parallel_executor.cc 17.5 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>
17
#include <tuple>
Q
qiaolongfei 已提交
18
#include <vector>
C
chengduo 已提交
19
#include "paddle/fluid/framework/ir/graph_helper.h"
Y
Yu Yang 已提交
20

X
clean  
Xin Pan 已提交
21
#include "paddle/fluid/framework/ir/graph.h"
X
Xin Pan 已提交
22

P
peizhilin 已提交
23
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Y
Yu Yang 已提交
24
#include "paddle/fluid/platform/nccl_helper.h"
Y
Yu Yang 已提交
25
#endif
Y
Yang Yang 已提交
26

Y
yuyang18 已提交
27
#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
28
#include "paddle/fluid/framework/details/multi_devices_helper.h"
Y
Yancey1989 已提交
29
#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
S
sneaxiy 已提交
30
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
Y
yuyang18 已提交
31
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
Y
Yu Yang 已提交
32
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
33
#include "paddle/fluid/platform/profiler.h"
Y
Yu Yang 已提交
34

Y
Yu Yang 已提交
35
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
36
#include "gperftools/profiler.h"
Y
Yu Yang 已提交
37
#endif
Y
Yu Yang 已提交
38
DEFINE_string(pe_profile_fname, "",
Y
Yu Yang 已提交
39 40 41
              "Profiler filename for PE, which generated by gperftools."
              "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable.");

Y
Yang Yang 已提交
42
namespace paddle {
Y
Yu Yang 已提交
43 44
namespace framework {

Y
Yu Yang 已提交
45
static std::once_flag gProfileOnce;
Y
Yu Yang 已提交
46
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
47
static bool gProfileStarted = false;
Y
Yu Yang 已提交
48
#endif
Y
Yu Yang 已提交
49 50 51
class ParallelExecutorPrivate {
 public:
  explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
Y
Yu Yang 已提交
52
      : places_(places) {
Y
Yu Yang 已提交
53
    if (!FLAGS_pe_profile_fname.empty()) {
Y
Yu Yang 已提交
54 55
      std::call_once(gProfileOnce, [] {
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
56
        ProfilerStart(FLAGS_pe_profile_fname.c_str());
Y
Yu Yang 已提交
57 58 59
        gProfileStarted = true;
#else
        LOG(WARNING) << "Paddle is not compiled with gperftools. "
Y
Yu Yang 已提交
60
                        "FLAGS_pe_profile_fname will be ignored";
Y
Yu Yang 已提交
61 62 63 64
#endif
      });
    }
  }
Y
Yu Yang 已提交
65

66 67 68 69 70 71 72 73 74 75 76
  ~ParallelExecutorPrivate() {
    if (own_local_scope_) {
      for (size_t i = 1; i < local_scopes_.size(); ++i) {
        // Skip the first scope, since it is the global scope.
        Scope *local_scope = local_scopes_[i];
        if (global_scope_->HasKid(local_scope)) {
          global_scope_->DeleteScope(local_scope);
        }
      }
    }
  }
S
sneaxiy 已提交
77

S
sneaxiy 已提交
78 79 80 81 82 83 84 85 86 87 88 89 90 91
  std::unique_ptr<ir::Graph> PrepareGCAndRefCnts(
      std::unique_ptr<ir::Graph> graph, size_t max_memory_size);

  inline bool HasGarbageCollectors() const { return !gcs_.empty(); }

  void ResetRuntimeReferenceCount(const std::vector<std::string> &fetch_tensors,
                                  const std::string &fetched_var_name) {
    for (size_t i = 0; i < runtime_ref_cnts_.size(); ++i) {
      for (auto &pair : global_ref_cnts_[i]) {
        runtime_ref_cnts_[i][pair.first] = pair.second;
      }

      for (auto &fetch_name : fetch_tensors) {
        runtime_ref_cnts_[i].erase(fetch_name);
S
sneaxiy 已提交
92
      }
S
sneaxiy 已提交
93
      runtime_ref_cnts_[i].erase(fetched_var_name);
S
sneaxiy 已提交
94 95 96
    }
  }

