parallel_executor.cc 14.5 KB
Newer Older
Y
Yang Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/framework/parallel_executor.h"
C
chengduoZH 已提交
16
#include <string>
17
#include <tuple>
Q
qiaolongfei 已提交
18
#include <vector>
C
chengduo 已提交
19
#include "paddle/fluid/framework/ir/graph_helper.h"
Y
Yu Yang 已提交
20

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

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

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

Y
Yang Yang 已提交
34
namespace paddle {
Y
Yu Yang 已提交
35 36
namespace framework {

Y
Yu Yang 已提交
37 38 39
class ParallelExecutorPrivate {
 public:
  explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
Y
Yu Yang 已提交
40
      : places_(places) {}
Y
Yu Yang 已提交
41

42 43 44 45 46 47 48 49 50 51 52
  ~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
53

S
sneaxiy 已提交
54 55 56 57 58 59 60 61 62 63 64 65 66 67
  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
68
      }
S
sneaxiy 已提交
69
      runtime_ref_cnts_[i].erase(fetched_var_name);
S
sneaxiy 已提交
70 71 72
    }
  }

Y
Yu Yang 已提交
73 74
  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
75
  Scope *global_scope_;  // not owned
Y
Yu Yang 已提交
76
  std::unique_ptr<details::SSAGraphExecutor> executor_;
Y
Yu Yang 已提交
77

P
peizhilin 已提交
78
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
Y
Yu Yang 已提交
79
  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
Y
Yu Yang 已提交
80
#endif
C
chengduoZH 已提交
81 82
  bool own_local_scope_;
  bool use_cuda_;
83
  bool use_all_reduce_;
S
sneaxiy 已提交
84

S
sneaxiy 已提交
85 86 87 88 89 90
  // 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
91 92
};

S
sneaxiy 已提交
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
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;
    }
#ifdef PADDLE_WITH_CUDA
    GarbageCollector<Tensor> *gc = nullptr;
    if (platform::is_gpu_place(place)) {
      if (IsFastEagerDeletionModeEnabled()) {
        gc = new UnsafeFastGPUGarbageCollector<Tensor>(
            boost::get<platform::CUDAPlace>(place), max_memory_size);
      } else {
        gc = new StreamGarbageCollector<Tensor>(
            boost::get<platform::CUDAPlace>(place), max_memory_size);
      }
      VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
    } else if (platform::is_cpu_place(place)) {
#endif
      gc = new CPUGarbageCollector<Tensor>(
          boost::get<platform::CPUPlace>(place), max_memory_size);
      VLOG(10) << "Created GarbageCollector at " << place;
#ifdef PADDLE_WITH_CUDA
    }
#endif

    if (gc) {
      gcs_[place] = std::unique_ptr<GarbageCollector<Tensor>>(gc);
    }
  }

  if (gcs_.empty()) {
    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";

    graph->SetNotOwned(details::kGarbageCollector, &gcs_);
  }

  return graph;
}

154 155 156 157
std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
  return member_->local_scopes_;
}

Y
Yu Yang 已提交
158
ParallelExecutor::ParallelExecutor(
159
    const std::vector<platform::Place> &places,
Y
Yu Yang 已提交
160
    const std::unordered_set<std::string> &params,
161 162
    const std::unordered_set<std::string> &bcast_vars,
    const ProgramDesc &main_program, const std::string &loss_var_name,
Y
yuyang18 已提交
163
    Scope *scope, const std::vector<Scope *> &local_scopes,
164
    const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy,
165
    size_t num_trainers, size_t trainer_id)
Y
Yu Yang 已提交
166
    : member_(new ParallelExecutorPrivate(places)) {
Y
Yu Yang 已提交
167
  member_->global_scope_ = scope;
168
  member_->use_cuda_ = exec_strategy.use_cuda_;
169 170 171 172 173 174 175 176
  member_->use_all_reduce_ =
      build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce;

  if (!member_->use_all_reduce_) {
    PADDLE_ENFORCE(places.size() > 1,
                   "If you set build_strategy.reduce with 'Reduce',"
                   "the number of places must be greater than 1.");
  }
Y
Yu Yang 已提交
177

