operator.cc 32.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Q
Qiao Longfei 已提交
2 3 4 5 6 7 8 9 10 11 12 13

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. */
D
dzhwinter 已提交
14

15 16
#include <gflags/gflags.h>
#include <glog/logging.h>
17

18
#include <algorithm>
19

Y
Yi Wang 已提交
20 21
#include "paddle/fluid/framework/data_transform.h"
#include "paddle/fluid/framework/executor.h"
22
#include "paddle/fluid/framework/lod_tensor.h"
23
#include "paddle/fluid/framework/operator.h"
Y
Yi Wang 已提交
24
#include "paddle/fluid/framework/shape_inference.h"
25
#include "paddle/fluid/framework/transfer_scope_cache.h"
Y
Yi Wang 已提交
26
#include "paddle/fluid/framework/var_type.h"
27
#include "paddle/fluid/platform/profiler.h"
Q
Qiao Longfei 已提交
28

D
dzhwinter 已提交
29
DECLARE_bool(benchmark);
C
chengduoZH 已提交
30 31 32
DEFINE_bool(check_nan_inf, false,
            "Checking whether operator produce NAN/INF or not. It will be "
            "extremely slow so please use this flag wisely.");
D
dzhwinter 已提交
33

Q
Qiao Longfei 已提交
34 35 36
namespace paddle {
namespace framework {

37 38 39 40 41 42
std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority = {
    std::make_tuple(platform::CUDAPlace(0), LibraryType::kCUDNN),
    std::make_tuple(platform::CUDAPlace(0), LibraryType::kPlain),
    std::make_tuple(platform::CPUPlace(), LibraryType::kMKLDNN),
    std::make_tuple(platform::CPUPlace(), LibraryType::kPlain),
};
D
dzhwinter 已提交
43

Q
qiaolongfei 已提交
44 45
proto::VarType::Type GetDataTypeOfVar(const Variable* var) {
  if (var->IsType<framework::LoDTensor>()) {
Y
Yu Yang 已提交
46
    return var->Get<framework::LoDTensor>().type();
Q
qiaolongfei 已提交
47
  } else if (var->IsType<framework::SelectedRows>()) {
Y
Yu Yang 已提交
48
    return var->Get<framework::SelectedRows>().value().type();
Q
qiaolongfei 已提交
49 50 51 52 53
  } else {
    PADDLE_THROW("Var should be LoDTensor or SelectedRows");
  }
}

54 55
static DDim GetDims(const Scope& scope, const std::string& name,
                    bool get_actual_dim = false) {
56
  Variable* var = scope.FindVar(name);
Q
qiaolongfei 已提交
57 58
  if (var == nullptr) {
    return DDim({-1});
Q
Qiao Longfei 已提交
59 60
  }

M
minqiyang 已提交
61 62
  if (var->IsType<LoDTensor>()) {
    const LoDTensor& tensor = var->Get<LoDTensor>();
M
minqiyang 已提交
63
    if (UNLIKELY(!tensor.IsInitialized())) {
64
      return DDim({-1});
65
    }
M
minqiyang 已提交
66 67 68 69 70 71 72
    return tensor.dims();
  } else if (var->IsType<SelectedRows>()) {
    if (get_actual_dim) {
      return var->Get<SelectedRows>().value().dims();
    } else {
      return var->Get<SelectedRows>().GetCompleteDims();
    }
73 74 75 76 77
  } else {
    return DDim({-1});
  }
}

Q
Qiao Longfei 已提交
78 79 80 81 82 83
static bool VarInited(const Scope& scope, const std::string& name) {
  Variable* var = scope.FindVar(name);
  if (var == nullptr) return false;
  return var->IsInitialized();
}

D
dzhwinter 已提交
84 85 86 87 88
static std::string GetDtype(const Scope& scope, const std::string& name) {
  Variable* var = scope.FindVar(name);
  if (var == nullptr) {
    return "";
  }
89

M
minqiyang 已提交
90 91 92
  if (var->IsType<LoDTensor>()) {
    const LoDTensor& tensor = var->Get<LoDTensor>();
    if (UNLIKELY(!tensor.IsInitialized())) {
93 94
      return "";
    }
Y
Yu Yang 已提交
95
    return DataTypeToString(tensor.type());
M
minqiyang 已提交
96
  } else if (var->IsType<SelectedRows>()) {
Q
Qiao Longfei 已提交
97 98 99 100
    auto tensor = var->Get<SelectedRows>().value();
    if (UNLIKELY(!tensor.IsInitialized())) {
      return "uninited";
    } else {
Y
Yu Yang 已提交
101
      return DataTypeToString(tensor.type());
Q
Qiao Longfei 已提交
102
    }
D
dzhwinter 已提交
103 104 105 106 107
  } else {
    return "";
  }
}

108 109 110 111 112 113
static int GetRowSize(const Scope& scope, const std::string& name) {
  Variable* var = scope.FindVar(name);
  if (var == nullptr) {
    return -1;
  }

