parallel_executor.cc 26.9 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>
Q
qingqing01 已提交
17
#include <memory>
C
chengduoZH 已提交
18
#include <string>
19
#include <tuple>
Q
Qiao Longfei 已提交
20
#include <utility>
Q
qiaolongfei 已提交
21
#include <vector>
Q
Qiao Longfei 已提交
22
#include "paddle/fluid/framework/details/async_ssa_graph_executor.h"
Y
yuyang18 已提交
23
#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
24
#include "paddle/fluid/framework/details/multi_devices_helper.h"
Y
Yancey1989 已提交
25
#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
Y
yuyang18 已提交
26
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
Y
Yu Yang 已提交
27
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
28 29
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
30
#include "paddle/fluid/framework/ir/memory_optimize_pass/reference_count_pass_helper.h"
31
#include "paddle/fluid/platform/profiler.h"
Y
Yu Yang 已提交
32

Y
Yu Yang 已提交
33
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
34
#include "gperftools/profiler.h"
Y
Yu Yang 已提交
35
#endif
Y
Yu Yang 已提交
36
DEFINE_string(pe_profile_fname, "",
Y
Yu Yang 已提交
37 38
              "Profiler filename for PE, which generated by gperftools."
              "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable.");
39
DEFINE_bool(enable_parallel_graph, false,
Y
Yancey1989 已提交
40
            "Force disable parallel graph execution mode if set false.");
Y
Yu Yang 已提交
41

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
49

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

67 68 69 70 71 72 73 74 75 76 77
  ~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
78

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

  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
    }
  }

97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
  void InitNCCLCtxs(framework::Scope *scope, const BuildStrategy &bst) {
    VLOG(1) << "nccl comm num:" << bst.nccl_comm_num_ << ", nranks:" << nranks_
            << ", num_trainers:" << bst.num_trainers_
            << ", trainer_id:" << bst.trainer_id_;

    if (bst.use_hierarchical_allreduce_) {
      VLOG(1) << ", use_hierarchical_allreduce:"
              << bst.use_hierarchical_allreduce_ << ", inter_trainers_num:"
              << bst.hierarchical_allreduce_inter_nranks_
              << ", exter_trainers_num:"
              << bst.hierarchical_allreduce_exter_nranks_;
    }

    std::vector<ncclUniqueId *> flat_nccl_ids;
    if (nranks_ == 1) {
      // FIXME(gongwb): need not to create ncclid when nranks==1
114 115
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134
      return;
    }

    if (bst.enable_parallel_graph_) {
      VLOG(1) << "use only one ncclid in pg model";

      ncclUniqueId *nccl_id = nullptr;

      std::string var_name = platform::GetFlatNCCLVarName(0);
      auto nccl_id_var = scope->FindVar(var_name);
      if (nccl_id_var) {
        nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
      } else {
        nccl_id = new ncclUniqueId();
        PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(nccl_id));
      }

      flat_nccl_ids.push_back(nccl_id);

135 136
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
137 138 139 140 141 142
      VLOG(1) << "init bst nccl context complete!";
      return;
    }

    // num_trainers ==1 && places > 1
    if (bst.num_trainers_ == 1) {
143 144
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
145 146 147 148 149 150 151 152 153 154 155
      return;
    }

    for (int i = 0; i < static_cast<int>(bst.nccl_comm_num_); i++) {
      std::string var_name = platform::GetFlatNCCLVarName(i);
      auto nccl_id_var = scope->FindVar(var_name);
      PADDLE_ENFORCE(nccl_id_var, "can't find %s nccl_id_var", var_name);
      auto nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
      flat_nccl_ids.push_back(nccl_id);
    }

