executor.cc 17.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Q
qijun 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

Y
Yi Wang 已提交
15
#include "paddle/fluid/framework/executor.h"
Y
Yang Yang 已提交
16

Y
Yi Wang 已提交
17 18 19 20 21
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
G
gongweibao 已提交
22
#include "paddle/fluid/operators/detail/macros.h"
Y
Yi Wang 已提交
23
#include "paddle/fluid/platform/place.h"
X
Xin Pan 已提交
24
#include "paddle/fluid/platform/profiler.h"
Y
Yang Yu 已提交
25

D
dzhwinter 已提交
26
DECLARE_bool(benchmark);
27
DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
Q
qijun 已提交
28 29 30

namespace paddle {
namespace framework {
X
Xin Pan 已提交
31 32 33 34 35
namespace {
// block id starts from 0. This id is used to represent the codeblock
// wrapping the first block 0.
int kProgramId = -1;
}  // namespace
Q
qijun 已提交
36

Q
Qiao Longfei 已提交
37 38
ExecutorPrepareContext::ExecutorPrepareContext(
    const framework::ProgramDesc& prog, size_t block_id)
S
sneaxiy 已提交
39 40 41 42 43
    : prog_(prog), block_id_(block_id) {
  if (GetEagerDeletionThreshold() >= 0) {
    ref_cnts_ = GetNonPersistableReferenceCount<int>(prog_, block_id_);
  }
}
Y
Yu Yang 已提交
44

Q
Qiao Longfei 已提交
45
ExecutorPrepareContext::~ExecutorPrepareContext() {
46
  VLOG(50) << "destroy ExecutorPrepareContext";
Q
Qiao Longfei 已提交
47
}
Y
Yu Yang 已提交
48

S
sneaxiy 已提交
49 50 51 52 53 54 55 56 57 58 59 60 61 62
template <typename RefCntMap>
static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
                                GarbageCollector<Tensor>* gc,
                                RefCntMap* ref_cnts) {
  std::unordered_set<Tensor*> erase_tensors;

  auto handler = [&](const VariableNameMap& name_map) {
    for (auto& name_pair : name_map) {
      for (auto& name : name_pair.second) {
        auto it = ref_cnts->find(name);
        if (it == ref_cnts->end()) continue;
        if ((it->second)-- == 1) {
          auto* var = scope.FindVar(name);
          if (var != nullptr) {
63
            VLOG(100) << "Erase tensor \'" << name << "\'";
S
sneaxiy 已提交
64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
            if (var->IsType<LoDTensor>()) {
              erase_tensors.insert(var->GetMutable<LoDTensor>());
            } else if (var->IsType<SelectedRows>()) {
              erase_tensors.insert(
                  var->GetMutable<SelectedRows>()->mutable_value());
            }
          }
        }
      }
    }
  };

  handler(op->Inputs());
  handler(op->Outputs());

  if (!erase_tensors.empty()) {
    gc->Add(erase_tensors);
  }
}

D
dzhwinter 已提交
84
Executor::Executor(const platform::Place& place) : place_(place) {}
Q
qijun 已提交
85

Y
Yancey1989 已提交
86
void Executor::Close() {
W
Wu Yi 已提交
87
#ifdef PADDLE_WITH_DISTRIBUTE
W
Wu Yi 已提交
88 89
  // TODO(typhoonzero): complete message will need to use real trainer_id,
  // except 0.
Y
Yancey1989 已提交
90
  ::paddle::operators::distributed::RPCClient::GetInstance<
W
Wu Yi 已提交
91
      ::paddle::operators::distributed::GRPCClient>(0)
Y
Yancey1989 已提交
92
      ->SendComplete();
W
Wu Yi 已提交
93
#endif
Y
Yancey1989 已提交
94
}
W
Wu Yi 已提交
95