Y
Yu Yang 已提交
97 98
  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
99
  Scope *global_scope_;  // not owned
Y
Yu Yang 已提交
100
  std::unique_ptr<details::SSAGraphExecutor> executor_;
Y
Yu Yang 已提交
101

P
peizhilin 已提交
102
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Y
Yu Yang 已提交
103
  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
Y
Yu Yang 已提交
104
#endif
C
chengduoZH 已提交
105 106
  bool own_local_scope_;
  bool use_cuda_;
107
  bool use_all_reduce_;
S
sneaxiy 已提交
108

S
sneaxiy 已提交
109 110 111 112 113 114
  // global_ref_cnts_ is only initialized when ParallelExecutor constructs, and
  // then keeps unchanged
  // Before each iteration, runtime_ref_cnts_ is reset to global_ref_cnts_
  std::vector<details::ReferenceCountMap> global_ref_cnts_;
  std::vector<details::AtomicReferenceCountMap> runtime_ref_cnts_;
  details::GarbageCollectorMap gcs_;
Y
Yu Yang 已提交
115 116
};

S
sneaxiy 已提交
117 118 119 120 121 122 123
std::unique_ptr<ir::Graph> ParallelExecutorPrivate::PrepareGCAndRefCnts(
    std::unique_ptr<ir::Graph> graph, size_t max_memory_size) {
  for (size_t i = 0; i < places_.size(); ++i) {
    auto &place = places_[i];
    if (gcs_.count(place) > 0) {
      continue;
    }
S
sneaxiy 已提交
124
    std::unique_ptr<GarbageCollector> gc;
S
sneaxiy 已提交
125
#ifdef PADDLE_WITH_CUDA
S
sneaxiy 已提交
126 127
    if (platform::is_gpu_place(place)) {
      if (IsFastEagerDeletionModeEnabled()) {
S
sneaxiy 已提交
128 129
        gc.reset(new UnsafeFastGPUGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
130
      } else {
S
sneaxiy 已提交
131 132
        gc.reset(new StreamGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
133 134
      }
      VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
S
sneaxiy 已提交
135
    } else {
S
sneaxiy 已提交
136
#endif
S
sneaxiy 已提交
137 138 139 140 141 142 143
      if (platform::is_cpu_place(place)) {
        gc.reset(new CPUGarbageCollector(boost::get<platform::CPUPlace>(place),
                                         max_memory_size));
        VLOG(10) << "Created GarbageCollector at " << place;
      } else {
        PADDLE_THROW("Unsupported place for garbage collection");
      }
S
sneaxiy 已提交
144 145 146 147
#ifdef PADDLE_WITH_CUDA
    }
#endif

S
sneaxiy 已提交
148
    gcs_.emplace(place, std::move(gc));
S
sneaxiy 已提交
149 150
  }

S
sneaxiy 已提交
151
  if (!gcs_.empty()) {
S
sneaxiy 已提交
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177
    std::vector<details::LastLiveOpsOfVars> last_live_ops_of_vars;

    auto ref_cnt_pass =
        ir::PassRegistry::Instance().Get("reference_count_pass");
    ref_cnt_pass->SetNotOwned(details::kGlobalReferenceCount,
                              &global_ref_cnts_);
    ref_cnt_pass->SetNotOwned(details::kLastLiveOpsOfVars,
                              &last_live_ops_of_vars);
    graph = ref_cnt_pass->Apply(std::move(graph));
    VLOG(10) << "ReferenceCountPass Applied";

    auto eager_deletion_pass =
        ir::PassRegistry::Instance().Get("eager_deletion_pass");
    eager_deletion_pass->SetNotOwned(details::kRuntimeReferenceCount,
                                     &runtime_ref_cnts_);
    eager_deletion_pass->SetNotOwned(details::kGarbageCollector, &gcs_);
    eager_deletion_pass->SetNotOwned(details::kLastLiveOpsOfVars,
                                     &last_live_ops_of_vars);
    eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_);
    graph = eager_deletion_pass->Apply(std::move(graph));
    VLOG(10) << "EagerDeletionPass Applied";
  }

  return graph;
}

178 179 180 181
std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
  return member_->local_scopes_;
}