156 157
    nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                             bst.trainer_id_);
158 159

    if (bst.use_hierarchical_allreduce_) {
G
gongweibao 已提交
160 161 162 163 164 165 166 167
      std::vector<ncclUniqueId *> inter_nccl_ids;
      for (int i = 0; i < static_cast<int>(bst.nccl_comm_num_); i++) {
        std::string var_name = platform::GetHierarchicalInterNCCLVarName(i);
        auto nccl_id_var = scope->FindVar(var_name);
        PADDLE_ENFORCE(nccl_id_var, "can't find %s nccl_id_var", var_name);
        auto inter_nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
        inter_nccl_ids.push_back(inter_nccl_id);
      }
168 169 170 171 172 173 174 175 176

      std::vector<ncclUniqueId *> exter_nccl_ids;
      for (int i = 0; i < static_cast<int>(bst.nccl_comm_num_); i++) {
        std::string var_name = platform::GetHierarchicalExterNCCLVarName(i);
        auto nccl_id_var = scope->FindVar(var_name);
        PADDLE_ENFORCE(nccl_id_var, "can't find %s nccl_id_var", var_name);
        auto nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
        exter_nccl_ids.push_back(nccl_id);
      }
G
gongweibao 已提交
177

178 179 180 181
      nccl_ctxs_->InitHierarchicalCtxs(
          places_, inter_nccl_ids, exter_nccl_ids, bst.num_trainers_,
          bst.trainer_id_, bst.hierarchical_allreduce_inter_nranks_,
          bst.hierarchical_allreduce_exter_nranks_);
182 183
    }
  }
184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201

  void InitOrGetNCCLCommunicator(framework::Scope *scope,
                                 const BuildStrategy &bst) {
    const std::string var_name = "NCCLCommunicator";
    auto var = scope->FindVar(var_name);
    if (var != nullptr) {
      PADDLE_ENFORCE(var->IsInitialized(),
                     "if %s exists, it must be initialized", var_name);
      VLOG(1) << "find " << var_name
              << " in scope, so use it and does not recreate!";
      nccl_ctxs_ = var->GetMutable<platform::NCCLCommunicator>();
      return;
    }

    VLOG(1) << "not find " << var_name << " in scope, so recreate it!";
    nccl_ctxs_ = scope->Var(var_name)->GetMutable<platform::NCCLCommunicator>();
    InitNCCLCtxs(scope, bst);
  }
202 203
#endif

D
dzhwinter 已提交
204
  BuildStrategy build_strategy_;
Y
Yu Yang 已提交
205 206
  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
207
  Scope *global_scope_;  // not owned
Y
Yu Yang 已提交
208
  std::unique_ptr<details::SSAGraphExecutor> executor_;
Y
Yu Yang 已提交
209

P
peizhilin 已提交
210
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
211
  platform::NCCLCommunicator *nccl_ctxs_{nullptr};
Y
Yu Yang 已提交
212
#endif
C
chengduoZH 已提交
213 214
  bool own_local_scope_;
  bool use_cuda_;
215
  bool use_all_reduce_;
216
  size_t nranks_;
S
sneaxiy 已提交
217

S
sneaxiy 已提交
218 219 220
  // 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_
221 222 223
  std::vector<ir::ReferenceCountMap> global_ref_cnts_;
  std::vector<ir::AtomicReferenceCountMap> runtime_ref_cnts_;
  ir::GarbageCollectorMap gcs_;
Y
Yu Yang 已提交
224 225
};

226 227
ir::Graph *ParallelExecutorPrivate::PrepareGCAndRefCnts(
    ir::Graph *graph, size_t max_memory_size) {
S
sneaxiy 已提交
228 229 230 231 232
  for (size_t i = 0; i < places_.size(); ++i) {
    auto &place = places_[i];
    if (gcs_.count(place) > 0) {
      continue;
    }
S
sneaxiy 已提交
233
    std::unique_ptr<GarbageCollector> gc;
S
sneaxiy 已提交
234
#ifdef PADDLE_WITH_CUDA
S
sneaxiy 已提交
235 236
    if (platform::is_gpu_place(place)) {
      if (IsFastEagerDeletionModeEnabled()) {
S
sneaxiy 已提交
237 238
        gc.reset(new UnsafeFastGPUGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
239
      } else {
S
sneaxiy 已提交
240 241
        gc.reset(new StreamGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
242 243
      }
      VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
S
sneaxiy 已提交
244
    } else {
S
sneaxiy 已提交
245
#endif
S
sneaxiy 已提交
246 247 248 249 250 251 252
      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
253 254 255 256
#ifdef PADDLE_WITH_CUDA
    }
#endif

S
sneaxiy 已提交
257
    gcs_.emplace(place, std::move(gc));
S
sneaxiy 已提交
258 259
  }