Y
Stash  
Yu Yang 已提交
96
void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
97
  if (var_type == proto::VarType::LOD_TENSOR) {
Q
QI JUN 已提交
98
    var->GetMutable<LoDTensor>();
99
  } else if (var_type == proto::VarType::SELECTED_ROWS) {
Q
QI JUN 已提交
100
    var->GetMutable<SelectedRows>();
101
  } else if (var_type == proto::VarType::FEED_MINIBATCH) {
Q
QI JUN 已提交
102
    var->GetMutable<FeedFetchList>();
103
  } else if (var_type == proto::VarType::FETCH_LIST) {
Q
QI JUN 已提交
104
    var->GetMutable<FeedFetchList>();
105
  } else if (var_type == proto::VarType::STEP_SCOPES) {
X
Xin Pan 已提交
106
    var->GetMutable<std::vector<framework::Scope*>>();
107
  } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
Y
Yu Yang 已提交
108
    var->GetMutable<LoDRankTable>();
109
  } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
Y
Yu Yang 已提交
110
    var->GetMutable<LoDTensorArray>();
111
  } else if (var_type == proto::VarType::PLACE_LIST) {
Y
Yang Yu 已提交
112
    var->GetMutable<platform::PlaceList>();
113
  } else if (var_type == proto::VarType::READER) {
F
fengjiayi 已提交
114
    var->GetMutable<ReaderHolder>();
T
typhoonzero 已提交
115 116
  } else if (var_type == proto::VarType::RAW) {
    // GetMutable will be called in operator
Q
QI JUN 已提交
117 118
  } else {
    PADDLE_THROW(
Y
Yu Yang 已提交
119
        "Variable type %d is not in "
F
fengjiayi 已提交
120
        "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
X
Xin Pan 已提交
121
        "LOD_RANK_TABLE, PLACE_LIST, READER, RAW]",
Y
Yu Yang 已提交
122
        var_type);
Q
QI JUN 已提交
123 124 125
  }
}

L
Liu Yiqun 已提交
126 127 128
void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
                               int block_id) {
  auto& global_block = pdesc.Block(block_id);
129 130 131 132 133 134 135 136 137 138 139 140 141 142

  const Scope* ancestor_scope = scope;
  while (ancestor_scope->parent()) {
    ancestor_scope = ancestor_scope->parent();
  }

  if (ancestor_scope != scope) {
    for (auto& var : global_block.AllVars()) {
      if (var->Name() == framework::kEmptyVarName) {
        continue;
      }

      if (var->Persistable()) {
        auto* ptr = const_cast<Scope*>(ancestor_scope)->Var(var->Name());
143
        InitializeVariable(ptr, var->GetType());
144 145
        VLOG(30) << "Create Variable " << var->Name()
                 << " global, which pointer is " << ptr;
146 147
      } else {
        auto* ptr = scope->Var(var->Name());
148
        InitializeVariable(ptr, var->GetType());
149 150
        VLOG(30) << "Create Variable " << var->Name()
                 << " locally, which pointer is " << ptr;
151 152 153 154 155
      }
    }
  } else {
    for (auto& var : global_block.AllVars()) {
      auto* ptr = scope->Var(var->Name());
156
      InitializeVariable(ptr, var->GetType());
157 158
      VLOG(30) << "Create variable " << var->Name() << ", which pointer is "
               << ptr;
159 160 161 162
    }
  }
}

Y
Yu Yang 已提交
163
void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
T
typhoonzero 已提交
164
                   bool create_local_scope, bool create_vars) {
X
Xin Pan 已提交
165
  platform::RecordBlock b(block_id);
166
  if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
Q
Qiao Longfei 已提交
167 168
  auto ctx = Prepare(pdesc, block_id);
  RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
Q
qijun 已提交
169 170
}

171 172 173 174 175 176 177
// Check whether the block already has feed operators and feed_holder.
// Return false if the block does not have any feed operators.
// If some feed operators have been prepended to the block, check that
// the info contained in these feed operators matches the feed_targets
// and feed_holder_name. Raise exception when any mismatch is found.
// Return true if the block has feed operators and holder of matching info.
static bool has_feed_operators(
178
    const BlockDesc& block,
L
Liu Yiqun 已提交
179
    const std::map<std::string, const LoDTensor*>& feed_targets,
180 181
    const std::string& feed_holder_name) {
  size_t feed_count = 0;
182
  for (auto* op : block.AllOps()) {
183 184
    if (op->Type() == kFeedOpType) {
      feed_count++;
L
Liu Yiqun 已提交
185
      // The input variable's name of feed_op should be feed_holder_name.
186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
      PADDLE_ENFORCE_EQ(op->Input("X")[0], feed_holder_name,
                        "Input to feed op should be '%s'", feed_holder_name);
      std::string feed_target_name = op->Output("Out")[0];
      PADDLE_ENFORCE(
          feed_targets.find(feed_target_name) != feed_targets.end(),
          "Feed operator output name '%s' cannot be found in 'feed_targets'",
          feed_target_name);
    }
  }

  if (feed_count > 0) {
    PADDLE_ENFORCE_EQ(
        feed_count, feed_targets.size(),
        "The number of feed operators should match 'feed_targets'");

