parallel_executor.h 2.4 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

Y
Yu Yang 已提交
17
#include <future>
Y
Yang Yang 已提交
18 19 20 21 22 23 24 25 26 27 28 29 30
#include <unordered_set>
#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/operators/nccl/nccl_gpu_common.h"
#include "paddle/fluid/platform/device_context.h"

namespace paddle {
namespace framework {

Y
Yu Yang 已提交
31 32 33
class ParallelExecutorPrivate;
class VarHandle;
class OpHandle;
Y
Yu Yang 已提交
34
class VarHandleBase;
Y
Use mtx  
Yu Yang 已提交
35

Y
Yang Yang 已提交
36
class ParallelExecutor {
Y
Yu Yang 已提交
37
 public:
Y
Yang Yang 已提交
38
  explicit ParallelExecutor(const std::vector<platform::Place>& places,
Y
Yu Yang 已提交
39 40 41 42 43
                            const std::unordered_set<std::string>& params,
                            const ProgramDesc& startup_program,
                            const ProgramDesc& main_program,
                            const std::string& loss_var_name, Scope* scope);

Y
Yu Yang 已提交
44 45
  void Run(const std::vector<std::string>& fetch_tensors,
           const std::string& fetched_var_name = "fetched_var");
Y
Yang Yang 已提交
46 47

 private:
Y
Yu Yang 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60
  ParallelExecutorPrivate* member_;

  void BCastParamsToGPUs(const ProgramDesc& startup_program) const;

  VarHandle* GetVarHandle(const std::string& each_var_name,
                          const platform::Place& place) const;

  void GenerateVar(OpHandle* op_handle, const std::string& each_var_name,
                   const platform::Place& place) const;

  void ConstructDependencyGraph(const std::unordered_set<std::string>& params,
                                const ProgramDesc& main_program,
                                const std::string& loss_var_name) const;
Y
Yu Yang 已提交
61 62

  void BuildNCCLCommunicator() const;
Y
Yu Yang 已提交
63

Y
Yu Yang 已提交
64 65 66
  void RunOp(
      std::unordered_map<VarHandleBase*, std::atomic<bool>>& pending_vars,
      OpHandle* op) const;
Y
Yu Yang 已提交
67

68
  void PolishGraphToSupportDataHazards() const;
Y
Yang Yang 已提交
69 70 71 72
};

}  // namespace framework
}  // namespace paddle