parallel_executor.cc 6.2 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"
X
Xin Pan 已提交
16
#include "paddle/fluid/platform/profiler.h"
Q
qiaolongfei 已提交
17

C
chengduoZH 已提交
18
#include <string>
Q
qiaolongfei 已提交
19
#include <vector>
Y
Yu Yang 已提交
20

Y
Yu Yang 已提交
21
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
22
#include "paddle/fluid/platform/nccl_helper.h"
Y
Yu Yang 已提交
23
#endif
Y
Yang Yang 已提交
24

Y
Yu Yang 已提交
25 26 27
#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
28
namespace paddle {
Y
Yu Yang 已提交
29 30
namespace framework {

Y
Yu Yang 已提交
31 32 33
class ParallelExecutorPrivate {
 public:
  explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
Y
Yu Yang 已提交
34
      : places_(places) {}
Y
Yu Yang 已提交
35 36 37 38

  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
  Scope *global_scope_;
Y
Yu Yang 已提交
39
  std::unique_ptr<details::SSAGraphExecutor> executor_;
Y
Yu Yang 已提交
40

Y
Yu Yang 已提交
41
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
42
  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
Y
Yu Yang 已提交
43
#endif
Y
Yu Yang 已提交
44 45
};

Y
Yu Yang 已提交
46
ParallelExecutor::ParallelExecutor(
47 48
    size_t num_threads, bool use_event,
    const std::vector<platform::Place> &places,
Y
Yu Yang 已提交
49 50
    const std::unordered_set<std::string> &params,
    const ProgramDesc &startup_program, const ProgramDesc &main_program,
X
Xin Pan 已提交
51
    const std::string &loss_var_name, Scope *scope, bool allow_op_delay)
Y
Yu Yang 已提交
52
    : member_(new ParallelExecutorPrivate(places)) {
Y
Yu Yang 已提交
53
  member_->global_scope_ = scope;
Y
Yu Yang 已提交
54

Y
Yu Yang 已提交
55 56 57 58
  // 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
59 60
  for (size_t i = 0; i < member_->places_.size(); ++i) {
    member_->local_scopes_.push_back(&scope->NewScope());
Y
Yu Yang 已提交
61 62
  }

Y
Yu Yang 已提交
63 64 65 66
// 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
67
  if (platform::is_gpu_place(places[0]) &&
Y
Yu Yang 已提交
68 69
      member_->local_scopes_.size() != 1) {  // Is CUDA
    BCastParamsToGPUs(startup_program);
Y
Yu Yang 已提交
70
  }
Y
Yu Yang 已提交
71
// Startup Program has been run. All local scopes has correct parameters.
Y
Yu Yang 已提交
72

Y
Yu Yang 已提交
73 74 75
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
76 77 78
  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
                                           params, member_->local_scopes_,
                                           member_->nccl_ctxs_.get());
Y
Yu Yang 已提交
79 80 81 82
#else
  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
                                           params, member_->local_scopes_);
#endif
Y
Yu Yang 已提交
83
  auto graph = builder.Build(main_program);
Y
Yu Yang 已提交
84

Y
Yu Yang 已提交
85
  member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
X
Xin Pan 已提交
86 87
      num_threads, use_event, member_->local_scopes_, places, std::move(graph),
      allow_op_delay));
Y
Yu Yang 已提交
88

Y
Yu Yang 已提交
89
  // Step 3. Create vars in each scope;
Y
Yu Yang 已提交
90
  for (auto *scope : member_->local_scopes_) {
Y
Yu Yang 已提交
91 92 93 94 95 96 97 98
    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
99 100 101 102
}

void ParallelExecutor::BCastParamsToGPUs(
    const ProgramDesc &startup_program) const {
Y
Yu Yang 已提交
103
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
104
  auto *main_scope = member_->local_scopes_[0];
Y
Yu Yang 已提交
105

Y
Yu Yang 已提交
106
  for (auto *var_desc : startup_program.Block(0).AllVars()) {
C
chengduoZH 已提交
107 108
    size_t idx = var_desc->Name().find("@GRAD");
    if (idx != std::string::npos) continue;
Y
Yu Yang 已提交
109 110 111 112
    if (var_desc->GetType() == proto::VarType::LOD_TENSOR) {
      auto &main_tensor =
          main_scope->FindVar(var_desc->Name())->Get<LoDTensor>();

C
chengduoZH 已提交
113
      auto &dims = main_tensor.dims();
Y
Yu Yang 已提交
114

C
chengduoZH 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
      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
138
          auto local_scope = member_->local_scopes_[i];
Y
Update  
Yu Yang 已提交
139 140
          auto *t = local_scope->Var(var_desc->Name())->GetMutable<LoDTensor>();
          t->Resize(dims);
C
chengduoZH 已提交
141 142
          t->mutable_data(cpu, main_tensor.type());
          paddle::framework::TensorCopy(main_tensor, cpu, t);
Y
Update  
Yu Yang 已提交
143
        }
Y
Yu Yang 已提交
144
      }
Y
Stash  
Yu Yang 已提交
145
    }
Y
Yu Yang 已提交
146
    member_->nccl_ctxs_->WaitAll();
Y
Stash  
Yu Yang 已提交
147
  }
Y
Yu Yang 已提交
148 149 150 151
#else
  PADDLE_THROW("Not compiled with CUDA");
#endif
}
Y
Yu Yang 已提交
152

X
Xin Pan 已提交
153 154 155 156
void ParallelExecutor::Run(
    const std::vector<std::string> &fetch_tensors,
    const std::string &fetched_var_name,
    const std::unordered_map<std::string, LoDTensor> &feed_tensors) {
X
Xin Pan 已提交
157
  platform::RecordBlock b(0);
X
Xin Pan 已提交
158
  SplitTensorToPlaces(feed_tensors);
Y
Yu Yang 已提交
159 160 161
  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
162
}
Y
Yu Yang 已提交
163

X
Xin Pan 已提交
164 165 166 167 168 169 170 171 172 173 174 175 176 177
void ParallelExecutor::SplitTensorToPlaces(
    const std::unordered_map<std::string, LoDTensor> &feed_tensors) {
  for (auto it : feed_tensors) {
    auto lod_tensors = it.second.SplitLoDTensor(member_->places_);
    for (size_t j = 0; j < member_->places_.size(); ++j) {
      // TODO(panxy0718): Do I need to delete this var?
      member_->local_scopes_[j]
          ->Var(it.first)
          ->GetMutable<LoDTensor>()
          ->ShareDataWith(lod_tensors[j]);
    }
  }
}

Y
Yu Yang 已提交
178
}  // namespace framework
Y
Yang Yang 已提交
179
}  // namespace paddle