parallel_executor.cc 20.9 KB
Newer Older
Y
Yang Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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"
D
dzhwinter 已提交
16
#include <algorithm>
C
chengduoZH 已提交
17
#include <string>
18
#include <tuple>
Q
Qiao Longfei 已提交
19
#include <utility>
Q
qiaolongfei 已提交
20
#include <vector>
C
chengduo 已提交
21
#include "paddle/fluid/framework/ir/graph_helper.h"
Y
Yu Yang 已提交
22

X
clean  
Xin Pan 已提交
23
#include "paddle/fluid/framework/ir/graph.h"
X
Xin Pan 已提交
24

Y
Yancey1989 已提交
25
#include "paddle/fluid/framework/details/all_reduce_deps_pass.h"
Q
Qiao Longfei 已提交
26
#include "paddle/fluid/framework/details/async_ssa_graph_executor.h"
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
              "Profiler filename for PE, which generated by gperftools."
              "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable.");
41
DEFINE_bool(enable_parallel_graph, false,
Y
Yancey1989 已提交
42
            "Force disable parallel graph execution mode if set false.");
Y
Yu Yang 已提交
43

Y
Yang Yang 已提交
44
namespace paddle {
Y
Yu Yang 已提交
45 46
namespace framework {

Y
Yu Yang 已提交
47
static std::once_flag gProfileOnce;
Y
Yu Yang 已提交
48
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
49
static bool gProfileStarted = false;
Y
Yu Yang 已提交
50
#endif
Y
Yu Yang 已提交
51 52 53
class ParallelExecutorPrivate {
 public:
  explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
Y
Yu Yang 已提交
54
      : places_(places) {
Y
Yu Yang 已提交
55
    if (!FLAGS_pe_profile_fname.empty()) {
Y
Yu Yang 已提交
56 57
      std::call_once(gProfileOnce, [] {
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
58
        ProfilerStart(FLAGS_pe_profile_fname.c_str());
Y
Yu Yang 已提交
59 60 61
        gProfileStarted = true;
#else
        LOG(WARNING) << "Paddle is not compiled with gperftools. "
Y
Yu Yang 已提交
62
                        "FLAGS_pe_profile_fname will be ignored";
Y
Yu Yang 已提交
63 64 65 66
#endif
      });
    }
  }
Y
Yu Yang 已提交
67

68 69 70 71 72 73 74 75 76 77 78
  ~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
79

S
sneaxiy 已提交
80 81 82 83 84 85 86 87 88 89 90 91 92 93
  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
94
      }
S
sneaxiy 已提交
95
      runtime_ref_cnts_[i].erase(fetched_var_name);
S
sneaxiy 已提交
96 97 98
    }
  }

D
dzhwinter 已提交
99
  BuildStrategy build_strategy_;
Y
Yu Yang 已提交
100 101
  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
102
  Scope *global_scope_;  // not owned
Y
Yu Yang 已提交
103
  std::unique_ptr<details::SSAGraphExecutor> executor_;
Y
Yu Yang 已提交
104

P
peizhilin 已提交
105
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Y
Yu Yang 已提交
106
  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
Y
Yu Yang 已提交
107
#endif
C
chengduoZH 已提交
108 109
  bool own_local_scope_;
  bool use_cuda_;
110
  bool use_all_reduce_;
111
  size_t nranks_;
S
sneaxiy 已提交
112

S
sneaxiy 已提交
113 114 115 116 117 118
  // 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
119 120
};

S
sneaxiy 已提交
121 122 123 124 125 126 127
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
128
    std::unique_ptr<GarbageCollector> gc;
S
sneaxiy 已提交
129
#ifdef PADDLE_WITH_CUDA
S
sneaxiy 已提交
130 131
    if (platform::is_gpu_place(place)) {
      if (IsFastEagerDeletionModeEnabled()) {
S
sneaxiy 已提交
132 133
        gc.reset(new UnsafeFastGPUGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
134
      } else {
S
sneaxiy 已提交
135 136
        gc.reset(new StreamGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
137 138
      }
      VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
S
sneaxiy 已提交
139
    } else {
S
sneaxiy 已提交
140
#endif
S
sneaxiy 已提交
141 142 143 144 145 146 147
      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
148 149 150 151
#ifdef PADDLE_WITH_CUDA
    }
#endif

