parallel_executor.cc 33.8 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"
25
#include "paddle/fluid/framework/details/op_handle_base.h"
Y
Yancey1989 已提交
26
#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
Y
yuyang18 已提交
27
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
Y
Yu Yang 已提交
28
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
29 30
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
31
#include "paddle/fluid/framework/ir/memory_optimize_pass/memory_optimization_var_info.h"
32
#include "paddle/fluid/framework/ir/memory_optimize_pass/reference_count_pass_helper.h"
33
#include "paddle/fluid/platform/profiler.h"
Y
Yu Yang 已提交
34

35 36
DECLARE_bool(use_ngraph);

37 38
DECLARE_double(eager_delete_tensor_gb);

Y
Yu Yang 已提交
39
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
40
#include "gperftools/profiler.h"
Y
Yu Yang 已提交
41
#endif
Y
Yu Yang 已提交
42
DEFINE_string(pe_profile_fname, "",
Y
Yu Yang 已提交
43 44
              "Profiler filename for PE, which generated by gperftools."
              "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable.");
45
DEFINE_bool(enable_parallel_graph, false,
Y
Yancey1989 已提交
46
            "Force disable parallel graph execution mode if set false.");
Y
Yu Yang 已提交
47

Y
Yang Yang 已提交
48
namespace paddle {
Y
Yu Yang 已提交
49 50
namespace framework {

Y
Yu Yang 已提交
51
static std::once_flag gProfileOnce;
Y
Yu Yang 已提交
52
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
53
static bool gProfileStarted = false;
Y
Yu Yang 已提交
54
#endif
55

Y
Yu Yang 已提交
56 57 58
class ParallelExecutorPrivate {
 public:
  explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
Y
Yu Yang 已提交
59
      : places_(places) {
Y
Yu Yang 已提交
60
    if (!FLAGS_pe_profile_fname.empty()) {
Y
Yu Yang 已提交
61 62
      std::call_once(gProfileOnce, [] {
#ifdef WITH_GPERFTOOLS
Y
Yu Yang 已提交
63
        ProfilerStart(FLAGS_pe_profile_fname.c_str());
Y
Yu Yang 已提交
64 65 66
        gProfileStarted = true;
#else
        LOG(WARNING) << "Paddle is not compiled with gperftools. "
67
          "FLAGS_pe_profile_fname will be ignored";
Y
Yu Yang 已提交
68 69 70 71
#endif
      });
    }
  }
Y
Yu Yang 已提交
72

73 74 75 76 77 78 79 80 81 82 83
  ~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
84

85
  ir::Graph *ApplyMemoryOptimizePass(ir::Graph *graph);
S
sneaxiy 已提交
86 87 88

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

89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
  /**
   * NOTE(zengjinle): the feeded variables of users should not be reused,
   * because users may feed them into another network. Changing the feeded
   * variables that users can visit may cause calculation wrong, which is
   * a very subtle bug when traning networks. However, these variables
   * can be garbage collected.
   *
   * ParallelExecutor provides 2 methods to feed variables:
   *
   *  - FeedTensorsIntoLocalScopes: this method would share memory of feeded
   *                                variables, so we have to skip these.
   *
   *  - FeedAndSplitTensorIntoLocalScopes: this method would copy data of feeded
   *                                       variables, so we do not need to skip
   *                                       them.
   */
  inline void SetSkipMemoryReuse(size_t scope_idx, const std::string &name) {
    auto iter = mem_opt_var_infos_[scope_idx].find(name);
    if (iter != mem_opt_var_infos_[scope_idx].end()) {
      iter->second->SetSkipMemoryReuse(true);
    }
  }

112
#if defined(PADDLE_WITH_NCCL)
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128
  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
129 130
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
131 132 133 134 135 136 137 138 139 140 141 142
      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>();
143
        VLOG(10) << "find nccl_id_var:" << var_name << ", nccl_id:" << nccl_id;
144 145 146
      } else {
        nccl_id = new ncclUniqueId();
        PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(nccl_id));
147 148
        VLOG(10) << "can't find nccl_id_var:" << var_name
                 << ", nccl_id:" << nccl_id;
149 150 151 152
      }

      flat_nccl_ids.push_back(nccl_id);

153 154
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
155 156 157 158 159 160
      VLOG(1) << "init bst nccl context complete!";
      return;
    }

    // num_trainers ==1 && places > 1
    if (bst.num_trainers_ == 1) {
161 162
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
163 164 165 166 167 168 169 170 171 172 173
      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);
    }