S
sneaxiy 已提交
260
  if (!gcs_.empty()) {
261
    std::vector<ir::LastLiveOpsOfVars> last_live_ops_of_vars;
S
sneaxiy 已提交
262 263 264

    auto ref_cnt_pass =
        ir::PassRegistry::Instance().Get("reference_count_pass");
265 266
    ref_cnt_pass->SetNotOwned(ir::kGlobalReferenceCount, &global_ref_cnts_);
    ref_cnt_pass->SetNotOwned(ir::kLastLiveOpsOfVars, &last_live_ops_of_vars);
267
    graph = ref_cnt_pass->Apply(graph);
S
sneaxiy 已提交
268 269 270 271
    VLOG(10) << "ReferenceCountPass Applied";

    auto eager_deletion_pass =
        ir::PassRegistry::Instance().Get("eager_deletion_pass");
272
    eager_deletion_pass->SetNotOwned(ir::kRuntimeReferenceCount,
S
sneaxiy 已提交
273
                                     &runtime_ref_cnts_);
274 275
    eager_deletion_pass->SetNotOwned(ir::kGarbageCollector, &gcs_);
    eager_deletion_pass->SetNotOwned(ir::kLastLiveOpsOfVars,
S
sneaxiy 已提交
276
                                     &last_live_ops_of_vars);
277
    eager_deletion_pass->SetNotOwned(ir::kAllPlaces, &places_);
278
    graph = eager_deletion_pass->Apply(graph);
S
sneaxiy 已提交
279 280 281 282 283
    VLOG(10) << "EagerDeletionPass Applied";
  }
  return graph;
}

284 285 286 287
std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
  return member_->local_scopes_;
}

288 289 290 291 292 293 294 295 296 297 298 299 300 301
void ParallelExecutor::DropLocalExeScopes() {
  auto executor = dynamic_cast<details::ScopeBufferedSSAGraphExecutor *>(
      member_->executor_.get());
  if (executor) {
    executor->DropLocalExeScopes();
  }
}

bool ParallelExecutor::NeedCreateLocalExeScope() {
  auto executor = dynamic_cast<details::ScopeBufferedSSAGraphExecutor *>(
      member_->executor_.get());
  return executor && executor->NeedCreateLocalExeScope();
}

Y
Yan Xu 已提交
302 303 304 305 306 307 308 309
ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
                                   const std::vector<std::string> &bcast_vars,
                                   const std::string &loss_var_name,
                                   Scope *scope,
                                   const std::vector<Scope *> &local_scopes,
                                   const ExecutionStrategy &exec_strategy,
                                   const BuildStrategy &build_strategy,
                                   ir::Graph *graph)
Y
Yu Yang 已提交
310
    : member_(new ParallelExecutorPrivate(places)) {
Y
Yu Yang 已提交
311
  member_->global_scope_ = scope;
312
  member_->use_cuda_ = exec_strategy.use_cuda_;
D
dzhwinter 已提交
313
  member_->build_strategy_ = build_strategy;
314 315
  member_->use_all_reduce_ =
      build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce;
X
Xin Pan 已提交
316
  member_->nranks_ = build_strategy.num_trainers_ * places.size();
317 318 319 320 321
#if defined(PADDLE_WITH_CUDA) && defined(_WIN32)
  if (member_->use_cuda_) {
    PADDLE_ENFORCE(places.size() == 1, "Windows can support Single GPU only.");
  }
#endif
322 323 324 325
  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
326 327
  }

328 329 330 331 332 333
  LOG(WARNING) << string::Sprintf(
      "The number of %s, which is used in ParallelExecutor, is %lu. And "
      "the Program will be copied %lu copies",
      (member_->use_cuda_ ? "CUDAPlace" : "CPUPlace"), places.size(),
      places.size());

334
  // Step 1. Bcast the bcast_vars to devs.
Y
Yu Yang 已提交
335
  // Create local scopes
336
  if (local_scopes.empty()) {
C
chengduoZH 已提交
337
    member_->own_local_scope_ = true;
Y
Yu Yang 已提交
338 339
    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
340
      member_->local_scopes_.emplace_back(&scope->NewScope());
341 342
    }
  } else {
C
chengduoZH 已提交
343
    member_->own_local_scope_ = false;
344 345
    PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
    for (size_t i = 0; i < member_->places_.size(); ++i) {
346
      member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
347
    }
Y
Yu Yang 已提交
348 349
  }

