executor.cc 17.9 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
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
B
baojun-nervana 已提交
20
#include "paddle/fluid/framework/ngraph_operator.h"
Y
Yi Wang 已提交
21 22
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
23
#include "paddle/fluid/framework/transfer_scope_cache.h"
G
gongweibao 已提交
24
#include "paddle/fluid/operators/detail/macros.h"
Y
Yi Wang 已提交
25
#include "paddle/fluid/platform/place.h"
X
Xin Pan 已提交
26
#include "paddle/fluid/platform/profiler.h"
Y
Yang Yu 已提交
27

D
dzhwinter 已提交
28
DECLARE_bool(benchmark);
29
DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
B
baojun-nervana 已提交
30
DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
Q
qijun 已提交
31 32 33

namespace paddle {
namespace framework {
X
Xin Pan 已提交
34 35 36 37 38
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
39

Q
Qiao Longfei 已提交
40 41
ExecutorPrepareContext::ExecutorPrepareContext(
    const framework::ProgramDesc& prog, size_t block_id)
S
sneaxiy 已提交
42 43 44 45 46
    : prog_(prog), block_id_(block_id) {
  if (GetEagerDeletionThreshold() >= 0) {
    ref_cnts_ = GetNonPersistableReferenceCount<int>(prog_, block_id_);
  }
}
Y
Yu Yang 已提交
47

Q
Qiao Longfei 已提交
48
ExecutorPrepareContext::~ExecutorPrepareContext() {
49
  VLOG(50) << "destroy ExecutorPrepareContext";
Q
Qiao Longfei 已提交
50
}
Y
Yu Yang 已提交
51

S
sneaxiy 已提交
52 53 54 55 56 57 58 59 60 61 62 63 64 65
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) {
66
            VLOG(100) << "Erase tensor \'" << name << "\'";
S
sneaxiy 已提交
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
            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);
  }
}

B
baojun-nervana 已提交
87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
static void EnableFusedOp(ExecutorPrepareContext* ctx) {
#ifdef PADDLE_WITH_NGRAPH
  VLOG(3) << "use_ngraph=True";
  auto intervals = FusedOperator::FusedOpIntervals(&ctx->ops_);
  for (auto& interval : intervals) {
    auto* fused_op = new FusedOperator(ctx->prog_, ctx->block_id_,
                                       interval.at(0), interval.at(1));
    *interval[0] = std::unique_ptr<OperatorBase>(fused_op);
  }
  for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
    ctx->ops_.erase(it->at(0) + 1, it->at(1));
  }
#else
  LOG(WARNING)
      << "'NGRAPH' is not supported, Please re-compile with WITH_NGRAPH option";
#endif
}

D
dzhwinter 已提交
105
Executor::Executor(const platform::Place& place) : place_(place) {}
Q
qijun 已提交
106

Y
Yancey1989 已提交
107
void Executor::Close() {
W
Wu Yi 已提交
108
#ifdef PADDLE_WITH_DISTRIBUTE
W
Wu Yi 已提交
109 110
  // TODO(typhoonzero): complete message will need to use real trainer_id,
  // except 0.
Y
Yancey1989 已提交
111
  ::paddle::operators::distributed::RPCClient::GetInstance<
W
Wu Yi 已提交
112
      ::paddle::operators::distributed::GRPCClient>(0)
Y
Yancey1989 已提交
113
      ->SendComplete();
W
Wu Yi 已提交
114
#endif
Y
Yancey1989 已提交
115
}
W
Wu Yi 已提交
116