Y
Yu Yang 已提交
182
ParallelExecutor::ParallelExecutor(
183
    const std::vector<platform::Place> &places,
Y
Yu Yang 已提交
184
    const std::unordered_set<std::string> &params,
185 186
    const std::unordered_set<std::string> &bcast_vars,
    const ProgramDesc &main_program, const std::string &loss_var_name,
Y
yuyang18 已提交
187
    Scope *scope, const std::vector<Scope *> &local_scopes,
188
    const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy,
189
    size_t num_trainers, size_t trainer_id)
Y
Yu Yang 已提交
190
    : member_(new ParallelExecutorPrivate(places)) {
Y
Yu Yang 已提交
191
  member_->global_scope_ = scope;
192
  member_->use_cuda_ = exec_strategy.use_cuda_;
193 194 195 196 197 198 199
  member_->use_all_reduce_ =
      build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce;

  if (!member_->use_all_reduce_) {
    PADDLE_ENFORCE(places.size() > 1,
                   "If you set build_strategy.reduce with 'Reduce',"
                   "the number of places must be greater than 1.");
Y
Yancey1989 已提交
200 201 202 203 204 205 206 207 208 209 210
  }

  if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
    PADDLE_ENFORCE(
        member_->use_all_reduce_,
        "build_strategy.reduce should be `AllReduce` if you want to use"
        "ParallelGraph executor.");
    PADDLE_ENFORCE(
        member_->use_cuda_,
        "execution_strategy.use_cuda should be True if you want to use"
        "ParallelGraph executor.");
211
  }
Y
Yu Yang 已提交
212

213
  // Step 1. Bcast the params to devs.
Y
Yu Yang 已提交
214
  // Create local scopes
215
  if (local_scopes.empty()) {
C
chengduoZH 已提交
216
    member_->own_local_scope_ = true;
Y
Yu Yang 已提交
217 218
    member_->local_scopes_.emplace_back(member_->global_scope_);
    for (size_t i = 1; i < member_->places_.size(); ++i) {
Y
Debug  
Yu Yang 已提交
219
      member_->local_scopes_.emplace_back(&scope->NewScope());
220 221
    }
  } else {
C
chengduoZH 已提交
222
    member_->own_local_scope_ = false;
223 224
    PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
    for (size_t i = 0; i < member_->places_.size(); ++i) {
225
      member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
226
    }
Y
Yu Yang 已提交
227 228
  }

C
chengduoZH 已提交
229
  if (member_->use_cuda_) {
Y
Yu Yang 已提交
230
// Bcast Parameters to all GPUs
P
peizhilin 已提交
231
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
C
chengduoZH 已提交
232
    auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
Y
Yancey1989 已提交
233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
    ncclUniqueId *nccl_id = nullptr;
    if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
      // parallel graph mode should initialize nccl by ncclCommInitRank since
      // it call nccl operator per device per thread.
      if (nccl_id_var == nullptr) {
        nccl_id = new ncclUniqueId();
        PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(nccl_id));
        *member_->global_scope_->Var(NCCL_ID_VARNAME)
             ->GetMutable<ncclUniqueId>() = *nccl_id;
      } else {
        nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
      }
    } else if (nccl_id_var != nullptr) {  // the other executor type.
      // the distributed training with nccl mode would initialize the nccl id in
      // startup_program.
      nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
Y
Yancey1989 已提交
249
    } else {
Y
Yancey1989 已提交
250
      // initlize NCCL by ncclCommInitAll, do not need to intialize the nccl_id.
C
chengduoZH 已提交
251
    }
Y
Yancey1989 已提交
252