S
sneaxiy 已提交
152
    gcs_.emplace(place, std::move(gc));
S
sneaxiy 已提交
153 154
  }

S
sneaxiy 已提交
155
  if (!gcs_.empty()) {
S
sneaxiy 已提交
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
    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;
}

182 183 184 185
std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
  return member_->local_scopes_;
}

Y
Yu Yang 已提交
186
ParallelExecutor::ParallelExecutor(
187
    const std::vector<platform::Place> &places,
188
    const std::unordered_set<std::string> &bcast_vars,
X
Xin Pan 已提交
189 190 191
    const std::string &loss_var_name, Scope *scope,
    const std::vector<Scope *> &local_scopes,
    const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy,
Q
Qiao Longfei 已提交
192
    ir::Graph *graph)
Y
Yu Yang 已提交
193
    : member_(new ParallelExecutorPrivate(places)) {
Y
Yu Yang 已提交
194
  member_->global_scope_ = scope;
195
  member_->use_cuda_ = exec_strategy.use_cuda_;
D
dzhwinter 已提交
196
  member_->build_strategy_ = build_strategy;
197 198
  member_->use_all_reduce_ =
      build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce;
X
Xin Pan 已提交
199
  member_->nranks_ = build_strategy.num_trainers_ * places.size();
200 201 202 203
  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
204 205
  }

206
  // Step 1. Bcast the bcast_vars to devs.
Y
Yu Yang 已提交
207
  // Create local scopes
208
  if (local_scopes.empty()) {
C
chengduoZH 已提交
209
    member_->own_local_scope_ = true;
Y
Yu Yang 已提交
210 211
    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
212
      member_->local_scopes_.emplace_back(&scope->NewScope());
213 214
    }
  } else {
C
chengduoZH 已提交
215
    member_->own_local_scope_ = false;
216 217
    PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
    for (size_t i = 0; i < member_->places_.size(); ++i) {
218
      member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
219
    }
Y
Yu Yang 已提交
220 221
  }

Q
Qiao Longfei 已提交
222
  std::vector<ir::Graph *> graphs;
Q
Qiao Longfei 已提交
223 224 225
  if (build_strategy.async_mode_) {
    PADDLE_ENFORCE(!member_->use_cuda_,
                   "gpu mode does not support async_mode_ now!");
Q
Qiao Longfei 已提交
226 227 228 229 230 231
    graphs.push_back(graph);
    for (int i = 1; i < places.size(); ++i) {
      auto *tmp_graph = new ir::Graph(graph->OriginProgram());
      async_graphs_.emplace_back(tmp_graph);
      graphs.push_back(tmp_graph);
    }
Q
Qiao Longfei 已提交
232
  }
Q
Qiao Longfei 已提交
233

X
Xin Pan 已提交
234
  std::unique_ptr<ir::Graph> temp_owned_graph(graph);
Q
Qiao Longfei 已提交
235

Y
Yancey1989 已提交
236 237 238
  // FIXME(Yancey1989): parallel graph mode get better performance
  // in GPU allreduce distributed training. Need an elegant way to
  // choice the execution strategy.
X
Xin Pan 已提交
239 240
  build_strategy.enable_parallel_graph_ = EnableParallelGraphExecution(
      *temp_owned_graph, exec_strategy, build_strategy);
Y
Yancey1989 已提交
241 242 243 244
  if (build_strategy.enable_parallel_graph_)
    VLOG(0) << "The Executor would execute the graph by ParallelGraph "
               "Execution which can get better performance,"
            << "you can force it off by env FLAGS_enable_parallel_graph=0";
Y
Yancey1989 已提交
245