Q
Qiao Longfei 已提交
350
  std::vector<ir::Graph *> graphs;
Q
Qiao Longfei 已提交
351 352 353
  if (build_strategy.async_mode_) {
    PADDLE_ENFORCE(!member_->use_cuda_,
                   "gpu mode does not support async_mode_ now!");
Q
Qiao Longfei 已提交
354
    graphs.push_back(graph);
D
dongdaxiang 已提交
355
    for (size_t i = 1; i < places.size(); ++i) {
Q
Qiao Longfei 已提交
356 357 358 359
      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
360
  }
Q
Qiao Longfei 已提交
361

Y
Yancey1989 已提交
362 363 364
  // FIXME(Yancey1989): parallel graph mode get better performance
  // in GPU allreduce distributed training. Need an elegant way to
  // choice the execution strategy.
365 366
  build_strategy.enable_parallel_graph_ =
      EnableParallelGraphExecution(*graph, exec_strategy, build_strategy);
Y
Yancey1989 已提交
367 368 369 370
  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
371

372
  if (member_->use_cuda_ && member_->nranks_ > 1) {
P
peizhilin 已提交
373
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
374
    member_->InitOrGetNCCLCommunicator(scope, build_strategy);
Q
qingqing01 已提交
375

W
Wu Yi 已提交
376 377 378
    // Initialize device context's nccl comm, will be used by normal
    // Operators like sync_batch_norm, and collective ops.
    // NOTE: more than one ParallelExecutor with same place, the nccl comm will
Q
qingqing01 已提交
379
    // be rewrite and there will be some problem.
W
Wu Yi 已提交
380 381 382
    // NOTE: NCCL group-calls and non-group-calls can not use the same
    // NCCL communicator, so for ParallelGraph and Multi-Process mode, re-use
    // same communicators.
383 384
    auto *nccl_ctxs =
        member_->nccl_ctxs_->GetSyncBatchNormCtx(scope, member_->places_);
Q
qingqing01 已提交
385 386 387 388 389
    for (size_t dev_id = 0; dev_id < member_->places_.size(); ++dev_id) {
      platform::DeviceContextPool &pool =
          platform::DeviceContextPool::Instance();
      auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
          pool.Get(member_->places_[dev_id]));
390
      auto &nccl_ctx = nccl_ctxs->at(member_->places_[dev_id]);
391
      dev_ctx->set_nccl_comm(nccl_ctx.comm());
Q
qingqing01 已提交
392
    }
Y
Yu Yang 已提交
393
#endif
C
chengduoZH 已提交
394
  }
Y
Yan Xu 已提交
395 396 397 398 399 400 401 402 403 404 405 406
  // broadcast parameters from the 0th device to others:
  auto need_broadcast = [&]() -> bool {
    if (build_strategy.num_trainers_ > 1) {
      // 1. num_tariners would be grater than 1 for nccl distributed training.
      return true;
    } else if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
      // 2. Only one trainer process, but ParallelExecutor hold multiple
      // devices.
      return true;
    }
    return false;
  };
407
  // Bcast Parameters to all GPUs
Y
Yan Xu 已提交
408 409
  if (need_broadcast()) {
    BCastParamsToDevices(bcast_vars, build_strategy.trainer_id_);
Y
Yu Yang 已提交
410
  }
411

Q
Qiao Longfei 已提交
412
  // Startup Program has been run. All local scopes has correct parameters.
Y
yuyang18 已提交
413