C
chengduoZH 已提交
253
    member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
Y
Yancey1989 已提交
254
        member_->places_, nccl_id, num_trainers, trainer_id));
C
chengduoZH 已提交
255 256
#else
    PADDLE_THROW("Not compiled with CUDA");
Y
Yu Yang 已提交
257
#endif
C
chengduoZH 已提交
258 259
  }
  if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
Y
Yancey1989 已提交
260
    BCastParamsToDevices(bcast_vars);
Y
Yu Yang 已提交
261
  }
Y
Yancey1989 已提交
262
  // Startup Program has been run. All local scopes has correct parameters.
Y
yuyang18 已提交
263

Y
Yancey1989 已提交
264 265 266
  // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
  // ncclOp
  std::vector<std::unique_ptr<ir::Graph>> graphs;
P
peizhilin 已提交
267
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Y
Yancey1989 已提交
268 269 270 271 272 273 274 275 276 277 278 279 280 281
  if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
    for (size_t i = 0; i < member_->places_.size(); ++i) {
      std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
          main_program, {member_->places_[i]}, loss_var_name, params,
          {member_->local_scopes_[i]}, member_->use_cuda_,
          member_->nccl_ctxs_.get());
      graphs.push_back(std::move(graph));
    }
  } else {
    std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
        main_program, member_->places_, loss_var_name, params,
        member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get());
    graphs.push_back(std::move(graph));
  }
C
chengduoZH 已提交
282
#else
283 284 285
  std::unique_ptr<ir::Graph> graph =
      build_strategy.Apply(main_program, member_->places_, loss_var_name,
                           params, member_->local_scopes_, member_->use_cuda_);
Y
Yancey1989 已提交
286
  graphs.push_back(std::move(graph));
Y
Yu Yang 已提交
287
#endif
X
Xin Pan 已提交
288

Y
Yancey1989 已提交
289 290 291 292 293 294 295 296
  auto max_memory_size = GetEagerDeletionThreshold();
  // TODO(Yancey1989): fix gc failed on ParallelGraph executor.
  if (max_memory_size >= 0 &&
      exec_strategy.type_ != ExecutionStrategy::kParallelGraph) {
    graphs[0] = member_->PrepareGCAndRefCnts(
        std::move(graphs[0]), static_cast<size_t>(max_memory_size));
  }

297 298
  // Step 3. Create vars in each scope. Passes may also create new vars.
  //         skip control vars and empty vars
Y
Yancey1989 已提交
299 300 301 302 303 304 305 306 307 308 309
  std::vector<details::VariableInfo> var_infos;
  for (auto &graph : graphs) {
    for (auto &node : graph->Nodes()) {
      if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
        var_infos.emplace_back();
        var_infos.back().name_ = node->Var()->Name();
        var_infos.back().type_ = node->Var()->GetType();
        var_infos.back().persistable_ = node->Var()->Persistable();
      }
    }
  }
Y
Yancey1989 已提交
310

W
Wu Yi 已提交
311 312
  // If the loss_var_name is given, the number of graph should be only one.
  if (loss_var_name.size()) {
Y
Yancey1989 已提交
313
    size_t graph_num = ir::GraphNum(*graphs[0]);
C
chengduo 已提交
314 315 316 317
    if (graph_num > 1) {
      LOG(WARNING)
          << "The number of graph should be only one, "
             "but the current graph has "
Y
Yancey1989 已提交
318
          << ir::GraphNum(*graphs[0])
C
chengduo 已提交
319 320 321 322 323
          << " sub_graphs. If you want to see the nodes of the "
             "sub_graphs, you should use 'FLAGS_print_sub_graph_dir' "
             "to specify the output dir. NOTES: if you not do training, "
             "please don't pass loss_var_name.";
    }
W
Wu Yi 已提交
324 325
  }