Y
Stash  
Yu Yang 已提交
117
void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
118
  if (var_type == proto::VarType::LOD_TENSOR) {
Q
QI JUN 已提交
119
    var->GetMutable<LoDTensor>();
120
  } else if (var_type == proto::VarType::SELECTED_ROWS) {
Q
QI JUN 已提交
121
    var->GetMutable<SelectedRows>();
122
  } else if (var_type == proto::VarType::FEED_MINIBATCH) {
Q
QI JUN 已提交
123
    var->GetMutable<FeedFetchList>();
124
  } else if (var_type == proto::VarType::FETCH_LIST) {
Q
QI JUN 已提交
125
    var->GetMutable<FeedFetchList>();
126
  } else if (var_type == proto::VarType::STEP_SCOPES) {
X
Xin Pan 已提交
127
    var->GetMutable<std::vector<framework::Scope*>>();
128
  } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
Y
Yu Yang 已提交
129
    var->GetMutable<LoDRankTable>();
130
  } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
Y
Yu Yang 已提交
131
    var->GetMutable<LoDTensorArray>();
132
  } else if (var_type == proto::VarType::PLACE_LIST) {
Y
Yang Yu 已提交
133
    var->GetMutable<platform::PlaceList>();
134
  } else if (var_type == proto::VarType::READER) {
F
fengjiayi 已提交
135
    var->GetMutable<ReaderHolder>();
T
typhoonzero 已提交
136 137
  } else if (var_type == proto::VarType::RAW) {
    // GetMutable will be called in operator
Q
QI JUN 已提交
138 139
  } else {
    PADDLE_THROW(
Y
Yu Yang 已提交
140
        "Variable type %d is not in "
F
fengjiayi 已提交
141
        "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
X
Xin Pan 已提交
142
        "LOD_RANK_TABLE, PLACE_LIST, READER, RAW]",
Y
Yu Yang 已提交
143
        var_type);
Q
QI JUN 已提交
144 145 146
  }
}

L
Liu Yiqun 已提交
147 148 149
void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
                               int block_id) {
  auto& global_block = pdesc.Block(block_id);
150 151 152 153 154 155 156 157 158 159 160 161 162 163

  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());
164
        InitializeVariable(ptr, var->GetType());
165 166
        VLOG(30) << "Create Variable " << var->Name()
                 << " global, which pointer is " << ptr;
167 168
      } else {
        auto* ptr = scope->Var(var->Name());
169
        InitializeVariable(ptr, var->GetType());
170 171
        VLOG(30) << "Create Variable " << var->Name()
                 << " locally, which pointer is " << ptr;
172 173 174 175 176
      }
    }
  } else {
    for (auto& var : global_block.AllVars()) {
      auto* ptr = scope->Var(var->Name());
177
      InitializeVariable(ptr, var->GetType());
178 179
      VLOG(30) << "Create variable " << var->Name() << ", which pointer is "
               << ptr;
180 181 182 183
    }
  }
}

Y
Yu Yang 已提交
184
void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
T
typhoonzero 已提交
185
                   bool create_local_scope, bool create_vars) {
X
Xin Pan 已提交
186
  platform::RecordBlock b(block_id);
187
  if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
Q
Qiao Longfei 已提交
188 189
  auto ctx = Prepare(pdesc, block_id);
  RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
Q
qijun 已提交
190 191
}

192 193 194 195 196 197 198
// 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(
199
    const BlockDesc& block,
L
Liu Yiqun 已提交
200
    const std::map<std::string, const LoDTensor*>& feed_targets,
201 202
    const std::string& feed_holder_name) {
  size_t feed_count = 0;
203
  for (auto* op : block.AllOps()) {
204 205
    if (op->Type() == kFeedOpType) {
      feed_count++;
L
Liu Yiqun 已提交
206
      // The input variable's name of feed_op should be feed_holder_name.
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
      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'");

222
    if (!feed_holder_name.empty()) {
L
Liu Yiqun 已提交
223
      // When feed operator are present, so should be feed_holder.
224 225 226 227 228 229 230
      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);
    }
231 232 233 234 235 236 237 238 239 240 241 242
  }

  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
243 244
    const BlockDesc& block,
    const std::map<std::string, LoDTensor*>& fetch_targets,
245 246
    const std::string& fetch_holder_name) {
  size_t fetch_count = 0;
247
  for (auto* op : block.AllOps()) {
248 249
    if (op->Type() == kFetchOpType) {
      fetch_count++;
L
Liu Yiqun 已提交
250
      // The output variable's name of fetch_op should be fetch_holder_name.
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265
      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'");