174 175
    nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                             bst.trainer_id_);
176 177

    if (bst.use_hierarchical_allreduce_) {
G
gongweibao 已提交
178 179 180 181 182 183 184 185
      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);
      }
186 187 188 189 190 191 192 193 194

      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
195

196 197 198 199
      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_);
200 201
    }
  }
202

203
  void InitOrGetNCCLCommunicator(framework::Scope *scope, BuildStrategy *bst) {
204 205 206 207 208 209 210 211 212 213 214
    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;
    }

215 216 217 218 219 220 221 222 223 224 225 226 227 228 229
    if (bst->use_hierarchical_allreduce_) {
      PADDLE_ENFORCE(bst->num_trainers_ > 1, "num_trainers:%llu < 1",
                     bst->num_trainers_);
      PADDLE_ENFORCE(bst->hierarchical_allreduce_inter_nranks_ > 1,
                     "inter_nranks:%d < 1",
                     bst->hierarchical_allreduce_inter_nranks_);
      PADDLE_ENFORCE(
          (bst->num_trainers_ % bst->hierarchical_allreduce_inter_nranks_ == 0),
          "num_trainers:%llu mod inter_nranks:%d != 0", bst->num_trainers_,
          bst->hierarchical_allreduce_inter_nranks_);

      bst->hierarchical_allreduce_exter_nranks_ =
          bst->num_trainers_ / bst->hierarchical_allreduce_inter_nranks_;
    }

230 231
    VLOG(1) << "not find " << var_name << " in scope, so recreate it!";
    nccl_ctxs_ = scope->Var(var_name)->GetMutable<platform::NCCLCommunicator>();
232
    InitNCCLCtxs(scope, *bst);
233
  }
234 235
#endif

236 237 238 239 240
  inline bool IsPersistable(const std::string &name) const {
    auto iter = is_persistable_.find(name);
    return iter != is_persistable_.end() && iter->second;
  }

D
dzhwinter 已提交
241
  BuildStrategy build_strategy_;
Y
Yu Yang 已提交
242 243
  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
244
  std::vector<Scope *> local_exec_scopes_;
245
  Scope *global_scope_;  // not owned
Y
Yu Yang 已提交
246
  std::unique_ptr<details::SSAGraphExecutor> executor_;
Y
Yu Yang 已提交
247

248 249
  std::unordered_map<std::string, bool> is_persistable_;

250
#if defined(PADDLE_WITH_NCCL)
251
  platform::NCCLCommunicator *nccl_ctxs_{nullptr};
Y
Yu Yang 已提交
252
#endif
C
chengduoZH 已提交
253 254
  bool own_local_scope_;
  bool use_cuda_;
255
  bool use_all_reduce_;
256
  size_t nranks_;
S
sneaxiy 已提交
257

258
  ir::MemOptVarInfoMapList mem_opt_var_infos_;
259
  ir::GarbageCollectorMap gcs_;
Y
Yu Yang 已提交
260 261
};

262
ir::Graph *ParallelExecutorPrivate::ApplyMemoryOptimizePass(ir::Graph *graph) {
263 264 265 266 267 268 269
  if (FLAGS_use_ngraph) {
    LOG_FIRST_N(WARNING, 1)
        << "FLAGS_use_ngraph=True, memory optimization strategy is "
           "disabled in ParallelExecutor";
    return graph;
  }

Z
Zeng Jinle 已提交
270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
  /**
   * NOTE(zengjinle): If BuildStrategy.memory_optimize = None in Python,
   * set BuildStrategy.memory_optimize according to whether gc is enabled.
   * If gc is enabled, BuildStrategy.memory_optimize = False.
   * If gc is disabled, BuildStrategy.memory_optimize = True.
   * This is because gc+memory_optimize is worse than gc only.
   *
   * As an option, users can enable BuildStrategy.memory_optimize forcely
   * by setting True, and disable it forcely by setting False.
   */
  bool is_gc_enabled = (GetEagerDeletionThreshold() >= 0);
  if (!build_strategy_.memory_optimize_) {
    build_strategy_.memory_optimize_ = !is_gc_enabled;
  }

  bool need_mem_opt = build_strategy_.enable_inplace_ ||
                      build_strategy_.memory_optimize_.get() || is_gc_enabled;

  if (!need_mem_opt) return graph;

290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306
  std::vector<ir::LastLiveOpsOfVars> last_live_ops_of_vars;

