parallel_executor.cc 5.6 KB
Newer Older
Y
Yang Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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. */

#include "paddle/fluid/framework/parallel_executor.h"
C
chengduoZH 已提交
16
#include <string>
Y
Yu Yang 已提交
17

Y
Yu Yang 已提交
18
#include "ThreadPool.h"
Y
Yu Yang 已提交
19

Y
Yu Yang 已提交
20
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
21
#include "paddle/fluid/platform/nccl_helper.h"
Y
Yu Yang 已提交
22
#endif
Y
Yang Yang 已提交
23

Y
Yu Yang 已提交
24 25 26
#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"

Y
Yang Yang 已提交
27
namespace paddle {
Y
Yu Yang 已提交
28 29
namespace framework {

Y
Yu Yang 已提交
30 31 32
class ParallelExecutorPrivate {
 public:
  explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
Y
Yu Yang 已提交
33
      : places_(places) {}
Y
Yu Yang 已提交
34 35 36 37

  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
  Scope *global_scope_;
Y
Yu Yang 已提交
38
  std::unique_ptr<details::SSAGraphExecutor> executor_;
Y
Yu Yang 已提交
39

Y
Yu Yang 已提交
40
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
41
  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
Y
Yu Yang 已提交
42
#endif
Y
Yu Yang 已提交
43 44
};

Y
Yu Yang 已提交
45
ParallelExecutor::ParallelExecutor(
46 47
    size_t num_threads, bool use_event,
    const std::vector<platform::Place> &places,
Y
Yu Yang 已提交
48 49 50
    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
51
    : member_(new ParallelExecutorPrivate(places)) {
Y
Yu Yang 已提交
52
  member_->global_scope_ = scope;
Y
Yu Yang 已提交
53

Y
Yu Yang 已提交
54 55 56 57
  // Step 1. RunStartupProgram and Bcast the params to devs.
  Executor exe(places[0]);
  exe.Run(startup_program, scope, 0);
  // Create local scopes
Y
Yu Yang 已提交
58 59
  for (size_t i = 0; i < member_->places_.size(); ++i) {
    member_->local_scopes_.push_back(&scope->NewScope());
Y
Yu Yang 已提交
60 61
  }

Y
Yu Yang 已提交
62 63 64 65
// Bcast Parameters to all GPUs
#ifdef PADDLE_WITH_CUDA
  member_->nccl_ctxs_.reset(new platform::NCCLContextMap(member_->places_));
#endif
Y
Yu Yang 已提交
66
  if (platform::is_gpu_place(places[0]) &&
Y
Yu Yang 已提交
67 68
      member_->local_scopes_.size() != 1) {  // Is CUDA
    BCastParamsToGPUs(startup_program);
Y
Yu Yang 已提交
69
  }
Y
Yu Yang 已提交
70
// Startup Program has been run. All local scopes has correct parameters.
Y
Yu Yang 已提交
71

Y
Yu Yang 已提交
72 73 74
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
75 76 77
  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
                                           params, member_->local_scopes_,
                                           member_->nccl_ctxs_.get());
Y
Yu Yang 已提交
78 79 80 81
#else
  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
                                           params, member_->local_scopes_);
#endif
Y
Yu Yang 已提交
82
  auto graph = builder.Build(main_program);
Y
Yu Yang 已提交
83

Y
Yu Yang 已提交
84
  member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
85 86
      num_threads, use_event, member_->local_scopes_, places,
      std::move(graph)));
Y
Yu Yang 已提交
87

Y
Yu Yang 已提交
88
  // Step 3. Create vars in each scope;
Y
Yu Yang 已提交
89
  for (auto *scope : member_->local_scopes_) {
Y
Yu Yang 已提交
90 91 92 93 94 95 96 97
    for (auto *var : main_program.Block(0).AllVars()) {
      if (scope->FindVar(var->Name()) != nullptr) {
        continue;
      }

      InitializeVariable(scope->Var(var->Name()), var->GetType());
    }
  }
Y
Yu Yang 已提交
98 99 100 101
}

void ParallelExecutor::BCastParamsToGPUs(
    const ProgramDesc &startup_program) const {
Y
Yu Yang 已提交
102
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
103
  auto *main_scope = member_->local_scopes_[0];
Y
Yu Yang 已提交
104

Y
Yu Yang 已提交
105
  for (auto *var_desc : startup_program.Block(0).AllVars()) {
C
chengduoZH 已提交
106 107
    size_t idx = var_desc->Name().find("@GRAD");
    if (idx != std::string::npos) continue;
Y
Yu Yang 已提交
108 109 110 111
    if (var_desc->GetType() == proto::VarType::LOD_TENSOR) {
      auto &main_tensor =
          main_scope->FindVar(var_desc->Name())->Get<LoDTensor>();

C
chengduoZH 已提交
112
      auto &dims = main_tensor.dims();
Y
Yu Yang 已提交
113

C
chengduoZH 已提交
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136
      if (paddle::platform::is_gpu_place(main_tensor.place())) {
        size_t numel = main_tensor.numel();
        ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());
        platform::NCCLGroupGuard guard;
        for (size_t i = 0; i < member_->places_.size(); ++i) {
          auto place = member_->places_[i];
          void *buffer;
          if (i == 0) {
            buffer = const_cast<void *>(main_tensor.data<void>());
          } else {
            auto local_scope = member_->local_scopes_[i];
            auto *t =
                local_scope->Var(var_desc->Name())->GetMutable<LoDTensor>();
            t->Resize(dims);
            buffer = t->mutable_data(place, main_tensor.type());
          }
          auto &nccl_ctx = member_->nccl_ctxs_->at(place);
          platform::dynload::ncclBcast(buffer, numel, data_type, 0,
                                       nccl_ctx.comm_, nccl_ctx.stream());
        }
      } else {
        platform::CPUPlace cpu;
        for (size_t i = 1; i < member_->places_.size(); ++i) {
Y
Yu Yang 已提交
137
          auto local_scope = member_->local_scopes_[i];
Y
Update  
Yu Yang 已提交
138 139
          auto *t = local_scope->Var(var_desc->Name())->GetMutable<LoDTensor>();
          t->Resize(dims);
C
chengduoZH 已提交
140 141
          t->mutable_data(cpu, main_tensor.type());
          paddle::framework::TensorCopy(main_tensor, cpu, t);
Y
Update  
Yu Yang 已提交
142
        }
Y
Yu Yang 已提交
143
      }
Y
Stash  
Yu Yang 已提交
144
    }
Y
Yu Yang 已提交
145
    member_->nccl_ctxs_->WaitAll();
Y
Stash  
Yu Yang 已提交
146
  }
Y
Yu Yang 已提交
147 148 149 150
#else
  PADDLE_THROW("Not compiled with CUDA");
#endif
}
Y
Yu Yang 已提交
151

Y
Yu Yang 已提交
152 153
void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                           const std::string &fetched_var_name) {
Y
Yu Yang 已提交
154 155 156
  auto fetch_data = member_->executor_->Run(fetch_tensors);
  *member_->global_scope_->Var(fetched_var_name)->GetMutable<FeedFetchList>() =
      fetch_data;
Y
Yu Yang 已提交
157
}
Y
Yu Yang 已提交
158

Y
Yu Yang 已提交
159
}  // namespace framework
Y
Yang Yang 已提交
160
}  // namespace paddle