C
chengduoZH 已提交
246
  if (member_->use_cuda_) {
Y
Yu Yang 已提交
247
// Bcast Parameters to all GPUs
P
peizhilin 已提交
248
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Y
Yancey1989 已提交
249 250 251
    ncclUniqueId *nccl_id = nullptr;
    // gen_nccl_id operator can broadcast the ncclUniqueId for nccl2 collective
    // distributed training
C
chengduoZH 已提交
252
    auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
Y
Yancey1989 已提交
253
    if (nccl_id_var != nullptr) {
Y
Yancey1989 已提交
254
      nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
Y
Yancey1989 已提交
255
    }
256
    if (build_strategy.enable_parallel_graph_ && member_->nranks_ > 1UL) {
Y
Yancey1989 已提交
257 258 259 260
      if (nccl_id == nullptr) {
        local_nccl_id_.reset(new ncclUniqueId());
        platform::dynload::ncclGetUniqueId(local_nccl_id_.get());
        nccl_id = local_nccl_id_.get();
Y
Yancey1989 已提交
261
      }
C
chengduoZH 已提交
262
    }
Y
Yancey1989 已提交
263

C
chengduoZH 已提交
264
    member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
265 266
        member_->places_, nccl_id, build_strategy.num_trainers_,
        build_strategy.trainer_id_));
C
chengduoZH 已提交
267 268
#else
    PADDLE_THROW("Not compiled with CUDA");
Y
Yu Yang 已提交
269
#endif
C
chengduoZH 已提交
270 271
  }
  if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
Y
Yancey1989 已提交
272
    BCastParamsToDevices(bcast_vars);
Y
Yu Yang 已提交
273
  }
Q
Qiao Longfei 已提交
274
  // Startup Program has been run. All local scopes has correct parameters.
Y
yuyang18 已提交
275

Q
Qiao Longfei 已提交
276 277 278
  // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
  // ncclOp
  std::vector<ir::Graph *> async_graphs(places.size());
P
peizhilin 已提交
279
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Q
Qiao Longfei 已提交
280
  if (build_strategy.async_mode_) {
Q
Qiao Longfei 已提交
281
    VLOG(3) << "use local async mode";
Q
Qiao Longfei 已提交
282 283 284 285 286 287 288 289 290 291 292 293
    temp_owned_graph =
        build_strategy.Apply(std::move(temp_owned_graph), {member_->places_[0]},
                             loss_var_name, {member_->local_scopes_[0]}, 1,
                             member_->use_cuda_, member_->nccl_ctxs_.get());
    for (int i = 1; i < member_->places_.size(); ++i) {
      std::unique_ptr<ir::Graph> temp_graph(graphs[i]);
      temp_graph =
          build_strategy.Apply(std::move(temp_graph), {member_->places_[i]},
                               loss_var_name, {member_->local_scopes_[i]}, 1,
                               member_->use_cuda_, member_->nccl_ctxs_.get());
      async_graphs[i] = temp_graph.release();
    }
Q
Qiao Longfei 已提交
294
  } else {
Q
Qiao Longfei 已提交
295 296 297 298
    temp_owned_graph = build_strategy.Apply(
        std::move(temp_owned_graph), member_->places_, loss_var_name,
        member_->local_scopes_, member_->nranks_, member_->use_cuda_,
        member_->nccl_ctxs_.get());
Q
Qiao Longfei 已提交
299
  }
C
chengduoZH 已提交
300
#else
Q
Qiao Longfei 已提交
301
  if (build_strategy.async_mode_) {
Q
Qiao Longfei 已提交
302
    VLOG(3) << "use local async mode";
Q
Qiao Longfei 已提交
303 304
    temp_owned_graph = build_strategy.Apply(
        std::move(temp_owned_graph), {member_->places_[0]}, loss_var_name,
Q
Qiao Longfei 已提交
305 306 307 308 309 310 311 312
        {member_->local_scopes_[0]}, 1, member_->use_cuda_);
    for (int i = 1; i < member_->places_.size(); ++i) {
      std::unique_ptr<ir::Graph> temp_graph(graphs[i]);
      temp_graph = build_strategy.Apply(
          std::move(temp_graph), {member_->places_[i]}, loss_var_name,
          {member_->local_scopes_[i]}, 1, member_->use_cuda_);
      async_graphs[i] = temp_graph.release();
    }
Q
can run  
Qiao Longfei 已提交
313
  } else {
Q
Qiao Longfei 已提交
314 315 316
    temp_owned_graph = build_strategy.Apply(
        std::move(temp_owned_graph), member_->places_, loss_var_name,
        member_->local_scopes_, member_->nranks_, member_->use_cuda_);
Q
can run  
Qiao Longfei 已提交
317
  }
X
Xin Pan 已提交
318