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

  if (build_strategy_.enable_inplace_) {
    auto inplace_pass =
        ir::PassRegistry::Instance().Get("buffer_shared_inplace_pass");
    inplace_pass->SetNotOwned(ir::kMemOptVarInfoMapList, &mem_opt_var_infos_);
    inplace_pass->SetNotOwned(ir::kLastLiveOpsOfVars, &last_live_ops_of_vars);
    inplace_pass->SetNotOwned(ir::kUseCuda, &use_cuda_);
    VLOG(10) << "Start to apply buffer_shared_inplace_pass";
    graph = inplace_pass->Apply(graph);
    VLOG(10) << "buffer_shared_inplace_pass Applied";
307 308
    LOG_FIRST_N(INFO, 1) << "Inplace strategy is enabled, when "
                            "build_strategy.enable_inplace = True";
309 310
  }

311
  if (build_strategy_.memory_optimize_.get()) {
312 313 314 315 316 317 318 319 320 321
    auto cross_op_memory_reuse_pass = ir::PassRegistry::Instance().Get(
        "buffer_shared_cross_op_memory_reuse_pass");
    cross_op_memory_reuse_pass->SetNotOwned(ir::kMemOptVarInfoMapList,
                                            &mem_opt_var_infos_);
    cross_op_memory_reuse_pass->SetNotOwned(ir::kLastLiveOpsOfVars,
                                            &last_live_ops_of_vars);
    cross_op_memory_reuse_pass->SetNotOwned(ir::kUseCuda, &use_cuda_);
    VLOG(10) << "Start to apply buffer_shared_cross_op_memory_reuse_pass";
    graph = cross_op_memory_reuse_pass->Apply(graph);
    VLOG(10) << "buffer_shared_cross_op_memory_reuse_pass Applied";
Z
Zeng Jinle 已提交
322 323 324
    LOG(INFO) << "Cross op memory reuse strategy is enabled, when "
                 "build_strategy.memory_optimize = True or garbage collection "
                 "strategy is disabled, which is not recommended";
325
  }
326

327
  if (!is_gc_enabled) {
328 329 330 331
    return graph;
  }
  size_t max_memory_size = static_cast<size_t>(GetEagerDeletionThreshold());

S
sneaxiy 已提交
332 333 334 335 336
  for (size_t i = 0; i < places_.size(); ++i) {
    auto &place = places_[i];
    if (gcs_.count(place) > 0) {
      continue;
    }
S
sneaxiy 已提交
337
    std::unique_ptr<GarbageCollector> gc;
S
sneaxiy 已提交
338
#ifdef PADDLE_WITH_CUDA
S
sneaxiy 已提交
339 340
    if (platform::is_gpu_place(place)) {
      if (IsFastEagerDeletionModeEnabled()) {
S
sneaxiy 已提交
341 342
        gc.reset(new UnsafeFastGPUGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
343
      } else {
S
sneaxiy 已提交
344 345
        gc.reset(new StreamGarbageCollector(
            boost::get<platform::CUDAPlace>(place), max_memory_size));
S
sneaxiy 已提交
346 347
      }
      VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
S
sneaxiy 已提交
348
    } else {
S
sneaxiy 已提交
349
#endif
S
sneaxiy 已提交
350 351 352 353 354 355 356
      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
357 358 359 360
#ifdef PADDLE_WITH_CUDA
    }
#endif

S
sneaxiy 已提交
361
    gcs_.emplace(place, std::move(gc));
S
sneaxiy 已提交
362 363
  }

S
sneaxiy 已提交
364
  if (!gcs_.empty()) {
S
sneaxiy 已提交
365 366
    auto eager_deletion_pass =
        ir::PassRegistry::Instance().Get("eager_deletion_pass");
367 368
    eager_deletion_pass->SetNotOwned(ir::kMemOptVarInfoMapList,
                                     &mem_opt_var_infos_);
369 370
    eager_deletion_pass->SetNotOwned(ir::kGarbageCollector, &gcs_);
    eager_deletion_pass->SetNotOwned(ir::kLastLiveOpsOfVars,
S
sneaxiy 已提交
371
                                     &last_live_ops_of_vars);
372
    eager_deletion_pass->SetNotOwned(ir::kAllPlaces, &places_);
373
    graph = eager_deletion_pass->Apply(graph);
S
sneaxiy 已提交
374
    VLOG(10) << "EagerDeletionPass Applied";
375 376 377
    LOG_FIRST_N(INFO, 1) << "Garbage collection strategy is enabled, when "
                         << "FLAGS_eager_delete_tensor_gb = "
                         << FLAGS_eager_delete_tensor_gb;
S
sneaxiy 已提交
378 379 380 381
  }
  return graph;
}