178
  // Step 1. Bcast the params to devs.
Y
Yu Yang 已提交
179
  // Create local scopes
180
  if (local_scopes.empty()) {
C
chengduoZH 已提交
181
    member_->own_local_scope_ = true;
Y
Yu Yang 已提交
182 183
    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
184
      member_->local_scopes_.emplace_back(&scope->NewScope());
185 186
    }
  } else {
C
chengduoZH 已提交
187
    member_->own_local_scope_ = false;
188 189
    PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
    for (size_t i = 0; i < member_->places_.size(); ++i) {
190
      member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
191
    }
Y
Yu Yang 已提交
192 193
  }

C
chengduoZH 已提交
194
  if (member_->use_cuda_) {
Y
Yu Yang 已提交
195
// Bcast Parameters to all GPUs
P
peizhilin 已提交
196
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
C
chengduoZH 已提交
197 198 199 200 201 202 203 204 205
    auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
    ncclUniqueId *nccl_id = nullptr;
    if (nccl_id_var != nullptr) {
      nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
    }
    member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
        member_->places_, nccl_id, num_trainers, trainer_id));
#else
    PADDLE_THROW("Not compiled with CUDA");
Y
Yu Yang 已提交
206
#endif
C
chengduoZH 已提交
207 208 209
  }

  if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
Y
Yancey1989 已提交
210
    BCastParamsToDevices(bcast_vars);
Y
Yu Yang 已提交
211
  }
212
// Startup Program has been run. All local scopes has correct parameters.
Y
yuyang18 已提交
213

214
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
X
Xin Pan 已提交
215
// ncclOp
P
peizhilin 已提交
216
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
217
  std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
X
Xin Pan 已提交
218
      main_program, member_->places_, loss_var_name, params,
219
      member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get());
S
sneaxiy 已提交
220 221 222 223 224
#else
  std::unique_ptr<ir::Graph> graph =
      build_strategy.Apply(main_program, member_->places_, loss_var_name,
                           params, member_->local_scopes_, member_->use_cuda_);
#endif
S
sneaxiy 已提交
225 226 227

  auto max_memory_size = GetEagerDeletionThreshold();
  if (max_memory_size >= 0) {
S
sneaxiy 已提交
228 229
    graph = member_->PrepareGCAndRefCnts(std::move(graph),
                                         static_cast<size_t>(max_memory_size));
S
sneaxiy 已提交
230
  }
X
Xin Pan 已提交
231

232 233 234 235 236 237 238 239 240 241 242
  // Step 3. Create vars in each scope. Passes may also create new vars.
  //         skip control vars and empty vars
  std::vector<details::VariableInfo> var_infos;
  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();
    }
  }
W
Wu Yi 已提交
243 244
  // If the loss_var_name is given, the number of graph should be only one.
  if (loss_var_name.size()) {
C
chengduo 已提交
245 246 247 248 249 250 251 252 253 254 255
    size_t graph_num = ir::GraphNum(*graph);
    if (graph_num > 1) {
      LOG(WARNING)
          << "The number of graph should be only one, "
             "but the current graph has "
          << ir::GraphNum(*graph)
          << " 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
256 257
  }

Y
yuyang18 已提交
258 259 260 261 262 263
  if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
    member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
        exec_strategy, member_->local_scopes_, places, std::move(graph)));
  } else {
    member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
        exec_strategy, member_->local_scopes_, places, std::move(graph)));
C
chengduoZH 已提交
264
  }
Y
yuyang18 已提交
265 266 267 268

  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
269 270
}

Y
Yancey1989 已提交
271
void ParallelExecutor::BCastParamsToDevices(
272
    const std::unordered_set<std::string> &vars) const {
X
Xin Pan 已提交
273
  // the initializing bcast, all vars would be bcast from device(0).
274
  for (auto &var : vars) {
X
Xin Pan 已提交
275
    framework::Variable *main_var = member_->local_scopes_[0]->FindVar(var);
J
JiayiFeng 已提交
276
    if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
277 278 279 280
      continue;
    }

    auto &main_tensor = main_var->Get<LoDTensor>();
281
    if (!main_tensor.IsInitialized()) {
M
minqiyang 已提交
282
      VLOG(3) << "one in var not inited, return!";
283 284
      continue;
    }
285 286
    auto &dims = main_tensor.dims();
    if (paddle::platform::is_gpu_place(main_tensor.place())) {
P
peizhilin 已提交
287
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
288
      std::vector<void *> buffers;
289 290 291 292 293
      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;
294

