parallel_executor.h 2.9 KB
Newer Older
Y
Yang Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* Copyright (c) 2016 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

X
Xin Pan 已提交
17
#include <string>
S
sneaxiy 已提交
18
#include <unordered_map>
Y
Yang Yang 已提交
19
#include <unordered_set>
X
Xin Pan 已提交
20
#include <vector>
21 22

#include "paddle/fluid/framework/details/build_strategy.h"
Y
yuyang18 已提交
23
#include "paddle/fluid/framework/details/execution_strategy.h"
Y
Yang Yang 已提交
24 25 26 27 28 29
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/device_context.h"
30

Y
Yancey1989 已提交
31 32 33 34
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif

Y
Yang Yang 已提交
35 36 37
namespace paddle {
namespace framework {

Y
Yu Yang 已提交
38
class ParallelExecutorPrivate;
Y
Use mtx  
Yu Yang 已提交
39

Y
yuyang18 已提交
40
using details::BuildStrategy;
Y
yuyang18 已提交
41 42
using details::ExecutionStrategy;

Y
Yang Yang 已提交
43
class ParallelExecutor {
Y
Yu Yang 已提交
44 45
  DISABLE_COPY_AND_ASSIGN(ParallelExecutor);

Y
Yu Yang 已提交
46
 public:
Y
yuyang18 已提交
47 48 49 50
  explicit ParallelExecutor(const std::vector<platform::Place> &places,
                            const std::unordered_set<std::string> &bcast_vars,
                            const std::string &loss_var_name, Scope *scope,
                            const std::vector<Scope *> &local_scopes,
Y
yuyang18 已提交
51
                            const ExecutionStrategy &exec_strategy,
X
Xin Pan 已提交
52
                            const BuildStrategy &build_strategy,
Q
Qiao Longfei 已提交
53
                            std::vector<ir::Graph *> graphs);
Y
Yu Yang 已提交
54

55 56
  ~ParallelExecutor();

Y
yuyang18 已提交
57
  std::vector<Scope *> &GetLocalScopes();
58

Y
Yu Yang 已提交
59 60 61 62 63
  /**
   * Feed tensors to local scopes. The size of tensors should be equal to the
   * size of local scopes.
   */
  void FeedTensorsIntoLocalScopes(
Y
yuyang18 已提交
64
      const std::vector<std::unordered_map<std::string, LoDTensor>> &tensors);
Y
Yu Yang 已提交
65 66

  void FeedAndSplitTensorIntoLocalScopes(
Y
yuyang18 已提交
67
      const std::unordered_map<std::string, LoDTensor> &tensors);
Y
Yu Yang 已提交
68

Y
yuyang18 已提交
69 70
  void Run(const std::vector<std::string> &fetch_tensors,
           const std::string &fetched_var_name);
Y
Yang Yang 已提交
71

X
Xin Pan 已提交
72
 private:
Y
Yancey1989 已提交
73
  void BCastParamsToDevices(const std::unordered_set<std::string> &vars) const;
X
Xin Pan 已提交
74 75 76
  bool EnableParallelGraphExecution(const ir::Graph &graph,
                                    const ExecutionStrategy &exec_strategy,
                                    const BuildStrategy &build_strategy) const;
T
typhoonzero 已提交
77

S
sneaxiy 已提交
78
  ParallelExecutorPrivate *member_;
Y
Yancey1989 已提交
79 80 81
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
  std::unique_ptr<ncclUniqueId> local_nccl_id_;
#endif
Y
Yang Yang 已提交
82 83 84 85
};

}  // namespace framework
}  // namespace paddle