M
minqiyang 已提交
114 115
  if (var->IsType<SelectedRows>()) {
    return var->Get<SelectedRows>().rows().size();
116 117 118 119 120
  }

  return -1;
}

Q
Qiao Longfei 已提交
121 122 123 124 125 126 127 128
static LoD GetLoD(const Scope& scope, const std::string& name) {
  Variable* var = scope.FindVar(name);
  auto default_lod = LoD({{}});

  if (var == nullptr) {
    return default_lod;
  }

M
minqiyang 已提交
129 130 131
  if (var->IsType<LoDTensor>()) {
    const LoDTensor& tensor = var->Get<LoDTensor>();
    if (UNLIKELY(!tensor.IsInitialized())) {
132 133
      return default_lod;
    }
M
minqiyang 已提交
134
    return tensor.lod();
Q
Qiao Longfei 已提交
135 136 137 138 139
  } else {
    return default_lod;
  }
}

X
Xin Pan 已提交
140 141 142 143 144 145
RuntimeContext::RuntimeContext(const VariableNameMap& innames,
                               const VariableNameMap& outnames,
                               const Scope& scope) {
  for (auto& var_name_item : innames) {
    std::vector<Variable*>& input_vars = inputs[var_name_item.first];
    for (auto& var_name : var_name_item.second) {
X
Xin Pan 已提交
146
      LOG(ERROR) << "first in " << var_name_item.first << ":" << var_name;
X
Xin Pan 已提交
147 148 149 150 151 152
      input_vars.push_back(scope.FindVar(var_name));
    }
  }
  for (auto& var_name_item : outnames) {
    std::vector<Variable*>& output_vars = outputs[var_name_item.first];
    for (auto& var_name : var_name_item.second) {
X
Xin Pan 已提交
153
      LOG(ERROR) << "first out " << var_name_item.first << ":" << var_name;
X
Xin Pan 已提交
154 155 156 157 158
      output_vars.push_back(scope.FindVar(var_name));
    }
  }
}

159
void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
M
minqiyang 已提交
160
  VLOG(4) << place << " " << DebugStringEx(&scope);
161
  if (platform::is_gpu_place(place)) {
162
#ifndef PADDLE_WITH_CUDA
163
    PADDLE_THROW("Cannot run operator on place %s", place);
164
#else
165 166
    auto dev_id = boost::get<platform::CUDAPlace>(place).device;
    platform::SetDeviceId(dev_id);
167 168
#endif
  }
169

P
peizhilin 已提交
170 171 172
  // The profile has a process-wide mutex, results in serious performance issue
  // in concurrency scenerio. Here use an `if` to fix this issue.
  // Please not remove the `if`, ask @Superjomn if there are any concern.
173 174 175 176
  if (platform::IsProfileEnabled()) {
    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
    platform::RecordEvent record_event(Type(), pool.Get(place));
    RunImpl(scope, place);
P
peizhilin 已提交
177
  } else {
178 179
    RunImpl(scope, place);
  }
M
minqiyang 已提交
180
  VLOG(3) << place << " " << DebugStringEx(&scope);
181 182
}

183 184 185 186 187 188 189 190
bool OperatorBase::HasInputs(const std::string& name) const {
  if (inputs_.find(name) != inputs_.end()) {
    return true;
  } else {
    return false;
  }
}

191
std::string OperatorBase::Input(const std::string& name) const {
Y
Yu Yang 已提交
192
  auto& ins = Inputs(name);
Y
Yu Yang 已提交
193
  PADDLE_ENFORCE_LE(ins.size(), 1UL,
194 195
                    "Operator %s's input %s should contain only one variable.",
                    type_, name);
Y
Yu Yang 已提交
196
  return ins.empty() ? kEmptyVarName : ins[0];
Y
Yan Chunwei 已提交
197 198
}

Y
Yu Yang 已提交
199 200
const std::vector<std::string>& OperatorBase::Inputs(
    const std::string& name) const {
Y
Yu Yang 已提交
201
  auto it = inputs_.find(name);
202 203
  PADDLE_ENFORCE(it != inputs_.end(), "Operator %s does not have the input %s.",
                 type_, name);
Y
Yu Yang 已提交
204
  return it->second;
Y
Yan Chunwei 已提交
205 206
}

207
bool OperatorBase::HasOutputs(const std::string& name) const {
208
  if (outputs_.find(name) != outputs_.end()) {
209 210 211 212 213 214
    return true;
  } else {
    return false;
  }
}

215
std::string OperatorBase::Output(const std::string& name) const {
Y
Yu Yang 已提交
216
  auto& outs = Outputs(name);
Y
Yu Yang 已提交
217
  PADDLE_ENFORCE_LE(outs.size(), 1UL,
218 219
                    "Operator %s's output %s should contain only one variable.",
                    type_, name);
Y
Yu Yang 已提交
220
  return outs.empty() ? kEmptyVarName : outs[0];
Y
Yan Chunwei 已提交
221 222
}

Y
Yu Yang 已提交
223 224
const std::vector<std::string>& OperatorBase::Outputs(
    const std::string& name) const {
Y
Yu Yang 已提交
225
  auto it = outputs_.find(name);
226 227
  PADDLE_ENFORCE(it != outputs_.end(),
                 "Operator %s does not have an output called %s.", type_, name);
Y
Yu Yang 已提交
228
  return it->second;
Y
Yan Chunwei 已提交
229 230
}