Y
Yu Yang 已提交
319
#endif
Y
Yancey1989 已提交
320
  auto max_memory_size = GetEagerDeletionThreshold();
D
dzhwinter 已提交
321 322
  VLOG(10) << "Eager Deletion Threshold "
           << static_cast<float>(max_memory_size) / (1 << 30);
Y
Yancey1989 已提交
323
  if (max_memory_size >= 0) {
X
Xin Pan 已提交
324 325 326 327 328 329
    graph = member_
                ->PrepareGCAndRefCnts(std::move(temp_owned_graph),
                                      static_cast<size_t>(max_memory_size))
                .release();
  } else {
    graph = temp_owned_graph.release();
Y
Yancey1989 已提交
330 331
  }

Q
Qiao Longfei 已提交
332 333
  async_graphs[0] = graph;

334 335
  // Step 3. Create vars in each scope. Passes may also create new vars.
  //         skip control vars and empty vars
Y
Yancey1989 已提交
336
  std::vector<details::VariableInfo> var_infos;
Q
Qiao Longfei 已提交
337 338 339 340 341 342
  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
343 344
    }
  }
Y
Yancey1989 已提交
345

W
Wu Yi 已提交
346 347
  // If the loss_var_name is given, the number of graph should be only one.
  if (loss_var_name.size()) {
Q
Qiao Longfei 已提交
348
    size_t graph_num = ir::GraphNum(*graph);
C
chengduo 已提交
349 350 351 352
    if (graph_num > 1) {
      LOG(WARNING)
          << "The number of graph should be only one, "
             "but the current graph has "
Q
Qiao Longfei 已提交
353
          << ir::GraphNum(*graph)
C
chengduo 已提交
354 355 356 357 358
          << " 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
359 360
  }

Q
Qiao Longfei 已提交
361
  if (build_strategy.async_mode_) {
Q
can run  
Qiao Longfei 已提交
362 363
    VLOG(3) << "use AsyncSSAGraphExecutor";
    member_->executor_.reset(new details::AsyncSSAGraphExecutor(
Q
Qiao Longfei 已提交
364
        exec_strategy, member_->local_scopes_, member_->places_, async_graphs));
Q
can run  
Qiao Longfei 已提交
365 366
  } else if (build_strategy.enable_parallel_graph_) {
    VLOG(3) << "use ParallelSSAGraphExecutor";
Y
Yancey1989 已提交
367
#ifdef PADDLE_WITH_CUDA
Y
Yancey1989 已提交
368 369
    // TODO(Yancey1989): Remove passing in the main_program when
    // allreduce_seq_pass doesn't need it as the attr.
Y
Yancey1989 已提交
370
    member_->executor_.reset(new details::ParallelSSAGraphExecutor(
X
Xin Pan 已提交
371
        exec_strategy, member_->local_scopes_, member_->places_, graph));
Y
Yancey1989 已提交
372 373 374 375
#else
    PADDLE_THROW(
        "Paddle should be compiled with CUDA for ParallelGraph Execution.");
#endif
Y
yuyang18 已提交
376
  } else {
Y
Yancey1989 已提交
377
    if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
Q
can run  
Qiao Longfei 已提交
378
      VLOG(3) << "use ThreadedSSAGraphExecutor";
Y
Yancey1989 已提交
379
      member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
X
Xin Pan 已提交
380
          exec_strategy, member_->local_scopes_, member_->places_, graph));
Y
Yancey1989 已提交
381
    } else {
Q
can run  
Qiao Longfei 已提交
382
      VLOG(3) << "use FastThreadedSSAGraphExecutor";
Y
Yancey1989 已提交
383
      member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
X
Xin Pan 已提交
384
          exec_strategy, member_->local_scopes_, member_->places_, graph));
Y
Yancey1989 已提交
385
    }
C
chengduoZH 已提交
386
  }
Y
yuyang18 已提交
387