382 383
size_t ParallelExecutor::DeviceCount() const { return member_->places_.size(); }

384 385 386 387
std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
  return member_->local_scopes_;
}

388 389 390 391 392 393 394 395 396 397 398 399 400 401
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
402 403 404 405 406 407 408 409
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
410
    : member_(new ParallelExecutorPrivate(places)) {
Y
Yu Yang 已提交
411
  member_->global_scope_ = scope;
412
  member_->use_cuda_ = exec_strategy.use_cuda_;
D
dzhwinter 已提交
413
  member_->build_strategy_ = build_strategy;
C
chengduo 已提交
414 415
  member_->use_all_reduce_ = member_->build_strategy_.reduce_ ==
                             BuildStrategy::ReduceStrategy::kAllReduce;
X
Xin Pan 已提交
416
  member_->nranks_ = build_strategy.num_trainers_ * places.size();
C
chengduo 已提交
417 418 419 420 421 422 423
  if (!member_->use_all_reduce_ && member_->nranks_ == 1) {
    LOG(INFO) << "If you set build_strategy.reduce with 'Reduce',"
                 "the number of places should be greater than 1.";
    member_->build_strategy_.reduce_ =
        BuildStrategy::ReduceStrategy::kAllReduce;
    member_->use_all_reduce_ = true;
  }
424 425 426 427 428
#if defined(PADDLE_WITH_CUDA) && defined(_WIN32)
  if (member_->use_cuda_) {
    PADDLE_ENFORCE(places.size() == 1, "Windows can support Single GPU only.");
  }
#endif
Y
Yancey1989 已提交
429

430 431 432 433 434 435 436 437 438 439
#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_NCCL)
  PADDLE_ENFORCE_EQ(
      places.size(), 1,
      platform::errors::PermissionDenied(
          "Your machine has multiple cards, "
          "but the WITH_NCCL option is not turned on during compilation, "
          "and you cannot use multi-card training or prediction. "
          "Please recompile and turn on the WITH_NCCL option."));
#endif

440
  LOG(INFO) << string::Sprintf(
441 442 443
      "The Program will be executed on %s using ParallelExecutor, %lu "
      "cards are used, so %lu programs are executed in parallel.",
      (member_->use_cuda_ ? "CUDA" : "CPU"), places.size(), places.size());
C
chengduo 已提交
444

445
  // Step 1. Bcast the bcast_vars to devs.
Y
Yu Yang 已提交
446
  // Create local scopes
447
  if (local_scopes.empty()) {
C
chengduoZH 已提交
448
    member_->own_local_scope_ = true;
Y
Yu Yang 已提交
449 450
    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
451
      member_->local_scopes_.emplace_back(&scope->NewScope());
452 453
    }
  } else {
C
chengduoZH 已提交
454
    member_->own_local_scope_ = false;
455 456
    PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
    for (size_t i = 0; i < member_->places_.size(); ++i) {
457
      member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
458
    }
Y
Yu Yang 已提交
459 460
  }

Q
Qiao Longfei 已提交
461
  std::vector<ir::Graph *> graphs;
C
chengduo 已提交
462
  if (member_->build_strategy_.async_mode_) {
Q
Qiao Longfei 已提交
463 464
    PADDLE_ENFORCE(!member_->use_cuda_,
                   "gpu mode does not support async_mode_ now!");
Q
Qiao Longfei 已提交
465
    graphs.push_back(graph);
D
dongdaxiang 已提交
466
    for (size_t i = 1; i < places.size(); ++i) {
Q
Qiao Longfei 已提交
467 468 469 470
      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
471
  }
Q
Qiao Longfei 已提交
472

Y
Yancey1989 已提交
473 474 475
  // FIXME(Yancey1989): parallel graph mode get better performance
  // in GPU allreduce distributed training. Need an elegant way to
  // choice the execution strategy.
C
chengduo 已提交
476 477 478 479
  member_->build_strategy_.enable_parallel_graph_ =
      EnableParallelGraphExecution(*graph, exec_strategy,
                                   member_->build_strategy_);
  if (member_->build_strategy_.enable_parallel_graph_) {
480 481 482 483
    LOG(INFO) << "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
484

485
  if (member_->use_cuda_ && member_->nranks_ > 1) {
486
#if defined(PADDLE_WITH_NCCL)
487
    member_->InitOrGetNCCLCommunicator(scope, &member_->build_strategy_);
Q
qingqing01 已提交
488