Q
Qiao Longfei 已提交
414 415 416
  // 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
417
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Q
Qiao Longfei 已提交
418
  if (build_strategy.async_mode_) {
Q
Qiao Longfei 已提交
419
    VLOG(3) << "use local async mode";
420 421
    graph = build_strategy.Apply(graph, {member_->places_[0]}, loss_var_name,
                                 {member_->local_scopes_[0]}, 1,
422
                                 member_->use_cuda_, member_->nccl_ctxs_);
D
dongdaxiang 已提交
423
    for (size_t i = 1; i < member_->places_.size(); ++i) {
424 425 426
      graphs[i] =
          build_strategy.Apply(graphs[i], {member_->places_[i]}, loss_var_name,
                               {member_->local_scopes_[i]}, 1,
427
                               member_->use_cuda_, member_->nccl_ctxs_);
428
      async_graphs[i] = graphs[i];
Q
Qiao Longfei 已提交
429
    }
Q
Qiao Longfei 已提交
430
  } else {
431 432
    graph = build_strategy.Apply(graph, member_->places_, loss_var_name,
                                 member_->local_scopes_, member_->nranks_,
433
                                 member_->use_cuda_, member_->nccl_ctxs_);
Q
Qiao Longfei 已提交
434
  }
C
chengduoZH 已提交
435
#else
Q
Qiao Longfei 已提交
436
  if (build_strategy.async_mode_) {
Q
Qiao Longfei 已提交
437
    VLOG(3) << "use local async mode";
438 439 440
    graph = build_strategy.Apply(graph, {member_->places_[0]}, loss_var_name,
                                 {member_->local_scopes_[0]}, 1,
                                 member_->use_cuda_);
441
    for (size_t i = 1; i < member_->places_.size(); ++i) {
442 443
      graphs[i] = build_strategy.Apply(
          graphs[i], {member_->places_[i]}, loss_var_name,
Q
Qiao Longfei 已提交
444
          {member_->local_scopes_[i]}, 1, member_->use_cuda_);
445
      async_graphs[i] = graphs[i];
Q
Qiao Longfei 已提交
446
    }
Q
can run  
Qiao Longfei 已提交
447
  } else {
448 449 450
    graph = build_strategy.Apply(graph, member_->places_, loss_var_name,
                                 member_->local_scopes_, member_->nranks_,
                                 member_->use_cuda_);
Q
can run  
Qiao Longfei 已提交
451
  }
Y
Yu Yang 已提交
452
#endif
453

Y
Yancey1989 已提交
454
  auto max_memory_size = GetEagerDeletionThreshold();
D
dzhwinter 已提交
455 456
  VLOG(10) << "Eager Deletion Threshold "
           << static_cast<float>(max_memory_size) / (1 << 30);
Y
Yancey1989 已提交
457
  if (max_memory_size >= 0) {
458 459
    graph = member_->PrepareGCAndRefCnts(graph,
                                         static_cast<size_t>(max_memory_size));
Y
Yancey1989 已提交
460 461
  }

Q
Qiao Longfei 已提交
462 463
  async_graphs[0] = graph;

464 465
  // Step 3. Create vars in each scope. Passes may also create new vars.
  //         skip control vars and empty vars
Y
Yancey1989 已提交
466
  std::vector<details::VariableInfo> var_infos;
Q
Qiao Longfei 已提交
467 468 469 470 471 472
  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
473 474
    }
  }
Y
Yancey1989 已提交
475

W
Wu Yi 已提交
476 477
  // 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
478
    size_t graph_num = ir::GraphNum(*graph);
C
chengduo 已提交
479 480 481 482
    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
483
          << ir::GraphNum(*graph)
C
chengduo 已提交
484 485 486 487 488
          << " 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
489 490
  }

Q
Qiao Longfei 已提交
491
  if (build_strategy.async_mode_) {
Q
can run  
Qiao Longfei 已提交
492 493
    VLOG(3) << "use AsyncSSAGraphExecutor";
    member_->executor_.reset(new details::AsyncSSAGraphExecutor(
Q
Qiao Longfei 已提交
494
        exec_strategy, member_->local_scopes_, member_->places_, async_graphs));
Q
can run  
Qiao Longfei 已提交
495 496
  } else if (build_strategy.enable_parallel_graph_) {
    VLOG(3) << "use ParallelSSAGraphExecutor";
Y
Yancey1989 已提交
497
#ifdef PADDLE_WITH_CUDA
Y
Yancey1989 已提交
498 499
    // TODO(Yancey1989): Remove passing in the main_program when
    // allreduce_seq_pass doesn't need it as the attr.
Y
Yancey1989 已提交
500
    member_->executor_.reset(new details::ParallelSSAGraphExecutor(
X
Xin Pan 已提交
501
        exec_strategy, member_->local_scopes_, member_->places_, graph));
Y
Yancey1989 已提交
502 503 504 505
#else
    PADDLE_THROW(
        "Paddle should be compiled with CUDA for ParallelGraph Execution.");
#endif
Y
yuyang18 已提交
506
  } else {
Y
Yancey1989 已提交
507
    if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
Q
can run  
Qiao Longfei 已提交
508
      VLOG(3) << "use ThreadedSSAGraphExecutor";
Y
Yancey1989 已提交
509
      member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
X
Xin Pan 已提交
510
          exec_strategy, member_->local_scopes_, member_->places_, graph));
Y
Yancey1989 已提交
511
    } else {
Q
can run  
Qiao Longfei 已提交
512
      VLOG(3) << "use FastThreadedSSAGraphExecutor";
Y
Yancey1989 已提交
513
      member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
X
Xin Pan 已提交
514
          exec_strategy, member_->local_scopes_, member_->places_, graph));
Y
Yancey1989 已提交
515
    }
C
chengduoZH 已提交
516
  }
Y
yuyang18 已提交
517