231
std::string OperatorBase::DebugStringEx(const Scope* scope) const {
Q
Qiao Longfei 已提交
232
  std::stringstream ss;
Y
Yu Yang 已提交
233
  ss << "Op(" << type_ << "), inputs:{";
Y
Yu Yang 已提交
234 235
  for (auto it = inputs_.begin(); it != inputs_.end();) {
    auto& input = *it;
Y
Yu Yang 已提交
236 237
    ss << input.first << "[";
    for (size_t i = 0; i < input.second.size(); ++i) {
Q
Qiao Longfei 已提交
238 239
      auto var_name = input.second[i];
      ss << var_name;
240
      if (scope) {
Q
Qiao Longfei 已提交
241 242 243 244 245 246 247 248 249 250 251
        if (!VarInited(*scope, var_name)) {
          ss << "[uninited]";
        } else {
          int row_size = GetRowSize(*scope, var_name);
          if (row_size >= 0) {
            ss << "[row_size=" << row_size << "]";
          }
          std::string dtype = GetDtype(*scope, var_name);
          ss << ":" << dtype;
          ss << "[" << GetDims(*scope, var_name, true) << "]";
          ss << "(" << GetLoD(*scope, var_name) << ")";
252
        }
253
      }
Y
Yu Yang 已提交
254 255 256
      if (i != input.second.size() - 1) {
        ss << ", ";
      }
257
    }
Y
Yu Yang 已提交
258
    ss << "]";
Y
Yu Yang 已提交
259 260
    ++it;
    if (it != inputs_.end()) {
261 262
      ss << ", ";
    }
Q
Qiao Longfei 已提交
263
  }
Y
Yu Yang 已提交
264
  ss << "}, outputs:{";
Y
Yu Yang 已提交
265 266
  for (auto it = outputs_.begin(); it != outputs_.end();) {
    auto& output = *it;
Y
Yu Yang 已提交
267 268
    ss << output.first << "[";
    for (size_t i = 0; i < output.second.size(); ++i) {
Q
Qiao Longfei 已提交
269 270
      auto var_name = output.second[i];
      ss << var_name;
271
      if (scope) {
Q
Qiao Longfei 已提交
272 273 274 275 276 277 278
        if (!VarInited(*scope, var_name)) {
          ss << "[uninited]";
        } else {
          int row_size = GetRowSize(*scope, output.second[i]);
          if (row_size >= 0) {
            ss << "[row_size=" << row_size << "]";
          }
C
chengduo 已提交
279 280
          std::string dtype = GetDtype(*scope, output.second[i]);
          ss << ":" << dtype;
Q
Qiao Longfei 已提交
281 282
          ss << "[" << GetDims(*scope, var_name, true) << "]";
          ss << "(" << GetLoD(*scope, var_name) << ")";
283
        }
284
      }
Y
Yu Yang 已提交
285 286 287
      if (i != output.second.size() - 1) {
        ss << ", ";
      }
288
    }
Y
Yu Yang 已提交
289
    ss << "]";
Y
Yu Yang 已提交
290 291
    ++it;
    if (it != outputs_.end()) {
292 293
      ss << ", ";
    }
Q
Qiao Longfei 已提交
294
  }
Y
Yu Yang 已提交
295
  ss << "}.";
Q
Qiao Longfei 已提交
296 297 298
  return ss.str();
}

Y
Yu Yang 已提交
299
OperatorBase::OperatorBase(const std::string& type,
Y
Yu Yang 已提交
300 301
                           const VariableNameMap& inputs,
                           const VariableNameMap& outputs,
Y
Yu Yang 已提交
302 303
                           const AttributeMap& attrs)
    : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
304 305
  GenerateTemporaryNames();
  CheckAllInputOutputSet();
Y
Yu Yang 已提交
306
}
307

Q
qijun 已提交
308 309
std::vector<std::string> OperatorBase::InputVars() const {
  std::vector<std::string> ret_val;
Y
Yu Yang 已提交
310
  for (auto& o : inputs_) {
Q
qijun 已提交
311 312 313 314 315 316
    ret_val.reserve(ret_val.size() + o.second.size());
    ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
  }
  return ret_val;
}

Y
Yu Yang 已提交
317 318 319 320 321 322 323 324 325 326
std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
  std::vector<std::string> ret_val;
  if (has_intermediate) {
    // push all outputs into ret_val
    for (auto& o : outputs_) {
      ret_val.reserve(ret_val.size() + o.second.size());
      ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
    }
    return ret_val;
  }
Y
Yu Yang 已提交
327
  auto& info = OpInfoMap::Instance().Get(Type());
Y
Yu Yang 已提交
328 329

  // get all OpProto::Var for outputs
Y
Yu Yang 已提交
330
  for (auto& o : info.Proto().outputs()) {
Y
Yu Yang 已提交
331 332 333 334 335 336 337 338 339
    // ignore all intermediate output
    if (o.intermediate()) continue;
    auto out = outputs_.find(o.name());
    if (out != outputs_.end()) {
      ret_val.reserve(ret_val.size() + out->second.size());
      ret_val.insert(ret_val.end(), out->second.begin(), out->second.end());
    }
  }
  return ret_val;
D
dongzhihong 已提交
340 341
}