266
    if (!fetch_holder_name.empty()) {
L
Liu Yiqun 已提交
267
      // When fetch operator are present, so should be fetch_holder.
268 269 270 271 272 273 274
      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);
    }
275 276 277 278 279 280
  }

  return fetch_count > 0;
}

void Executor::Run(const ProgramDesc& program, Scope* scope,
281 282
                   std::map<std::string, const LoDTensor*>* feed_targets,
                   std::map<std::string, LoDTensor*>* fetch_targets,
W
Wu Yi 已提交
283 284
                   bool create_local_scope, bool create_vars,
                   const std::string& feed_holder_name,
285
                   const std::string& fetch_holder_name) {
X
Xin Pan 已提交
286
  platform::RecordBlock b(kProgramId);
287
  if (FLAGS_use_mkldnn) EnableMKLDNN(program);
288
  bool has_feed_ops =
289
      has_feed_operators(program.Block(0), *feed_targets, feed_holder_name);
290
  bool has_fetch_ops =
291
      has_fetch_operators(program.Block(0), *fetch_targets, fetch_holder_name);
292 293

  ProgramDesc* copy_program = const_cast<ProgramDesc*>(&program);
S
sneaxiy 已提交
294
  std::unique_ptr<ProgramDesc> unique_ptr_of_copy_program;
295
  if (!has_feed_ops || !has_fetch_ops) {
S
sneaxiy 已提交
296 297
    unique_ptr_of_copy_program.reset(new ProgramDesc(program));
    copy_program = unique_ptr_of_copy_program.get();
298
  }
299 300
  auto* global_block = copy_program->MutableBlock(0);

301
  if (!has_feed_ops) {
302 303
    // create feed_holder variable
    auto* feed_holder = global_block->Var(feed_holder_name);
304
    feed_holder->SetType(proto::VarType::FEED_MINIBATCH);
305 306 307
    feed_holder->SetPersistable(true);

    int i = 0;
308
    for (auto& feed_target : (*feed_targets)) {
309
      std::string var_name = feed_target.first;
310
      VLOG(30) << "feed target's name: " << var_name;
311 312 313 314 315 316 317 318 319 320 321 322 323

      // 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++;
    }
  }

324
  if (!has_fetch_ops) {
325 326
    // create fetch_holder variable
    auto* fetch_holder = global_block->Var(fetch_holder_name);
327
    fetch_holder->SetType(proto::VarType::FETCH_LIST);
328 329 330
    fetch_holder->SetPersistable(true);

    int i = 0;
331
    for (auto& fetch_target : (*fetch_targets)) {
332
      std::string var_name = fetch_target.first;
333
      VLOG(30) << "fetch target's name: " << var_name;
334 335 336 337 338 339 340 341 342 343 344 345 346

      // 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++;
    }
  }

347
  auto ctx = Prepare(*copy_program, 0);
W
Wu Yi 已提交
348 349 350
  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets,
                     create_local_scope, create_vars, feed_holder_name,
                     fetch_holder_name);
351 352
}

Q
Qiao Longfei 已提交
353 354
std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
    const ProgramDesc& program, int block_id) {
Q
Qiyang Min 已提交
355 356
  std::unique_ptr<ExecutorPrepareContext> ctx(
      new ExecutorPrepareContext(program, block_id));
Y
Yu Yang 已提交
357 358 359 360 361
  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));
  }
B
baojun-nervana 已提交
362
  if (FLAGS_use_ngraph) EnableFusedOp(ctx.get());
Q
Qiyang Min 已提交
363
  return ctx;
Y
Yu Yang 已提交
364 365
}