Y
yuyang18 已提交
326 327
  if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
    member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
Y
Yancey1989 已提交
328 329
        exec_strategy, member_->local_scopes_, member_->places_,
        std::move(graphs[0])));
Y
Yancey1989 已提交
330 331
  } else if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
    member_->executor_.reset(new details::ParallelSSAGraphExecutor(
Y
Yancey1989 已提交
332 333
        exec_strategy, member_->local_scopes_, member_->places_,
        std::move(graphs)));
Y
yuyang18 已提交
334 335
  } else {
    member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
Y
Yancey1989 已提交
336 337
        exec_strategy, member_->local_scopes_, member_->places_,
        std::move(graphs[0])));
C
chengduoZH 已提交
338
  }
Y
yuyang18 已提交
339 340

  member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
Y
Yancey1989 已提交
341
      exec_strategy, member_->local_scopes_, std::move(var_infos),
Y
yuyang18 已提交
342
      member_->places_, std::move(member_->executor_)));
Y
Yu Yang 已提交
343 344
}

Y
Yancey1989 已提交
345
void ParallelExecutor::BCastParamsToDevices(
346
    const std::unordered_set<std::string> &vars) const {
X
Xin Pan 已提交
347
  // the initializing bcast, all vars would be bcast from device(0).
348
  for (auto &var : vars) {
X
Xin Pan 已提交
349
    framework::Variable *main_var = member_->local_scopes_[0]->FindVar(var);
J
JiayiFeng 已提交
350
    if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
351 352 353 354
      continue;
    }

    auto &main_tensor = main_var->Get<LoDTensor>();
355
    if (!main_tensor.IsInitialized()) {
M
minqiyang 已提交
356
      VLOG(3) << "one in var not inited, return!";
357 358
      continue;
    }
359 360
    auto &dims = main_tensor.dims();
    if (paddle::platform::is_gpu_place(main_tensor.place())) {
P
peizhilin 已提交
361
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
362
      std::vector<void *> buffers;
363 364 365 366 367
      size_t numel = main_tensor.numel();
      ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());
      for (size_t i = 0; i < member_->places_.size(); ++i) {
        auto place = member_->places_[i];
        void *buffer;
368

X
Xin Pan 已提交
369
        if (i == 0) {
370 371
          buffer = const_cast<void *>(main_tensor.data<void>());
        } else {
Y
Yu Yang 已提交
372
          auto local_scope = member_->local_scopes_[i];
373
          auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
Y
Update  
Yu Yang 已提交
374
          t->Resize(dims);
375
          buffer = t->mutable_data(place, main_tensor.type());
Y
Update  
Yu Yang 已提交
376
        }
377
        buffers.push_back(buffer);
378
      }
379

380 381 382 383 384 385
      PADDLE_ENFORCE_EQ(member_->places_.size(), buffers.size(),
                        "variables' buffer size to bcast NOT equal to places");
      {
        platform::NCCLGroupGuard guard;
        for (size_t i = 0; i < member_->places_.size(); ++i) {
          auto &nccl_ctx = member_->nccl_ctxs_->at(member_->places_[i]);
X
Xin Pan 已提交
386 387
          platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
                                       nccl_ctx.comm_, nccl_ctx.stream());
388
        }
389
        member_->nccl_ctxs_->WaitAll();
390
      }
C
chengduoZH 已提交
391 392 393
#else
      PADDLE_THROW("Not compiled with CUDA");
#endif
394 395
    } else {
      platform::CPUPlace cpu;
Y
Yancey1989 已提交
396
      for (size_t i = 0; i < member_->places_.size(); ++i) {
X
Xin Pan 已提交
397
        if (i == 0) continue;
Y
Yancey1989 已提交
398