342 343 344
void OperatorBase::CheckAllInputOutputSet() const {
  auto& info_map = OpInfoMap::Instance();
  auto* op_info = info_map.GetNullable(Type());
Y
Yu Yang 已提交
345
  if (op_info == nullptr || op_info->proto_ == nullptr) return;
346 347

  for (auto& in : op_info->Proto().inputs()) {
348 349 350 351
    if (!in.dispensable()) {
      PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(),
                     "Operator %s's input, %s, is not set", Type(), in.name());
    }
352 353 354
  }

  for (auto& out : op_info->Proto().outputs()) {
355 356 357 358 359
    if (!out.dispensable()) {
      PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(),
                     "Operator %s's output, %s, is not set", Type(),
                     out.name());
    }
360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375
  }
}

void OperatorBase::GenerateTemporaryNames() {
  static std::atomic<size_t> gUniqId(0UL);
  for (auto& output : outputs_) {
    for (auto& output_name : output.second) {
      if (output_name == kTempVarName) {
        output_name += type_;
        output_name += "@";
        output_name += std::to_string(gUniqId.fetch_add(1));
      }
    }
  }
}

B
baojun-nervana 已提交
376
static bool VarIsTensor(const Variable& var) {
C
chengduo 已提交
377
  return var.IsType<LoDTensor>() || var.IsType<SelectedRows>();
378 379
}

C
chengduo 已提交
380
const Tensor* GetLoDTensorOrSelectedRowsValueFromVar(const Variable& var) {
C
chengduo 已提交
381 382 383 384
  if (var.IsType<LoDTensor>()) {
    return static_cast<const Tensor*>(&(var.Get<LoDTensor>()));
  } else if (var.IsType<SelectedRows>()) {
    return &(var.Get<SelectedRows>().value());
Q
QI JUN 已提交
385
  } else {
Y
Yang Yang 已提交
386
    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
C
chengduo 已提交
387
                 var.Type().name());
Q
QI JUN 已提交
388 389 390
  }
}

C
chengduo 已提交
391
Tensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var) {
Q
QI JUN 已提交
392
  if (var->IsType<LoDTensor>()) {
393
    return var->GetMutable<LoDTensor>();
Q
QI JUN 已提交
394
  } else if (var->IsType<SelectedRows>()) {
395
    return var->GetMutable<SelectedRows>()->mutable_value();
Q
QI JUN 已提交
396
  } else {
Y
Yang Yang 已提交
397 398
    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
                 var->Type().name());
Q
QI JUN 已提交
399 400 401
  }
}

402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433
bool ExecutionContext::HasInput(const std::string& name) const {
  if (!op_.HasInputs(name)) {
    return false;
  }
  auto& ins = Inputs(name);
  size_t length = ins.size();
  if (length == 0) {
    return false;
  }
  PADDLE_ENFORCE_EQ(length, 1UL,
                    "Input %s should not have more than one inputs", name);
  auto arg = ins[0];
  auto* var = arg == kEmptyVarName ? nullptr : scope_.FindVar(arg);
  return var != nullptr;
}

bool ExecutionContext::HasOutput(const std::string& name) const {
  if (!op_.HasOutputs(name)) {
    return false;
  }
  auto& outs = Outputs(name);
  size_t length = outs.size();
  if (length == 0) {
    return false;
  }
  PADDLE_ENFORCE_EQ(length, 1UL,
                    "Output %s should not have more than one inputs", name);
  auto arg = outs[0];
  auto* var = arg == kEmptyVarName ? nullptr : scope_.FindVar(arg);
  return var != nullptr;
}

X
Xin Pan 已提交
434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468
const Variable* ExecutionContext::InputVar(const std::string& name) const {
  auto it = ctx_.inputs.find(name);
  if (it == ctx_.inputs.end()) return nullptr;

  PADDLE_ENFORCE_LE(it->second.size(), 1UL,
                    "Operator %s's input %s should contain only one variable.",
                    op_.Type(), name);
  return it->second.empty() ? nullptr : it->second[0];
}

Variable* ExecutionContext::OutputVar(const std::string& name) const {
  auto opt = op_.Output(name);
  return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt);
}

const Variable* ExecutionContext::FastInputVar(const std::string& name) const {
  auto it = ctx_.inputs.find(name);
  if (it == ctx_.inputs.end()) return nullptr;

  PADDLE_ENFORCE_LE(it->second.size(), 1UL,
                    "Operator %s's input %s should contain only one variable.",
                    op_.Type(), name);
  return it->second.empty() ? nullptr : it->second[0];
}

Variable* ExecutionContext::FastOutputVar(const std::string& name) const {
  auto it = ctx_.outputs.find(name);
  if (it == ctx_.outputs.end()) return nullptr;

  PADDLE_ENFORCE_LE(it->second.size(), 1UL,
                    "Operator %s's output %s should contain only one variable.",
                    op_.Type(), name);
  return it->second.empty() ? nullptr : it->second[0];
}

469
template <>
470
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
C
chengduo 已提交
471
  return Input<LoDTensor>(name);
472 473
}