T
refine  
typhoonzero 已提交
366
std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
T
typhoonzero 已提交
367 368 369 370 371 372 373 374 375 376 377 378 379 380
    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
381
void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
Q
qiaolongfei 已提交
382 383
                                  bool create_local_scope, bool create_vars,
                                  bool keep_kids) {
384
  PADDLE_ENFORCE_NOT_NULL(scope);
Y
Yu Yang 已提交
385 386 387 388
  Scope* local_scope = scope;
  if (create_vars) {
    if (create_local_scope) {
      local_scope = &scope->NewScope();
389 390
    }
    CreateVariables(ctx->prog_, local_scope, ctx->block_id_);
L
Liu Yiqun 已提交
391
  }
Y
Yu Yang 已提交
392

S
sneaxiy 已提交
393 394
  int64_t max_memory_size = GetEagerDeletionThreshold();
  std::unique_ptr<GarbageCollector<Tensor>> gc;
395 396
  // WhileOp would set keep_kids to true,
  // because WhileGradOp needs the scopes created in WhileOp.
S
sneaxiy 已提交
397 398 399 400
  // 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
401
    ctx->ResetReferenceCount();
S
sneaxiy 已提交
402 403 404 405 406 407 408 409 410 411 412 413 414
#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
415
  for (auto& op : ctx->ops_) {
416
    op->Run(*local_scope, place_);
S
sneaxiy 已提交
417 418

    if (gc != nullptr) {
S
sneaxiy 已提交
419 420
      DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
                          &(ctx->cur_ref_cnts_));
S
sneaxiy 已提交
421
    }
Y
Yu Yang 已提交
422
  }
S
sneaxiy 已提交
423

S
sneaxiy 已提交
424
  if (gc != nullptr) {
S
sneaxiy 已提交
425
    gc->Wait();
S
sneaxiy 已提交
426
  } else {
S
sneaxiy 已提交
427
    platform::DeviceContextPool::Instance().Get(place_)->Wait();
S
sneaxiy 已提交
428
  }
S
sneaxiy 已提交
429

Q
qiaolongfei 已提交
430
  if (local_scope != scope) {
Y
Yu Yang 已提交
431
    scope->DeleteScope(local_scope);
432
  } else {
Q
qiaolongfei 已提交
433 434 435 436 437
    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
438 439
      // 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
440 441
      scope->DropKids();
    }
Y
Yu Yang 已提交
442 443 444
  }
}

445 446
void Executor::RunPreparedContext(
    ExecutorPrepareContext* ctx, Scope* scope,
447
    std::map<std::string, const LoDTensor*>* feed_targets,
W
Wu Yi 已提交
448 449 450
    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) {
451 452
  auto& global_block = ctx->prog_.Block(ctx->block_id_);

453
  PADDLE_ENFORCE(
454
      has_feed_operators(global_block, *feed_targets, feed_holder_name),
455 456
      "Program in ExecutorPrepareContext should has feed_ops.");
  PADDLE_ENFORCE(
457
      has_fetch_operators(global_block, *fetch_targets, fetch_holder_name),
458 459
      "Program in the prepared context should has fetch_ops.");

460 461 462 463 464
  // 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"));
465 466
      SetFeedVariable(scope, *(*feed_targets)[feed_target_name],
                      feed_holder_name, idx);
467 468 469
    }
  }

W
Wu Yi 已提交
470
  RunPreparedContext(ctx, scope, create_local_scope, create_vars);
471 472 473 474 475 476

  // 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"));
477
      *(*fetch_targets)[fetch_target_name] =
478 479 480 481 482
          GetFetchVariable(*scope, fetch_holder_name, idx);
    }
  }
}

483 484
void Executor::EnableMKLDNN(const ProgramDesc& program) {
#ifdef PADDLE_WITH_MKLDNN
485
  VLOG(30) << "use_mkldnn=True";
486 487 488 489 490 491 492 493
  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);
      }
    }
  }
494 495 496
#else
  LOG(WARNING)
      << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
497 498
#endif
}
Q
qijun 已提交
499 500
}  // namespace framework
}  // namespace paddle