X
Xin Pan 已提交
295
        if (i == 0) {
296 297
          buffer = const_cast<void *>(main_tensor.data<void>());
        } else {
Y
Yu Yang 已提交
298
          auto local_scope = member_->local_scopes_[i];
299
          auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
Y
Update  
Yu Yang 已提交
300
          t->Resize(dims);
301
          buffer = t->mutable_data(place, main_tensor.type());
Y
Update  
Yu Yang 已提交
302
        }
303
        buffers.push_back(buffer);
304
      }
305

306 307 308 309 310 311
      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
312 313
          platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
                                       nccl_ctx.comm_, nccl_ctx.stream());
314
        }
315
        member_->nccl_ctxs_->WaitAll();
316
      }
C
chengduoZH 已提交
317 318 319
#else
      PADDLE_THROW("Not compiled with CUDA");
#endif
320 321
    } else {
      platform::CPUPlace cpu;
Y
Yancey1989 已提交
322
      for (size_t i = 0; i < member_->places_.size(); ++i) {
X
Xin Pan 已提交
323
        if (i == 0) continue;
Y
Yancey1989 已提交
324

325 326
        auto local_scope = member_->local_scopes_[i];
        auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
C
chengduo 已提交
327 328 329 330

        // FIXME(zcd): LR_DECAY_COUNTER should not be shared. This is a hot fix.
        if (member_->use_all_reduce_ || member_->use_cuda_ ||
            var == "@LR_DECAY_COUNTER@") {
331 332 333 334 335 336
          t->Resize(dims);
          t->mutable_data(cpu, main_tensor.type());
          paddle::framework::TensorCopy(main_tensor, cpu, t);
        } else {
          t->ShareDataWith(main_tensor);
        }
Y
Yu Yang 已提交
337
      }
Y
Stash  
Yu Yang 已提交
338 339
    }
  }
Y
Yu Yang 已提交
340
}
Y
Yu Yang 已提交
341

Y
Yu Yang 已提交
342 343
void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                           const std::string &fetched_var_name) {
X
Xin Pan 已提交
344
  platform::RecordBlock b(0);
S
sneaxiy 已提交
345 346
  if (member_->HasGarbageCollectors()) {
    member_->ResetRuntimeReferenceCount(fetch_tensors, fetched_var_name);
S
sneaxiy 已提交
347
  }
S
sneaxiy 已提交
348 349 350
  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
351
}
Y
Yu Yang 已提交
352

Y
Yu Yang 已提交
353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371
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_);
372 373 374 375 376
    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
377 378
    for (size_t j = 0; j < member_->places_.size(); ++j) {
      // TODO(panxy0718): Do I need to delete this var?
379
      auto t =
Y
Yu Yang 已提交
380
          member_->local_scopes_[j]->Var(pair.first)->GetMutable<LoDTensor>();
381 382
      t->ShareDataWith(lod_tensors[j]);
      t->set_lod(lod_tensors[j].lod());
X
Xin Pan 已提交
383 384 385 386
    }
  }
}

387
ParallelExecutor::~ParallelExecutor() {
388 389
  for (auto &p : member_->places_) {
    platform::DeviceContextPool::Instance().Get(p)->Wait();
C
chengduozh 已提交
390
  }
S
sneaxiy 已提交
391
  delete member_;
392 393
}

Y
Yu Yang 已提交
394
}  // namespace framework
Y
Yang Yang 已提交
395
}  // namespace paddle
S
sneaxiy 已提交
396

S
sneaxiy 已提交
397
USE_PASS(reference_count_pass);
S
sneaxiy 已提交
398
USE_PASS(eager_deletion_pass);