X
Xin Pan 已提交
474 475 476 477 478 479
template <>
const Tensor* ExecutionContext::FastInput<Tensor>(
    const std::string& name) const {
  return FastInput<LoDTensor>(name);
}

480
template <>
481
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
482 483 484 485
    const std::string& name) const {
  auto names = op().Inputs(name);
  std::vector<const Tensor*> res;
  res.reserve(names.size());
486
  std::transform(names.begin(), names.end(), std::back_inserter(res),
C
chengduo 已提交
487
                 [&](const std::string& sub_name) -> const Tensor* {
488
                   auto var = scope_.FindVar(sub_name);
C
chengduo 已提交
489 490 491 492 493 494
                   if (var == nullptr) return nullptr;
                   PADDLE_ENFORCE(
                       var->IsType<LoDTensor>(),
                       "%s should be LoDTensor, but the received type is %s",
                       sub_name, var->Type().name());
                   return &(var->Get<LoDTensor>());
495
                 });
496 497 498 499
  return res;
}

template <>
500
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
C
chengduo 已提交
501
  return Output<LoDTensor>(name);
502 503
}

X
Xin Pan 已提交
504 505 506 507 508
template <>
Tensor* ExecutionContext::FastOutput<Tensor>(const std::string& name) const {
  return FastOutput<LoDTensor>(name);
}

509
template <>
510
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
511 512 513 514
    const std::string& name) const {
  auto names = op().Outputs(name);
  std::vector<Tensor*> res;
  res.reserve(names.size());
515
  std::transform(names.begin(), names.end(), std::back_inserter(res),
C
chengduo 已提交
516
                 [&](const std::string& sub_name) -> Tensor* {
517
                   auto var = scope_.FindVar(sub_name);
C
chengduo 已提交
518 519 520 521 522 523
                   if (var == nullptr) return nullptr;
                   PADDLE_ENFORCE(
                       var->IsType<LoDTensor>(),
                       "%s should be LoDTensor, but the received type is %s",
                       sub_name, var->Type().name());
                   return var->GetMutable<LoDTensor>();
524
                 });
525 526 527
  return res;
}

Y
Yu Yang 已提交
528 529 530 531 532 533 534 535 536 537 538 539 540 541 542
bool OpSupportGPU(const std::string& op_type) {
  auto& all_kernels = OperatorWithKernel::AllOpKernels();
  auto it = all_kernels.find(op_type);
  if (it == all_kernels.end()) {
    // All control operator must support GPU
    return true;
  }
  for (auto& kern_pair : it->second) {
    if (platform::is_gpu_place(kern_pair.first.place_)) {
      return true;
    }
  }
  return false;
}

543 544
class RuntimeInferShapeContext : public InferShapeContext {
 public:
X
Xin Pan 已提交
545 546 547
  RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope,
                           const RuntimeContext& ctx)
      : op_(op), scope_(scope), ctx_(ctx) {}
548 549

  bool HasInput(const std::string& name) const override {
550
    // has only one input
X
Xin Pan 已提交
551
    const auto& ins = ctx_.inputs;
552 553
    auto it = ins.find(name);
    if (it == ins.end()) {
554 555
      return false;
    }
556
    const auto& in = it->second;
X
Xin Pan 已提交
557
    if (in.size() == 0) return false;
T
tensor-tang 已提交
558
    PADDLE_ENFORCE_EQ(in.size(), 1UL,
F
fengjiayi 已提交
559
                      "Input %s should not have more than one inputs", name);
X
Xin Pan 已提交
560
    return in[0] != nullptr;
561 562 563
  }

  bool HasOutput(const std::string& name) const override {
564 565 566 567
    // has only one output
    const auto& outs = op_.Outputs();
    auto it = outs.find(name);
    if (it == outs.end()) {
568 569
      return false;
    }
570
    const auto& out = it->second;
T
tensor-tang 已提交
571
    if (out.size() == 0 || out[0] == kEmptyVarName) {
572 573
      return false;
    }
T
tensor-tang 已提交
574 575
    PADDLE_ENFORCE_EQ(out.size(), 1UL,
                      "Output %s should not have more than one outputs", name);
576
    return scope_.FindVar(out[0]) != nullptr;
577 578 579
  }

  bool HasInputs(const std::string& name) const override {
580 581 582
    if (!op_.HasInputs(name)) {
      return false;
    }
583 584 585 586 587 588 589 590 591 592 593 594 595
    auto inputs = op_.Inputs(name);
    if (inputs.empty()) {
      return false;
    }
    for (auto& input : inputs) {
      if (scope_.FindVar(input) == nullptr) {
        return false;
      }
    }
    return true;
  }

  bool HasOutputs(const std::string& name) const override {
596 597 598
    if (!op_.HasOutputs(name)) {
      return false;
    }
599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622
    auto outputs = op_.Outputs(name);
    if (outputs.empty()) {
      return false;
    }
    for (auto& output : outputs) {
      if (scope_.FindVar(output) == nullptr) {
        return false;
      }
    }
    return true;
  }

  AttrReader Attrs() const override { return AttrReader(op_.Attrs()); }

  const std::vector<std::string>& Inputs(
      const std::string& name) const override {
    return op_.Inputs(name);
  }

  const std::vector<std::string>& Outputs(
      const std::string& name) const override {
    return op_.Outputs(name);
  }