Q
can run  
Qiao Longfei 已提交
388
  VLOG(3) << "use ScopeBufferedSSAGraphExecutor";
Q
Qiao Longfei 已提交
389 390 391 392 393
  if (!build_strategy.async_mode_) {
    member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
        exec_strategy, member_->local_scopes_, std::move(var_infos),
        member_->places_, std::move(member_->executor_)));
  }
Y
Yu Yang 已提交
394 395
}

Y
Yancey1989 已提交
396
void ParallelExecutor::BCastParamsToDevices(
397
    const std::unordered_set<std::string> &vars) const {
Q
Qiao Longfei 已提交
398
  VLOG(3) << "BCastParamsToDevices";
X
Xin Pan 已提交
399
  // the initializing bcast, all vars would be bcast from device(0).
400
  for (auto &var : vars) {
X
Xin Pan 已提交
401
    framework::Variable *main_var = member_->local_scopes_[0]->FindVar(var);
J
JiayiFeng 已提交
402
    if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
403 404 405 406
      continue;
    }

    auto &main_tensor = main_var->Get<LoDTensor>();
407
    if (!main_tensor.IsInitialized()) {
M
minqiyang 已提交
408
      VLOG(3) << "one in var not inited, return!";
409 410
      continue;
    }
411 412
    auto &dims = main_tensor.dims();
    if (paddle::platform::is_gpu_place(main_tensor.place())) {
P
peizhilin 已提交
413
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
414
      std::vector<void *> buffers;
C
chengduo 已提交
415
      buffers.reserve(member_->places_.size());
416 417 418 419 420
      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;
421

X
Xin Pan 已提交
422
        if (i == 0) {
423 424
          buffer = const_cast<void *>(main_tensor.data<void>());
        } else {
Y
Yu Yang 已提交
425
          auto local_scope = member_->local_scopes_[i];
426
          auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
Y
Update  
Yu Yang 已提交
427
          t->Resize(dims);
428
          buffer = t->mutable_data(place, main_tensor.type());
Y
Update  
Yu Yang 已提交
429
        }
430
        buffers.push_back(buffer);
431
      }
432

433 434 435 436 437 438
      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
439 440
          platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
                                       nccl_ctx.comm_, nccl_ctx.stream());
441
        }
442
        member_->nccl_ctxs_->WaitAll();
443
      }
C
chengduoZH 已提交
444 445 446
#else
      PADDLE_THROW("Not compiled with CUDA");
#endif
447 448
    } else {
      platform::CPUPlace cpu;
C
chengduo 已提交
449
      for (size_t i = 1; i < member_->places_.size(); ++i) {
450 451
        auto local_scope = member_->local_scopes_[i];
        auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
C
chengduo 已提交
452

Q
Qiao Longfei 已提交
453
        auto copy_memory = [&] {
454 455 456
          t->Resize(dims);
          t->mutable_data(cpu, main_tensor.type());
          paddle::framework::TensorCopy(main_tensor, cpu, t);
Q
can run  
Qiao Longfei 已提交
457 458
        };

Q
Qiao Longfei 已提交
459
        auto share_memory = [&] { t->ShareDataWith(main_tensor); };
Q
can run  
Qiao Longfei 已提交
460 461 462 463 464 465 466