W
Wu Yi 已提交
489 490 491
    // 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
492
    // be rewrite and there will be some problem.
W
Wu Yi 已提交
493 494 495
    // 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.
496 497
    auto *nccl_ctxs =
        member_->nccl_ctxs_->GetSyncBatchNormCtx(scope, member_->places_);
498
    auto &pool = platform::DeviceContextPool::Instance();
Q
qingqing01 已提交
499 500 501
    for (size_t dev_id = 0; dev_id < member_->places_.size(); ++dev_id) {
      auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
          pool.Get(member_->places_[dev_id]));
502
      auto &nccl_ctx = nccl_ctxs->at(member_->places_[dev_id]);
503
      dev_ctx->set_nccl_comm(nccl_ctx.comm());
Q
qingqing01 已提交
504
    }
Y
Yu Yang 已提交
505
#endif
C
chengduoZH 已提交
506
  }
Y
Yan Xu 已提交
507 508
  // broadcast parameters from the 0th device to others:
  auto need_broadcast = [&]() -> bool {
C
chengduo 已提交
509
    if (member_->build_strategy_.num_trainers_ > 1) {
Y
Yan Xu 已提交
510 511 512 513 514 515 516 517 518
      // 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;
  };
519
  // Bcast Parameters to all GPUs
Y
Yan Xu 已提交
520
  if (need_broadcast()) {
C
chengduo 已提交
521
    BCastParamsToDevices(bcast_vars, member_->build_strategy_.trainer_id_);
Y
Yu Yang 已提交
522
  }
523

Q
Qiao Longfei 已提交
524
  // Startup Program has been run. All local scopes has correct parameters.
Y
yuyang18 已提交
525

Q
Qiao Longfei 已提交
526 527 528
  // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
  // ncclOp
  std::vector<ir::Graph *> async_graphs(places.size());
529
#if defined(PADDLE_WITH_NCCL)
C
chengduo 已提交
530
  if (member_->build_strategy_.async_mode_) {
Q
Qiao Longfei 已提交
531
    VLOG(3) << "use local async mode";
C
chengduo 已提交
532 533 534 535
    graph = member_->build_strategy_.Apply(
        graph, {member_->places_[0]}, loss_var_name,
        {member_->local_scopes_[0]}, 1, member_->use_cuda_,
        member_->nccl_ctxs_);
D
dongdaxiang 已提交
536
    for (size_t i = 1; i < member_->places_.size(); ++i) {
C
chengduo 已提交
537 538 539 540
      graphs[i] = member_->build_strategy_.Apply(
          graphs[i], {member_->places_[i]}, loss_var_name,
          {member_->local_scopes_[i]}, 1, member_->use_cuda_,
          member_->nccl_ctxs_);
541
      async_graphs[i] = graphs[i];
Q
Qiao Longfei 已提交
542
    }
Q
Qiao Longfei 已提交
543
  } else {
C
chengduo 已提交
544 545 546
    graph = member_->build_strategy_.Apply(
        graph, member_->places_, loss_var_name, member_->local_scopes_,
        member_->nranks_, member_->use_cuda_, member_->nccl_ctxs_);
Q
Qiao Longfei 已提交
547
  }
C
chengduoZH 已提交
548
#else
C
chengduo 已提交
549
  if (member_->build_strategy_.async_mode_) {
Q
Qiao Longfei 已提交
550
    VLOG(3) << "use local async mode";
C
chengduo 已提交
551 552 553
    graph = member_->build_strategy_.Apply(
        graph, {member_->places_[0]}, loss_var_name,
        {member_->local_scopes_[0]}, 1, member_->use_cuda_);
554
    for (size_t i = 1; i < member_->places_.size(); ++i) {
C
chengduo 已提交
555
      graphs[i] = member_->build_strategy_.Apply(
556
          graphs[i], {member_->places_[i]}, loss_var_name,
Q
Qiao Longfei 已提交
557
          {member_->local_scopes_[i]}, 1, member_->use_cuda_);
558
      async_graphs[i] = graphs[i];
Q
Qiao Longfei 已提交
559
    }
Q
can run  
Qiao Longfei 已提交
560
  } else {
C
chengduo 已提交
561 562 563
    graph = member_->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
564
  }
Y
Yu Yang 已提交
565
#endif
566

567
  graph = member_->ApplyMemoryOptimizePass(graph);
Y
Yancey1989 已提交
568

Q
Qiao Longfei 已提交
569 570
  async_graphs[0] = graph;