623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652
  void ShareDim(const std::string& in, const std::string& out, size_t i = 0,
                size_t j = 0) override {
    PADDLE_ENFORCE_LT(i, Inputs(in).size());
    PADDLE_ENFORCE_LT(j, Outputs(out).size());
    const std::string& input_n = Inputs(in)[i];
    const std::string& output_n = Outputs(out)[j];

    Variable* in_var = scope_.FindVar(input_n);
    Variable* out_var = scope_.FindVar(output_n);
    PADDLE_ENFORCE(in_var->Type() == out_var->Type(),
                   "The type of %s and %s is not the same.", output_n,
                   GetDim(input_n));

    if (in_var->IsType<framework::SelectedRows>()) {
      auto& in_sele_rows = in_var->Get<framework::SelectedRows>();
      auto out_sele_rows = out_var->GetMutable<framework::SelectedRows>();
      out_sele_rows->mutable_value()->Resize(in_sele_rows.value().dims());
      out_sele_rows->set_rows(in_sele_rows.rows());
      out_sele_rows->set_height(in_sele_rows.height());
    } else if (in_var->IsType<framework::LoDTensor>()) {
      auto& in_lod_tensor = in_var->Get<framework::LoDTensor>();
      auto* out_lod_tensor = out_var->GetMutable<framework::LoDTensor>();
      out_lod_tensor->Resize(in_lod_tensor.dims());
    } else {
      PADDLE_THROW(
          "Currently, the input type of ShareDim only can be LoDTensor "
          "or SelectedRows.");
    }
  }

Q
Qiao Longfei 已提交
653 654
  void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
                size_t j = 0) const override {
X
Xin Pan 已提交
655 656 657 658 659
    const std::vector<std::string>& inputs = Inputs(in);
    const std::vector<std::string>& outputs = Outputs(out);
    PADDLE_ENFORCE_LT(i, inputs.size());
    PADDLE_ENFORCE_LT(j, outputs.size());
    Variable* in_var = scope_.FindVar(inputs.at(i));
Q
Qiao Longfei 已提交
660
    if (!in_var->IsType<LoDTensor>()) return;
X
Xin Pan 已提交
661
    Variable* out_var = scope_.FindVar(outputs.at(j));
Q
Qiao Longfei 已提交
662 663 664 665 666
    PADDLE_ENFORCE(out_var->IsType<LoDTensor>(),
                   "The %d-th output of Output(%s) must be LoDTensor.", j, out);
    auto in_tensor = in_var->Get<LoDTensor>();
    auto* out_tensor = out_var->GetMutable<LoDTensor>();
    out_tensor->set_lod(in_tensor.lod());
D
dzhwinter 已提交
667

M
mozga-intel 已提交
668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686
// TODO(dzhwinter) : reuse ShareLoD in most operators.
// Need to call ShareLayout explicitly in sequence related ops.
// Shall we have a better method to shared info between in/out Tensor?
#ifdef PADDLE_WITH_MKLDNN
    // Fix me: ugly workaround below
    // Correct solution:
    //    set_layout() should NOT be called here (i.e. ShareLoD). Instead,
    //    layout of output tensor should be set "manually" in Compute()
    //    of each OPKernel. The reason layout should NOT be shared between
    //    input and output "automatically" (now by InferShape()->ShareLoD())
    //    is that layout transform may occur after InferShape().
    // Workaround:
    //    Skip set_layout() when input layout is kMKLDNN
    //    This is to avoid kMKLDNN is populated wrongly into a non-MKLDNN
    //    OPKernel. In all MKLDNN OPkernel, set_layout(kMKLDNN) should be called
    //    in Compute()
    if (in_tensor.layout() != DataLayout::kMKLDNN)
#endif
      out_tensor->set_layout(in_tensor.layout());
D
dzhwinter 已提交
687 688
  }

C
chengduo 已提交
689 690 691 692 693
  void DecreaseLoDLevel(const std::string& in, const std::string& out,
                        size_t i = 0, size_t j = 0) const override {
    PADDLE_THROW("DecreaseLoDLevel is only used in compile time.");
  }

694 695 696
  bool IsRuntime() const override { return true; }

 protected:
697 698
  DDim GetDim(const std::string& name) const override {
    Variable* var = scope_.FindVar(name);
F
fengjiayi 已提交
699
    PADDLE_ENFORCE_NOT_NULL(var);
700 701 702 703 704
    if (var->IsType<LoDTensor>()) {
      return var->Get<LoDTensor>().dims();
    } else if (var->IsType<SelectedRows>()) {
      return var->Get<SelectedRows>().GetCompleteDims();
    } else {
F
fengjiayi 已提交
705 706 707 708 709 710 711
      PADDLE_THROW(
          "Only LoDTensor/SelectedRows support 'GetDim', but Variable %s's "
          "type_id is %s.",
          name, var->Type().name());
    }
  }