Q
can run  
Qiao Longfei 已提交
518
  VLOG(3) << "use ScopeBufferedSSAGraphExecutor";
Q
Qiao Longfei 已提交
519 520 521 522 523
  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
524 525
}

Y
Yancey1989 已提交
526
void ParallelExecutor::BCastParamsToDevices(
Y
Yan Xu 已提交
527
    const std::vector<std::string> &vars, int trainer_id) const {
Q
Qiao Longfei 已提交
528
  VLOG(3) << "BCastParamsToDevices";
X
Xin Pan 已提交
529
  // the initializing bcast, all vars would be bcast from device(0).
530
  for (auto &var : vars) {
X
Xin Pan 已提交
531
    framework::Variable *main_var = member_->local_scopes_[0]->FindVar(var);
J
JiayiFeng 已提交
532
    if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
533 534 535 536
      continue;
    }

    auto &main_tensor = main_var->Get<LoDTensor>();
537
    if (!main_tensor.IsInitialized()) {
M
minqiyang 已提交
538
      VLOG(3) << "one in var not inited, return!";
539 540
      continue;
    }
541 542
    auto &dims = main_tensor.dims();
    if (paddle::platform::is_gpu_place(main_tensor.place())) {
P
peizhilin 已提交
543
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
544
      std::vector<void *> buffers;
C
chengduo 已提交
545
      buffers.reserve(member_->places_.size());
546 547 548 549 550
      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;
551

Y
Yan Xu 已提交
552
        if (i == 0 && trainer_id == 0) {
553 554
          buffer = const_cast<void *>(main_tensor.data<void>());
        } else {
Y
Yu Yang 已提交
555
          auto local_scope = member_->local_scopes_[i];
556
          auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
Y
Update  
Yu Yang 已提交
557
          t->Resize(dims);
558
          buffer = t->mutable_data(place, main_tensor.type());
Y
Update  
Yu Yang 已提交
559
        }
560
        buffers.push_back(buffer);
561
      }
562

563 564 565
      PADDLE_ENFORCE_EQ(member_->places_.size(), buffers.size(),
                        "variables' buffer size to bcast NOT equal to places");
      {
566
        auto *nccl_ctxs = member_->nccl_ctxs_->DefaultFlatCtx();
567 568
        platform::NCCLGroupGuard guard;
        for (size_t i = 0; i < member_->places_.size(); ++i) {
569
          auto &nccl_ctx = nccl_ctxs->at(member_->places_[i]);
X
Xin Pan 已提交
570 571
          platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
                                       nccl_ctx.comm_, nccl_ctx.stream());
572
        }
573
        nccl_ctxs->WaitAll();
574
      }
C
chengduoZH 已提交
575
#endif
576 577
    } else {
      platform::CPUPlace cpu;
C
chengduo 已提交
578
      for (size_t i = 1; i < member_->places_.size(); ++i) {
579 580
        auto local_scope = member_->local_scopes_[i];
        auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
C
chengduo 已提交
581

Q
Qiao Longfei 已提交
582
        auto copy_memory = [&] {
583 584 585
          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
586 587
        };

Q
Qiao Longfei 已提交
588
        auto share_memory = [&] { t->ShareDataWith(main_tensor); };
Q
can run  
Qiao Longfei 已提交
589 590 591 592 593 594 595