571 572
  // Step 3. Create vars in each scope. Passes may also create new vars.
  //         skip control vars and empty vars
Y
Yancey1989 已提交
573
  std::vector<details::VariableInfo> var_infos;
Q
Qiao Longfei 已提交
574 575 576 577 578 579
  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();
580 581 582

      member_->is_persistable_.emplace(node->Var()->Name(),
                                       node->Var()->Persistable());
Y
Yancey1989 已提交
583 584
    }
  }
Y
Yancey1989 已提交
585

586 587 588 589 590 591 592 593 594 595 596 597
  std::unordered_map<Scope *, Scope *> scope_map;
  for (auto *scope : member_->local_scopes_) {
    auto &local_exec_scope = scope->NewScope();
    member_->local_exec_scopes_.emplace_back(&local_exec_scope);
    scope_map.emplace(scope, &local_exec_scope);
  }

  PADDLE_ENFORCE_EQ(member_->local_scopes_.size(),
                    member_->local_exec_scopes_.size());

  std::vector<ir::Graph *> final_graphs;

C
chengduo 已提交
598
  if (member_->build_strategy_.async_mode_) {
Q
can run  
Qiao Longfei 已提交
599 600
    VLOG(3) << "use AsyncSSAGraphExecutor";
    member_->executor_.reset(new details::AsyncSSAGraphExecutor(
601 602 603
        exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
        member_->places_, async_graphs));
    final_graphs = async_graphs;
C
chengduo 已提交
604
  } else if (member_->build_strategy_.enable_parallel_graph_) {
Q
can run  
Qiao Longfei 已提交
605
    VLOG(3) << "use ParallelSSAGraphExecutor";
Y
Yancey1989 已提交
606
#ifdef PADDLE_WITH_CUDA
Y
Yancey1989 已提交
607 608
    // TODO(Yancey1989): Remove passing in the main_program when
    // allreduce_seq_pass doesn't need it as the attr.
609 610 611 612 613
    auto *pg_exe = new details::ParallelSSAGraphExecutor(
        exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
        member_->places_, graph);
    final_graphs = pg_exe->Graphs();
    member_->executor_.reset(pg_exe);
Y
Yancey1989 已提交
614 615 616 617
#else
    PADDLE_THROW(
        "Paddle should be compiled with CUDA for ParallelGraph Execution.");
#endif
Y
yuyang18 已提交
618
  } else {
Y
Yancey1989 已提交
619
    if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
Q
can run  
Qiao Longfei 已提交
620
      VLOG(3) << "use ThreadedSSAGraphExecutor";
Y
Yancey1989 已提交
621
      member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
622 623
          exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
          member_->places_, graph));
Y
Yancey1989 已提交
624
    } else {
Q
can run  
Qiao Longfei 已提交
625
      VLOG(3) << "use FastThreadedSSAGraphExecutor";
Y
Yancey1989 已提交
626
      member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
627 628
          exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
          member_->places_, graph));
Y
Yancey1989 已提交
629
    }
630
    final_graphs.emplace_back(graph);
C
chengduoZH 已提交
631
  }
Y
yuyang18 已提交
632

Q
can run  
Qiao Longfei 已提交
633
  VLOG(3) << "use ScopeBufferedSSAGraphExecutor";
C
chengduo 已提交
634
  if (!member_->build_strategy_.async_mode_) {
Q
Qiao Longfei 已提交
635
    member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
636 637 638 639 640 641 642 643 644
        exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
        std::move(var_infos), member_->places_, std::move(member_->executor_)));
  }

  for (auto *g : final_graphs) {
    auto ops = ir::FilterByNodeWrapper<details::OpHandleBase>(*g);
    for (auto *op : ops) {
      op->SetLocalExecScopes(scope_map);
    }
Q
Qiao Longfei 已提交
645
  }
Y
Yu Yang 已提交
646 647
}

Y
Yancey1989 已提交
648
void ParallelExecutor::BCastParamsToDevices(
Y
Yan Xu 已提交
649
    const std::vector<std::string> &vars, int trainer_id) const {
Q
Qiao Longfei 已提交
650
  VLOG(3) << "BCastParamsToDevices";
X
Xin Pan 已提交
651
  // the initializing bcast, all vars would be bcast from device(0).
652
  for (auto &var : vars) {
X
Xin Pan 已提交
653
    framework::Variable *main_var = member_->local_scopes_[0]->FindVar(var);
J
JiayiFeng 已提交
654
    if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
655 656 657 658
      continue;
    }

    auto &main_tensor = main_var->Get<LoDTensor>();
659
    if (!main_tensor.IsInitialized()) {
M
minqiyang 已提交
660
      VLOG(3) << "one in var not inited, return!";
661 662
      continue;
    }
663 664
    auto &dims = main_tensor.dims();
    if (paddle::platform::is_gpu_place(main_tensor.place())) {
665
#if defined(PADDLE_WITH_NCCL)
666
      std::vector<void *> buffers;
C
chengduo 已提交
667
      buffers.reserve(member_->places_.size());
668 669 670 671 672
      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;
673