F
fengjiayi 已提交
712
  std::vector<DDim> GetRepeatedDims(const std::string& name) const override {
Y
Yu Yang 已提交
713
    PADDLE_THROW("Only compile time support this method");
714 715 716 717 718 719 720 721 722
  }

  void SetDim(const std::string& name, const DDim& dim) override {
    Variable* var = scope_.FindVar(name);
    if (var->IsType<LoDTensor>()) {
      var->GetMutable<LoDTensor>()->Resize(dim);
    } else if (var->IsType<SelectedRows>()) {
      var->GetMutable<SelectedRows>()->set_height(dim[0]);
    } else {
Y
Yang Yang 已提交
723 724
      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
                   name, var->Type().name());
725 726 727
    }
  }

F
fengjiayi 已提交
728 729
  void SetRepeatedDims(const std::string& name,
                       const std::vector<DDim>& dims) override {
Y
Yu Yang 已提交
730
    PADDLE_THROW("Only compile time support this method");
F
fengjiayi 已提交
731 732
  }

733
  proto::VarType::Type GetVarType(const std::string& name) const override {
734 735 736 737
    auto* var = scope_.FindVar(name);
    return ToVarType(var->Type());
  }

F
fengjiayi 已提交
738 739 740 741
  InferShapeVarPtr GetVarPtr(const std::string& name) override {
    return scope_.FindVar(name);
  }

742
 private:
743 744
  const OperatorBase& op_;
  const Scope& scope_;
X
Xin Pan 已提交
745
  const RuntimeContext& ctx_;
746 747
};

C
chengduoZH 已提交
748 749 750 751 752
static void CheckTensorNANOrInf(const std::string& name,
                                const framework::Tensor& tensor) {
  if (tensor.memory_size() == 0) {
    return;
  }
Y
Yu Yang 已提交
753 754
  if (tensor.type() != proto::VarType::FP32 &&
      tensor.type() != proto::VarType::FP64) {
C
chengduoZH 已提交
755 756 757 758 759 760 761 762
    return;
  }
  PADDLE_ENFORCE(!framework::TensorContainsInf(tensor),
                 "Tensor %s contains Inf", name);
  PADDLE_ENFORCE(!framework::TensorContainsNAN(tensor),
                 "Tensor %s contains NAN", name);
}

B
baojun-nervana 已提交
763
void OperatorWithKernel::RuntimeInferShape(const Scope& scope,
X
Xin Pan 已提交
764 765 766
                                           const platform::Place& place,
                                           const RuntimeContext& ctx) const {
  RuntimeInferShapeContext infer_shape_ctx(*this, scope, ctx);
B
baojun-nervana 已提交
767 768 769
  this->InferShape(&infer_shape_ctx);
}

770 771
void OperatorWithKernel::RunImpl(const Scope& scope,
                                 const platform::Place& place) const {
X
Xin Pan 已提交
772
  RuntimeContext ctx(Inputs(), Outputs(), scope);
Y
Yu Yang 已提交
773
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
774
  auto* dev_ctx = pool.Get(place);
775

776 777 778 779
  // check if op[type] has kernel registered.
  auto& all_op_kernels = AllOpKernels();
  auto kernels_iter = all_op_kernels.find(type_);
  if (kernels_iter == all_op_kernels.end()) {
Y
Yu Yang 已提交
780 781
    PADDLE_THROW(
        "There are no kernels which are registered in the %s operator.", type_);
782 783
  }

Q
qiaolongfei 已提交
784 785
  OpKernelMap& kernels = kernels_iter->second;

X
Xin Pan 已提交
786 787
  auto expected_kernel_key = this->GetExpectedKernelType(
      ExecutionContext(*this, scope, *dev_ctx, ctx));
M
minqiyang 已提交
788
  VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
Q
qiaolongfei 已提交
789

790
  auto kernel_iter = kernels.find(expected_kernel_key);
791
#ifdef PADDLE_WITH_MKLDNN
P
Paweł Żelazko 已提交
792
  // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
793 794
  if (kernel_iter == kernels.end() &&
      expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
M
minqiyang 已提交
795
    VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
796 797 798 799 800
    expected_kernel_key.library_type_ = LibraryType::kPlain;
    expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
    kernel_iter = kernels.find(expected_kernel_key);
  }
#endif
801 802 803 804 805
  if (kernel_iter == kernels.end()) {
    PADDLE_THROW("op %s does not have kernel for %s", type_,
                 KernelTypeToString(expected_kernel_key));
  }

Y
yuyang18 已提交
806 807 808
  // do data transformScope &transfer_scope;
  std::vector<std::string> transfered_inplace_vars;
  auto* transfer_scope =
X
Xin Pan 已提交
809
      PrepareData(scope, expected_kernel_key, &transfered_inplace_vars, &ctx);
810

Y
yuyang18 已提交
811 812 813 814 815 816
  // exec scope is the scope that kernel actually executed on.
  const Scope& exec_scope =
      (transfer_scope == nullptr ? scope : *transfer_scope);

  if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) {
    dev_ctx = pool.Get(expected_kernel_key.place_);
817
  }
Q
QI JUN 已提交
818

X
Xin Pan 已提交
819
  RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, ctx);
X
Xin Pan 已提交
820
  this->InferShape(&infer_shape_ctx);
X
Xin Pan 已提交
821
  kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx, ctx));
D
dzhwinter 已提交
822