        // FIXME(zcd): LR_DECAY_COUNTER should not be shared. This is a hot fix.
        if (member_->build_strategy_.async_mode_) {
          share_memory();
        } else if (member_->use_all_reduce_ || member_->use_cuda_ ||
                   var == "@LR_DECAY_COUNTER@") {
          copy_memory();
467
        } else {
Q
can run  
Qiao Longfei 已提交
468
          share_memory();
469
        }
Y
Yu Yang 已提交
470
      }
Y
Stash  
Yu Yang 已提交
471 472
    }
  }
Y
Yu Yang 已提交
473
}
Y
Yu Yang 已提交
474

Y
Yu Yang 已提交
475 476
void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                           const std::string &fetched_var_name) {
Y
Yu Yang 已提交
477 478 479
#ifdef WITH_GPERFTOOLS
  if (gProfileStarted) {
    ProfilerFlush();
S
sneaxiy 已提交
480 481
  }
#endif
Y
Yu Yang 已提交
482

X
Xin Pan 已提交
483
  platform::RecordBlock b(0);
S
sneaxiy 已提交
484 485
  if (member_->HasGarbageCollectors()) {
    member_->ResetRuntimeReferenceCount(fetch_tensors, fetched_var_name);
S
sneaxiy 已提交
486
  }
S
sneaxiy 已提交
487 488 489
  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
490
}
Y
Yu Yang 已提交
491

Y
Yu Yang 已提交
492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510
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_);
511 512 513 514 515
    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
516 517
    for (size_t j = 0; j < member_->places_.size(); ++j) {
      // TODO(panxy0718): Do I need to delete this var?
518
      auto t =
Y
Yu Yang 已提交
519
          member_->local_scopes_[j]->Var(pair.first)->GetMutable<LoDTensor>();
520 521
      t->ShareDataWith(lod_tensors[j]);
      t->set_lod(lod_tensors[j].lod());
X
Xin Pan 已提交
522 523 524 525
    }
  }
}

X
Xin Pan 已提交
526 527 528 529 530 531 532
ParallelExecutor::~ParallelExecutor() {
  for (auto &p : member_->places_) {
    platform::DeviceContextPool::Instance().Get(p)->Wait();
  }
  delete member_;
}

533
bool ParallelExecutor::EnableParallelGraphExecution(
X
Xin Pan 已提交
534
    const ir::Graph &graph, const ExecutionStrategy &exec_strategy,
535
    const BuildStrategy &build_strategy) const {
Y
Yancey1989 已提交
536
  if (!FLAGS_enable_parallel_graph) return false;
537

Y
Yancey1989 已提交
538
  bool enable_parallel_graph = true;
539

X
Xin Pan 已提交
540 541 542 543 544 545 546 547 548 549 550 551 552
  for (ir::Node *node : graph.Nodes()) {
    if (node->IsVar() && node->Var()) {
      // TODO(Yancey1989): support sparse update in ParallelGraph mode.
      if (node->Var()->GetType() == proto::VarType::SELECTED_ROWS) {
        enable_parallel_graph = false;
        break;
      }
    } else if (node->IsOp() && node->Op()) {
      // TODO(Yancey1989): support pserver mode
      if (node->Op()->Type() == "send" || node->Op()->Type() == "recv") {
        enable_parallel_graph = false;
        break;
      }
553 554 555 556 557
    }
  }

  if (!member_->use_all_reduce_ || !member_->use_cuda_)

Y
Yancey1989 已提交
558 559 560
    if (build_strategy.enable_sequential_execution_ ||
        exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental)
      enable_parallel_graph = false;
Y
Yancey1989 已提交
561
  return enable_parallel_graph;
562 563
}

Y
Yu Yang 已提交
564
}  // namespace framework
Y
Yang Yang 已提交
565
}  // namespace paddle
S
sneaxiy 已提交
566

S
sneaxiy 已提交
567
USE_PASS(reference_count_pass);
S
sneaxiy 已提交
568
USE_PASS(eager_deletion_pass);