201
    if (!feed_holder_name.empty()) {
L
Liu Yiqun 已提交
202
      // When feed operator are present, so should be feed_holder.
203 204 205 206 207 208 209
      auto var = block.FindVar(feed_holder_name);
      PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable",
                              feed_holder_name);
      PADDLE_ENFORCE_EQ(var->GetType(), proto::VarType::FEED_MINIBATCH,
                        "'%s' variable should be 'FEED_MINIBATCH' type",
                        feed_holder_name);
    }
210 211 212 213 214 215 216 217 218 219 220 221
  }

  return feed_count > 0;
}

// Check whether the block already has fetch operators and fetch_holder.
// Return false if the block does not have any fetch operators.
// If some fetch operators have been appended to the block, check that
// the info contained in these fetch operators matches the fetch_targets
// and fetch_holder_name. Raise exception when any mismatch is found.
// Return true if the block has fetch operators and holder of matching info.
static bool has_fetch_operators(
L
Liu Yiqun 已提交
222 223
    const BlockDesc& block,
    const std::map<std::string, LoDTensor*>& fetch_targets,
224 225
    const std::string& fetch_holder_name) {
  size_t fetch_count = 0;
226
  for (auto* op : block.AllOps()) {
227 228
    if (op->Type() == kFetchOpType) {
      fetch_count++;
L
Liu Yiqun 已提交
229
      // The output variable's name of fetch_op should be fetch_holder_name.
230 231 232 233 234 235 236 237 238 239 240 241 242 243 244
      PADDLE_ENFORCE_EQ(op->Output("Out")[0], fetch_holder_name,
                        "Output of fetch op should be '%s'", fetch_holder_name);
      std::string fetch_target_name = op->Input("X")[0];
      PADDLE_ENFORCE(
          fetch_targets.find(fetch_target_name) != fetch_targets.end(),
          "Fetch operator input name '%s' cannot be found in 'fetch_targets'",
          fetch_target_name);
    }
  }

  if (fetch_count > 0) {
    PADDLE_ENFORCE_EQ(
        fetch_count, fetch_targets.size(),
        "The number of fetch operators should match 'fetch_targets'");

245
    if (!fetch_holder_name.empty()) {
L
Liu Yiqun 已提交
246
      // When fetch operator are present, so should be fetch_holder.
247 248 249 250 251 252 253
      auto var = block.FindVar(fetch_holder_name);
      PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable",
                              fetch_holder_name);
      PADDLE_ENFORCE_EQ(var->GetType(), proto::VarType::FETCH_LIST,
                        "'%s' variable should be 'FETCH_LIST' type",
                        fetch_holder_name);
    }
254 255 256 257 258 259
  }

  return fetch_count > 0;
}

void Executor::Run(const ProgramDesc& program, Scope* scope,
260 261
                   std::map<std::string, const LoDTensor*>* feed_targets,
                   std::map<std::string, LoDTensor*>* fetch_targets,
W
Wu Yi 已提交
262 263
                   bool create_local_scope, bool create_vars,
                   const std::string& feed_holder_name,
264
                   const std::string& fetch_holder_name) {
X
Xin Pan 已提交
265
  platform::RecordBlock b(kProgramId);
266
  if (FLAGS_use_mkldnn) EnableMKLDNN(program);
267
  bool has_feed_ops =
268
      has_feed_operators(program.Block(0), *feed_targets, feed_holder_name);
269
  bool has_fetch_ops =
270
      has_fetch_operators(program.Block(0), *fetch_targets, fetch_holder_name);
271 272

  ProgramDesc* copy_program = const_cast<ProgramDesc*>(&program);
S
sneaxiy 已提交
273
  std::unique_ptr<ProgramDesc> unique_ptr_of_copy_program;
274
  if (!has_feed_ops || !has_fetch_ops) {
S
sneaxiy 已提交
275 276
    unique_ptr_of_copy_program.reset(new ProgramDesc(program));
    copy_program = unique_ptr_of_copy_program.get();
277
  }
278 279
  auto* global_block = copy_program->MutableBlock(0);

280
  if (!has_feed_ops) {
281 282
    // create feed_holder variable
    auto* feed_holder = global_block->Var(feed_holder_name);
283
    feed_holder->SetType(proto::VarType::FEED_MINIBATCH);
284 285 286
    feed_holder->SetPersistable(true);

    int i = 0;
287
    for (auto& feed_target : (*feed_targets)) {
288
      std::string var_name = feed_target.first;
289
      VLOG(30) << "feed target's name: " << var_name;
290 291 292 293 294 295 296 297 298 299 300 301 302

