build_strategy.h 4.1 KB
Newer Older
Y
yuyang18 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// 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

Y
yuyang18 已提交
17
#include <string>
18 19 20 21 22 23 24 25 26 27 28
#include <vector>

#include "paddle/fluid/framework/ir/pass_builder.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"

#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/nccl_helper.h"
#endif
Y
yuyang18 已提交
29

Y
yuyang18 已提交
30 31 32 33 34
namespace paddle {
namespace framework {
namespace details {

struct BuildStrategy {
C
chengduo 已提交
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
  // ParallelExecutor supports two modes of ReduceStrategy, kAllReduce and
  // kReduce, for CPU and GPU. If you use kAllReduce, different threads
  // optimize their parameters separately. If you use kReduce, the optimizations
  // of parameters are distributed to different threads.
  // For example, a model has 100 parameters and is running with four threads,
  // if you choose kAllReduce, every thread is to optimize 100 parameters
  // separately, if you choose kReduce, every thread is to optimize 25
  // parameters.
  // Of particular note is, if you use kReduce when using CPU training,
  // all the parameters are shared between different threads. This feature will
  // save memory.
  // FIXME(zcd): The result of the two modes(kAllReduce and kReduce) maybe not
  // equal for GPU. Because, the result of the different order of summing maybe
  // different, for example, the result of `a+b+c+d` may be different with the
  // result of `c+a+b+d`.
  // For GPU, the implementation of kAllReduce and kReduce is adopted NCCL,
  // so the result of kAllReduce and kReduce maybe not equal.
  // For CPU, if you want to fix the order of summing to make the result
  // of kAllReduce and kReduce no diff, you can add
  // `FLAGS_cpu_deterministic=true` to env.
Y
yuyang18 已提交
55 56 57 58 59 60 61 62
  enum class ReduceStrategy { kAllReduce = 0, kReduce = 1 };

  enum class GradientScaleStrategy {
    kCoeffNumDevice = 0,
    kOne = 1,
    kCustomized = 2,
  };

Y
yuyang18 已提交
63
  ReduceStrategy reduce_{ReduceStrategy::kAllReduce};
Y
yuyang18 已提交
64
  GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice};
Y
yuyang18 已提交
65 66

  std::string debug_graphviz_path_{""};
F
fengjiayi 已提交
67

C
chengduo 已提交
68 69
  bool fuse_elewise_add_act_ops_{false};

Y
yuyang18 已提交
70
  bool enable_data_balance_{false};
71

S
sneaxiy 已提交
72
  bool enable_sequential_execution_{false};
S
sneaxiy 已提交
73

74 75
  bool fuse_broadcast_op_{false};

S
sneaxiy 已提交
76 77
  bool remove_unnecessary_lock_{false};

X
Xin Pan 已提交
78 79 80 81 82
  // NOTE:
  // Before you add new options, think if it's a general strategy that works
  // with other strategy. If not, the strategy should be created through
  // CreatePassesFromStrategy and the pass can be managed separately.

X
Xin Pan 已提交
83
  // User normally doesn't need to call this API.
X
Xin Pan 已提交
84
  // The PassBuilder allows for more customized insert, remove of passes
X
Xin Pan 已提交
85 86 87
  // from python side.
  // A new PassBuilder is created based on configs defined above and
  // passes are owned by the PassBuilder.
88
  std::shared_ptr<ir::PassBuilder> CreatePassesFromStrategy(
X
Xin Pan 已提交
89 90 91
      bool finalize_strategy) const;

  bool IsFinalized() const { return is_finalized_; }
92

X
Xin Pan 已提交
93 94
  // Apply the passes built by the pass_builder_. The passes will be
  // applied to the Program and output an ir::Graph.
95 96 97 98 99 100 101 102 103 104 105 106 107
  std::unique_ptr<ir::Graph> Apply(
      const ProgramDesc &main_program,
      const std::vector<platform::Place> &places,
      const std::string &loss_var_name,
      const std::unordered_set<std::string> &param_names,
      const std::vector<Scope *> &local_scopes,
#ifdef PADDLE_WITH_CUDA
      const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const;
#else
      const bool use_cuda) const;
#endif

 private:
X
Xin Pan 已提交
108
  mutable bool is_finalized_ = false;
109
  mutable std::shared_ptr<ir::PassBuilder> pass_builder_;
Y
yuyang18 已提交
110 111 112 113 114
};

}  // namespace details
}  // namespace framework
}  // namespace paddle