Y
Yan Xu 已提交
674
        if (i == 0 && trainer_id == 0) {
675 676
          buffer = const_cast<void *>(main_tensor.data<void>());
        } else {
Y
Yu Yang 已提交
677
          auto local_scope = member_->local_scopes_[i];
678
          auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
Y
Update  
Yu Yang 已提交
679
          t->Resize(dims);
680
          buffer = t->mutable_data(place, main_tensor.type());
Y
Update  
Yu Yang 已提交
681
        }
682
        buffers.push_back(buffer);
683
      }
684

685 686 687
      PADDLE_ENFORCE_EQ(member_->places_.size(), buffers.size(),
                        "variables' buffer size to bcast NOT equal to places");
      {
688
        auto *nccl_ctxs = member_->nccl_ctxs_->DefaultFlatCtx();
689 690
        platform::NCCLGroupGuard guard;
        for (size_t i = 0; i < member_->places_.size(); ++i) {
691
          auto &nccl_ctx = nccl_ctxs->at(member_->places_[i]);
X
Xin Pan 已提交
692 693
          platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
                                       nccl_ctx.comm_, nccl_ctx.stream());
694
        }
695
        nccl_ctxs->WaitAll();
696
      }
C
chengduoZH 已提交
697
#endif
698 699
    } else {
      platform::CPUPlace cpu;
C
chengduo 已提交
700
      for (size_t i = 1; i < member_->places_.size(); ++i) {
701 702
        auto local_scope = member_->local_scopes_[i];
        auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
C
chengduo 已提交
703

Q
Qiao Longfei 已提交
704
        auto copy_memory = [&] {
705 706 707
          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
708 709
        };

Q
Qiao Longfei 已提交
710
        auto share_memory = [&] { t->ShareDataWith(main_tensor); };
Q
can run  
Qiao Longfei 已提交
711 712 713 714 715 716 717

        // 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();
718
        } else {
Q
can run  
Qiao Longfei 已提交
719
          share_memory();
720
        }
Y
Yu Yang 已提交
721
      }
Y
Stash  
Yu Yang 已提交
722 723
    }
  }
Y
Yu Yang 已提交
724
}
Y
Yu Yang 已提交
725

726 727
FeedFetchList ParallelExecutor::Run(
    const std::vector<std::string> &fetch_tensors) {
728
  VLOG(3) << "enter ParallelExecutor Run";
Y
Yu Yang 已提交
729 730 731
#ifdef WITH_GPERFTOOLS
  if (gProfileStarted) {
    ProfilerFlush();
S
sneaxiy 已提交
732 733
  }
#endif
Y
Yu Yang 已提交
734

X
Xin Pan 已提交
735
  platform::RecordBlock b(0);
736 737 738

  ir::SkipMemOptVarsGuard guard(&(member_->mem_opt_var_infos_), fetch_tensors,
                                member_->HasGarbageCollectors());
739 740

  VLOG(3) << "ParallelExecutor begin to run member_->executor_->Run";
S
sneaxiy 已提交
741
  auto fetch_data = member_->executor_->Run(fetch_tensors);
742
  return fetch_data;
Y
Yu Yang 已提交
743
}
Y
Yu Yang 已提交
744

Y
Yu Yang 已提交
745 746 747 748 749 750 751
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];
    for (auto &pair : map) {
752
      bool is_persistable = member_->IsPersistable(pair.first);
753 754 755
      if (!is_persistable) {
        member_->SetSkipMemoryReuse(i, pair.first);
      }
756 757 758 759 760
      auto *feed_scope = is_persistable ? member_->local_scopes_[i]
                                        : member_->local_exec_scopes_[i];
      auto *feed_var = feed_scope->Var(pair.first);

      auto *trg = feed_var->GetMutable<LoDTensor>();
Y
Yu Yang 已提交
761 762 763 764 765 766 767 768
      trg->ShareDataWith(pair.second);
      trg->set_lod(pair.second.lod());
    }
  }
}