      // prepend feed op
      auto* op = global_block->PrependOp();
      op->SetType(kFeedOpType);
      op->SetInput("X", {feed_holder_name});
      op->SetOutput("Out", {var_name});
      op->SetAttr("col", {static_cast<int>(i)});
      op->CheckAttrs();

      i++;
    }
  }

303
  if (!has_fetch_ops) {
304 305
    // create fetch_holder variable
    auto* fetch_holder = global_block->Var(fetch_holder_name);
306
    fetch_holder->SetType(proto::VarType::FETCH_LIST);
307 308 309
    fetch_holder->SetPersistable(true);

    int i = 0;
310
    for (auto& fetch_target : (*fetch_targets)) {
311
      std::string var_name = fetch_target.first;
312
      VLOG(30) << "fetch target's name: " << var_name;
313 314 315 316 317 318 319 320 321 322 323 324 325

      // append fetch op
      auto* op = global_block->AppendOp();
      op->SetType(kFetchOpType);
      op->SetInput("X", {var_name});
      op->SetOutput("Out", {fetch_holder_name});
      op->SetAttr("col", {static_cast<int>(i)});
      op->CheckAttrs();

      i++;
    }
  }

326
  auto ctx = Prepare(*copy_program, 0);
W
Wu Yi 已提交
327 328 329
  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets,
                     create_local_scope, create_vars, feed_holder_name,
                     fetch_holder_name);
330 331
}

Q
Qiao Longfei 已提交
332 333
std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
    const ProgramDesc& program, int block_id) {
Q
Qiyang Min 已提交
334 335
  std::unique_ptr<ExecutorPrepareContext> ctx(
      new ExecutorPrepareContext(program, block_id));
Y
Yu Yang 已提交
336 337 338 339 340
  PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
  auto& block = program.Block(block_id);
  for (auto& op_desc : block.AllOps()) {
    ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
  }
Q
Qiyang Min 已提交
341
  return ctx;
Y
Yu Yang 已提交
342 343
}

T
refine  
typhoonzero 已提交
344
std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
T
typhoonzero 已提交
345 346 347 348 349 350 351 352 353 354 355 356 357 358
    const ProgramDesc& program, const std::vector<int>& block_ids) {
  std::vector<std::shared_ptr<ExecutorPrepareContext>> result;
  for (auto& bid : block_ids) {
    auto* ctx = new ExecutorPrepareContext(program, bid);
    PADDLE_ENFORCE_LT(static_cast<size_t>(bid), program.Size());
    auto& block = program.Block(bid);
    for (auto& op_desc : block.AllOps()) {
      ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
    }
    result.push_back(std::shared_ptr<ExecutorPrepareContext>(ctx));
  }
  return result;
}

Y
Yu Yang 已提交
359
void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
Q
qiaolongfei 已提交
360 361
                                  bool create_local_scope, bool create_vars,
                                  bool keep_kids) {
Y
Yu Yang 已提交
362 363 364 365
  Scope* local_scope = scope;
  if (create_vars) {
    if (create_local_scope) {
      local_scope = &scope->NewScope();
366 367
    }
    CreateVariables(ctx->prog_, local_scope, ctx->block_id_);
L
Liu Yiqun 已提交
368
  }
Y
Yu Yang 已提交
369

S
sneaxiy 已提交
370 371
  int64_t max_memory_size = GetEagerDeletionThreshold();
  std::unique_ptr<GarbageCollector<Tensor>> gc;
S
sneaxiy 已提交
372 373 374 375 376 377
  // WhileOp would set keep_kids to false
  // WhileGradOp would need the scopes created in WhileOp
  // Perhaps, we should not perform eager deletion in WhileOp
  // The scopes and variables created by WhileOp would be deleted
  // in WhileGradOp.
  if (max_memory_size >= 0 && !keep_kids) {
S
sneaxiy 已提交
378
    ctx->ResetReferenceCount();
S
sneaxiy 已提交
379 380 381 382 383 384 385 386 387 388 389 390 391
#ifdef PADDLE_WITH_CUDA
    if (platform::is_gpu_place(place_)) {
      gc.reset(new DefaultStreamGarbageCollector<Tensor>(
          boost::get<platform::CUDAPlace>(place_), max_memory_size));
    } else {
#endif
      gc.reset(new CPUGarbageCollector<Tensor>(
          boost::get<platform::CPUPlace>(place_), max_memory_size));
#ifdef PADDLE_WITH_CUDA
    }
#endif
  }

Y
Yu Yang 已提交
392
  for (auto& op : ctx->ops_) {
393
    op->Run(*local_scope, place_);
S
sneaxiy 已提交
394 395

    if (gc != nullptr) {
S
sneaxiy 已提交
396 397
      DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
                          &(ctx->cur_ref_cnts_));
S
sneaxiy 已提交
398
    }
Y
Yang Yang 已提交
399

