgraph, 3 as default value.
  int tensorrt_min_subgraph_size_{3};
  Precision tensorrt_precision_mode_{Precision::kFloat32};
  bool trt_use_static_engine_{false};
  bool trt_use_calib_mode_{true};
  std::map<std::string, std::vector<int>> min_input_shape_{};
  std::map<std::string, std::vector<int>> max_input_shape_{};
  std::map<std::string, std::vector<int>> optim_input_shape_{};
  bool disable_trt_plugin_fp16_{false};

  // memory reuse related.
  bool enable_memory_optim_{false};

  bool use_mkldnn_{false};
  std::unordered_set<std::string> mkldnn_enabled_op_types_;

  bool model_from_memory_{false};

  bool enable_ir_optim_{true};
  bool use_feed_fetch_ops_{true};
  bool ir_debug_{false};

  bool specify_input_name_{false};

  int cpu_math_library_num_threads_{1};

  bool with_profile_{false};

  bool with_glog_info_{true};

  // A runtime cache, shouldn't be transferred to others.
  std::string serialized_info_cache_;

  mutable std::unique_ptr<PassStrategy> pass_builder_;

  bool use_lite_{false};
  std::vector<std::string> lite_passes_filter_;
  std::vector<std::string> lite_ops_filter_;
  Precision lite_precision_mode_;

  // mkldnn related.
  int mkldnn_cache_capacity_{0};
  bool use_mkldnn_quantizer_{false};
  std::shared_ptr<MkldnnQuantizerConfig> mkldnn_quantizer_config_;

  // If the config is already used on a predictor, it becomes invalid.
  // Any config can only be used with one predictor.
  // Variables held by config can take up a lot of memory in some cases.
  // So we release the memory when the predictor is set up.
  mutable bool is_valid_{true};
  std::string opt_cache_dir_;
};

}  // namespace paddle