void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
    const std::unordered_map<std::string, LoDTensor> &tensors) {
769
  size_t num_places = member_->places_.size();
770
  for (auto &pair : tensors) {
771 772 773 774
    bool is_persistable = member_->IsPersistable(pair.first);
    VLOG(3) << "Split " << (is_persistable ? "persistable" : "no persistable")
            << " data (" << pair.first << "), dim:" << pair.second.dims()
            << ", place: " << pair.second.place();
Y
Yu Yang 已提交
775
    auto lod_tensors = pair.second.SplitLoDTensor(member_->places_);
776 777
    bool is_cpu_place = platform::is_cpu_place(member_->places_.front());
    if (!is_persistable && num_places != lod_tensors.size()) {
C
chengduo 已提交
778
      auto error_info = string::Sprintf(
779 780 781
          "The number(%d) of samples[%s] of current batch is less than the "
          "count(%d) of devices(%s), currently, it is not allowed. ",
          lod_tensors.size(), pair.first, num_places,
C
chengduo 已提交
782 783 784 785 786 787 788
          (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);
789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814
    } else if (is_persistable) {
      if (lod_tensors.size() == 1) {
        lod_tensors.reserve(num_places);
        auto &tensor = lod_tensors.front();
        PADDLE_ENFORCE_EQ(tensor.dims(), pair.second.dims(),
                          "The dim doesn't match.");
        PADDLE_ENFORCE_EQ(tensor.place(), member_->places_.at(0),
                          "The place doesn't match.");
        for (size_t i = 1; i < num_places; ++i) {
          lod_tensors.emplace_back();
          auto &tmp = lod_tensors.back();
          framework::TensorCopy(pair.second, member_->places_.at(i), &tmp);
        }
      }
      if (lod_tensors.size() != num_places) {
        auto error_info = string::Sprintf(
            "The number(%d) of samples[%s] of the current batch does not match "
            "the count(%d) of devices(%s). Because that %s is a persistable "
            "variable, you can feed just one sample, in that case, the input "
            "sample will be copied in %d copies and be sent to different "
            "places separately. If you need that different place has different "
            "value, you should feed %d samples.",
            lod_tensors.size(), pair.first, num_places,
            (is_cpu_place ? "CPU" : "GPU"), pair.first, num_places, num_places);
        PADDLE_THROW(error_info);
      }
C
chengduo 已提交
815
    }
816

817
    for (size_t j = 0; j < num_places; ++j) {
818 819 820 821 822
      auto *feed_scope = is_persistable ? member_->local_scopes_[j]
                                        : member_->local_exec_scopes_[j];
      auto *feed_var = feed_scope->Var(pair.first);

      auto t = feed_var->GetMutable<LoDTensor>();
823 824
      t->ShareDataWith(lod_tensors[j]);
      t->set_lod(lod_tensors[j].lod());
X
Xin Pan 已提交
825 826 827 828
    }
  }
}

X
Xin Pan 已提交
829 830 831 832 833 834 835
ParallelExecutor::~ParallelExecutor() {
  for (auto &p : member_->places_) {
    platform::DeviceContextPool::Instance().Get(p)->Wait();
  }
  delete member_;
}

836
bool ParallelExecutor::EnableParallelGraphExecution(
X
Xin Pan 已提交
837
    const ir::Graph &graph, const ExecutionStrategy &exec_strategy,
838
    const BuildStrategy &build_strategy) const {
839 840 841
  if (!FLAGS_enable_parallel_graph) {
    return false;
  }
842

Y
Yancey1989 已提交
843
  bool enable_parallel_graph = true;
844

X
Xin Pan 已提交
845 846 847 848 849 850 851 852 853 854 855 856 857
  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;
      }
858 859 860
    }
  }

861
  if (!member_->use_all_reduce_ || !member_->use_cuda_) {
Y
Yancey1989 已提交
862
    if (build_strategy.enable_sequential_execution_ ||
863
        exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental) {
Y
Yancey1989 已提交
864
      enable_parallel_graph = false;
865 866 867 868 869 870 871 872 873
    }
  }

#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
874
  return enable_parallel_graph;
875 876
}

Y
Yu Yang 已提交
877
}  // namespace framework
Y
Yang Yang 已提交
878
}  // namespace paddle
S
sneaxiy 已提交
879

S
sneaxiy 已提交
880
USE_PASS(reference_count_pass);
S
sneaxiy 已提交
881
USE_PASS(eager_deletion_pass);
882
USE_PASS(buffer_shared_inplace_pass);
883
USE_PASS(buffer_shared_cross_op_memory_reuse_pass);