        // 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();
596
        } else {
Q
can run  
Qiao Longfei 已提交
597
          share_memory();
598
        }
Y
Yu Yang 已提交
599
      }
Y
Stash  
Yu Yang 已提交
600 601
    }
  }
Y
Yu Yang 已提交
602
}
Y
Yu Yang 已提交
603

Y
Yu Yang 已提交
604 605
void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                           const std::string &fetched_var_name) {
606
  VLOG(3) << "enter ParallelExecutor Run";
Y
Yu Yang 已提交
607 608 609
#ifdef WITH_GPERFTOOLS
  if (gProfileStarted) {
    ProfilerFlush();
S
sneaxiy 已提交
610 611
  }
#endif
Y
Yu Yang 已提交
612

X
Xin Pan 已提交
613
  platform::RecordBlock b(0);
S
sneaxiy 已提交
614
  if (member_->HasGarbageCollectors()) {
615
    platform::RecordEvent event("PrepareGarbageCollectors");
S
sneaxiy 已提交
616
    member_->ResetRuntimeReferenceCount(fetch_tensors, fetched_var_name);
S
sneaxiy 已提交
617
  }
618 619

  VLOG(3) << "ParallelExecutor begin to run member_->executor_->Run";
S
sneaxiy 已提交
620 621 622
  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
623
}
Y
Yu Yang 已提交
624

Y
Yu Yang 已提交
625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643
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_);
C
chengduo 已提交
644 645 646 647 648 649 650 651 652 653 654 655 656 657 658
    if (member_->places_.size() != lod_tensors.size()) {
      bool is_cpu_place = platform::is_cpu_place(member_->places_.front());
      auto error_info = string::Sprintf(
          "The number(%d) of samples of "
          "current batch is less than the count(%d) of "
          "devices(%s), currently, it is not allowed. ",
          member_->places_.size(), lod_tensors.size(),
          (is_cpu_place ? "CPU" : "GPU"));
      if (is_cpu_place) {
        error_info +=
            "You should set the environment variable CPU_NUM in the system "
            "to determine the number of devices you need.";
      }
      PADDLE_THROW(error_info);
    }
X
Xin Pan 已提交
659 660
    for (size_t j = 0; j < member_->places_.size(); ++j) {
      // TODO(panxy0718): Do I need to delete this var?
661
      auto t =
Y
Yu Yang 已提交
662
          member_->local_scopes_[j]->Var(pair.first)->GetMutable<LoDTensor>();
663 664
      t->ShareDataWith(lod_tensors[j]);
      t->set_lod(lod_tensors[j].lod());
X
Xin Pan 已提交
665 666 667 668
    }
  }
}

X
Xin Pan 已提交
669 670 671 672 673 674 675
ParallelExecutor::~ParallelExecutor() {
  for (auto &p : member_->places_) {
    platform::DeviceContextPool::Instance().Get(p)->Wait();
  }
  delete member_;
}

676
bool ParallelExecutor::EnableParallelGraphExecution(
X
Xin Pan 已提交
677
    const ir::Graph &graph, const ExecutionStrategy &exec_strategy,
678
    const BuildStrategy &build_strategy) const {
679 680 681
  if (!FLAGS_enable_parallel_graph) {
    return false;
  }
682

Y
Yancey1989 已提交
683
  bool enable_parallel_graph = true;
684

X
Xin Pan 已提交
685 686 687 688 689 690 691 692 693 694 695 696 697
  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;
      }
698 699 700
    }
  }

701
  if (!member_->use_all_reduce_ || !member_->use_cuda_) {
Y
Yancey1989 已提交
702
    if (build_strategy.enable_sequential_execution_ ||
703
        exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental) {
Y
Yancey1989 已提交
704
      enable_parallel_graph = false;
705 706 707 708 709 710 711 712 713
    }
  }

#ifdef WIN32
  VLOG(1) << "Windows has no support to parallel graph, enable_parallel_graph "
             "would be forced to false.";
  enable_parallel_graph = false;
#endif

Y
Yancey1989 已提交
714
  return enable_parallel_graph;
715 716
}

Y
Yu Yang 已提交
717
}  // namespace framework
Y
Yang Yang 已提交
718
}  // namespace paddle
S
sneaxiy 已提交
719

S
sneaxiy 已提交
720
USE_PASS(reference_count_pass);
S
sneaxiy 已提交
721
USE_PASS(eager_deletion_pass);