399 400
        auto local_scope = member_->local_scopes_[i];
        auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
C
chengduo 已提交
401 402 403 404

        // FIXME(zcd): LR_DECAY_COUNTER should not be shared. This is a hot fix.
        if (member_->use_all_reduce_ || member_->use_cuda_ ||
            var == "@LR_DECAY_COUNTER@") {
405 406 407 408 409 410
          t->Resize(dims);
          t->mutable_data(cpu, main_tensor.type());
          paddle::framework::TensorCopy(main_tensor, cpu, t);
        } else {
          t->ShareDataWith(main_tensor);
        }
Y
Yu Yang 已提交
411
      }
Y
Stash  
Yu Yang 已提交
412 413
    }
  }
Y
Yu Yang 已提交
414
}
Y
Yu Yang 已提交
415

Y
Yu Yang 已提交
416 417
void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                           const std::string &fetched_var_name) {
Y
Yu Yang 已提交
418 419 420
#ifdef WITH_GPERFTOOLS
  if (gProfileStarted) {
    ProfilerFlush();
S
sneaxiy 已提交
421 422
  }
#endif
Y
Yu Yang 已提交
423

X
Xin Pan 已提交
424
  platform::RecordBlock b(0);
S
sneaxiy 已提交
425 426
  if (member_->HasGarbageCollectors()) {
    member_->ResetRuntimeReferenceCount(fetch_tensors, fetched_var_name);
S
sneaxiy 已提交
427
  }
S
sneaxiy 已提交
428 429 430
  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
431
}
Y
Yu Yang 已提交
432

Y
Yu Yang 已提交
433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
void ParallelExecutor::FeedTensorsIntoLocalScopes(
    const std::vector<std::unordered_map<std::string, LoDTensor>> &tensors) {
  PADDLE_ENFORCE_EQ(member_->local_scopes_.size(), tensors.size());

  for (size_t i = 0; i < tensors.size(); ++i) {
    auto &map = tensors[i];
    auto *scope = member_->local_scopes_[i];
    for (auto &pair : map) {
      auto *trg = scope->Var(pair.first)->GetMutable<LoDTensor>();
      trg->ShareDataWith(pair.second);
      trg->set_lod(pair.second.lod());
    }
  }
}

void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
    const std::unordered_map<std::string, LoDTensor> &tensors) {
  for (auto pair : tensors) {
    auto lod_tensors = pair.second.SplitLoDTensor(member_->places_);
452 453 454 455 456
    PADDLE_ENFORCE_EQ(
        member_->places_.size(), lod_tensors.size(),
        "The number of samples of current batch is less than the count of "
        "devices, currently, it is not allowed. (%d vs %d)",
        member_->places_.size(), lod_tensors.size());
X
Xin Pan 已提交
457 458
    for (size_t j = 0; j < member_->places_.size(); ++j) {
      // TODO(panxy0718): Do I need to delete this var?
459
      auto t =
Y
Yu Yang 已提交
460
          member_->local_scopes_[j]->Var(pair.first)->GetMutable<LoDTensor>();
461 462
      t->ShareDataWith(lod_tensors[j]);
      t->set_lod(lod_tensors[j].lod());
X
Xin Pan 已提交
463 464 465 466
    }
  }
}

467
ParallelExecutor::~ParallelExecutor() {
468 469
  for (auto &p : member_->places_) {
    platform::DeviceContextPool::Instance().Get(p)->Wait();
C
chengduozh 已提交
470
  }
S
sneaxiy 已提交
471
  delete member_;
472 473
}

Y
Yu Yang 已提交
474
}  // namespace framework
Y
Yang Yang 已提交
475
}  // namespace paddle
S
sneaxiy 已提交
476

S
sneaxiy 已提交
477
USE_PASS(reference_count_pass);
S
sneaxiy 已提交
478
USE_PASS(eager_deletion_pass);