Y
Yu Yang 已提交
400
    if (FLAGS_benchmark) {
401 402
      VLOG(20) << "Memory used after operator " + op->Type() + " running: "
               << memory::memory_usage(place_);
Y
Yu Yang 已提交
403 404
    }
  }
S
sneaxiy 已提交
405

S
sneaxiy 已提交
406
  if (gc != nullptr) {
S
sneaxiy 已提交
407
    gc->Wait();
S
sneaxiy 已提交
408
  } else {
S
sneaxiy 已提交
409
    platform::DeviceContextPool::Instance().Get(place_)->Wait();
S
sneaxiy 已提交
410
  }
S
sneaxiy 已提交
411

Q
qiaolongfei 已提交
412
  if (local_scope != scope) {
Y
Yu Yang 已提交
413
    scope->DeleteScope(local_scope);
414
  } else {
Q
qiaolongfei 已提交
415 416 417 418 419
    if (!keep_kids) {
      // By default, we should delete all kid scopes after run executor because
      // some operators may create local scope when running, such as while_op.
      // But when while_op also create a local executor to run it's sub block,
      // the sub scopes it created should not be dropped immediately, because
Q
qiaolongfei 已提交
420 421
      // while_grad_op will use some variables created during while_op run, so
      // we need to keep the kids and wait for the outer executor to drop them.
Q
qiaolongfei 已提交
422 423
      scope->DropKids();
    }
Y
Yu Yang 已提交
424
  }
Q
qiaolongfei 已提交
425

Y
Yu Yang 已提交
426
  if (FLAGS_benchmark) {
427 428 429 430
    VLOG(20) << "-------------------------------------------------------";
    VLOG(20) << "Memory used after deleting local scope: "
             << memory::memory_usage(place_);
    VLOG(20) << "-------------------------------------------------------";
Y
Yu Yang 已提交
431 432 433
  }
}

434 435
void Executor::RunPreparedContext(
    ExecutorPrepareContext* ctx, Scope* scope,
436
    std::map<std::string, const LoDTensor*>* feed_targets,
W
Wu Yi 已提交
437 438 439
    std::map<std::string, LoDTensor*>* fetch_targets, bool create_local_scope,
    bool create_vars, const std::string& feed_holder_name,
    const std::string& fetch_holder_name) {
440 441
  auto& global_block = ctx->prog_.Block(ctx->block_id_);

442
  PADDLE_ENFORCE(
443
      has_feed_operators(global_block, *feed_targets, feed_holder_name),
444 445
      "Program in ExecutorPrepareContext should has feed_ops.");
  PADDLE_ENFORCE(
446
      has_fetch_operators(global_block, *fetch_targets, fetch_holder_name),
447 448
      "Program in the prepared context should has fetch_ops.");

449 450 451 452 453
  // map the data of feed_targets to feed_holder
  for (auto* op : global_block.AllOps()) {
    if (op->Type() == kFeedOpType) {
      std::string feed_target_name = op->Output("Out")[0];
      int idx = boost::get<int>(op->GetAttr("col"));
454 455
      SetFeedVariable(scope, *(*feed_targets)[feed_target_name],
                      feed_holder_name, idx);
456 457 458
    }
  }

W
Wu Yi 已提交
459
  RunPreparedContext(ctx, scope, create_local_scope, create_vars);
460 461 462 463 464 465

  // obtain the data of fetch_targets from fetch_holder
  for (auto* op : global_block.AllOps()) {
    if (op->Type() == kFetchOpType) {
      std::string fetch_target_name = op->Input("X")[0];
      int idx = boost::get<int>(op->GetAttr("col"));
466
      *(*fetch_targets)[fetch_target_name] =
467 468 469 470 471
          GetFetchVariable(*scope, fetch_holder_name, idx);
    }
  }
}

472 473
void Executor::EnableMKLDNN(const ProgramDesc& program) {
#ifdef PADDLE_WITH_MKLDNN
474
  VLOG(30) << "use_mkldnn=True";
475 476 477 478 479 480 481 482
  for (size_t bid = 0; bid < program.Size(); ++bid) {
    auto* block = const_cast<ProgramDesc&>(program).MutableBlock(bid);
    for (auto* op : block->AllOps()) {
      if (op->HasAttr("use_mkldnn")) {
        op->SetAttr("use_mkldnn", true);
      }
    }
  }
483 484 485
#else
  LOG(WARNING)
      << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
486 487 488
#endif
}

Q
qijun 已提交
489 490
}  // namespace framework
}  // namespace paddle