Y
yuyang18 已提交
823 824 825
  if (!transfered_inplace_vars.empty()) {
    // there is inplace variable has been transfered.
    TransferInplaceVarsBack(scope, transfered_inplace_vars, *transfer_scope);
826 827
  }

D
dzhwinter 已提交
828
  /*For profiling/benchmark only*/
D
dzhwinter 已提交
829
  if (FLAGS_benchmark) {
Y
yuyang18 已提交
830
    dev_ctx->Wait();
D
dzhwinter 已提交
831
  }
C
chengduoZH 已提交
832 833 834

  if (FLAGS_check_nan_inf) {
    for (auto& vname : OutputVars(true)) {
Y
yuyang18 已提交
835
      auto* var = exec_scope.FindVar(vname);
C
chengduoZH 已提交
836 837 838
      if (var == nullptr) continue;
      if (var->IsType<framework::LoDTensor>()) {
        CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
839 840
      } else if (var->IsType<framework::SelectedRows>()) {
        CheckTensorNANOrInf(vname, var->Get<framework::SelectedRows>().value());
C
chengduoZH 已提交
841 842 843
      }
    }
  }
Q
Qiao Longfei 已提交
844
}
X
Xin Pan 已提交
845

Y
yuyang18 已提交
846 847 848 849
void OperatorWithKernel::TransferInplaceVarsBack(
    const Scope& scope, const std::vector<std::string>& inplace_vars,
    const Scope& transfer_scope) const {
  for (auto& var_name : inplace_vars) {
M
minqiyang 已提交
850
    VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
C
chengduo 已提交
851 852
    auto* original_tensor =
        GetMutableLoDTensorOrSelectedRowsValueFromVar(scope.FindVar(var_name));
C
chengduo 已提交
853 854 855
    auto* var = transfer_scope.FindVar(var_name);
    PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr",
                   var_name);
C
chengduo 已提交
856
    auto* transformed_tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var);
Y
yuyang18 已提交
857 858 859 860
    original_tensor->ShareDataWith(*transformed_tensor);
  }
}

X
Xin Pan 已提交
861
Scope* OperatorWithKernel::PrepareData(
Y
yuyang18 已提交
862
    const Scope& scope, const OpKernelType& expected_kernel_key,
X
Xin Pan 已提交
863 864
    std::vector<std::string>* transfered_inplace_vars,
    RuntimeContext* ctx) const {
Y
yuyang18 已提交
865 866
  Scope* new_scope = nullptr;
  for (auto& var_name_item : Inputs()) {
X
Xin Pan 已提交
867 868 869 870
    std::vector<Variable*>& input_vars = ctx->inputs[var_name_item.first];

    for (size_t i = 0; i < var_name_item.second.size(); ++i) {
      auto& var_name = var_name_item.second[i];
Y
yuyang18 已提交
871
      auto* var = scope.FindVar(var_name);
X
Xin Pan 已提交
872
      input_vars[i] = var;
X
Xin Pan 已提交
873
      LOG(ERROR) << "second in " << var_name_item.first << ":" << var_name;
X
Xin Pan 已提交
874

Y
yuyang18 已提交
875
      // Only tensor can be tranfer to another device.
C
chengduo 已提交
876
      if (var == nullptr || !VarIsTensor(*var)) {
Y
yuyang18 已提交
877 878 879
        continue;
      }

C
chengduo 已提交
880
      auto* tensor_in = GetLoDTensorOrSelectedRowsValueFromVar(*var);
Y
yuyang18 已提交
881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897
      if (!tensor_in->IsInitialized()) {
        continue;
      }

      auto kernel_type_for_var = GetKernelTypeForVar(
          var_name_item.first, *tensor_in, expected_kernel_key);

      if (!NeedTransform(kernel_type_for_var, expected_kernel_key)) {
        continue;
      }

      auto out_var_names = OutputVars(true);
      if (std::find(out_var_names.begin(), out_var_names.end(), var_name) !=
          out_var_names.end()) {
        transfered_inplace_vars->emplace_back(var_name);
      }

M
minqiyang 已提交
898 899
      VLOG(3) << "Transform Variable " << var_name << " from "
              << kernel_type_for_var << " to " << expected_kernel_key;
Y
yuyang18 已提交
900

901 902 903 904 905 906 907 908 909 910 911 912 913 914 915
      // In the inference scenerio, the scopes will be reused across the
      // batches, so the `new_scope` here will result in GPU memroy explosion
      // over the  running of operators.
      // We use a thread_local cache to fix that issue, the key in the cache is
      // the combination of the `scope` argument, from_kernel_type,
      // target_kernel_type.
      // Have a discussion with @Superjomn or the inference developers if some
      // changes on this logic for this macro might not tested on the other
      // scenerios.
      // If this op is not called by an Executor or ParallelExecutor, it should
      // called by a NaiveExecutor, the NaiveExecutor will cache the scopes and
      // variables, that behavior a lot different.
      if (!run_by_executor_) {
        new_scope = TryCreateTransferScope(kernel_type_for_var,
                                           expected_kernel_key, &scope);
916
      }
917
      if (!new_scope) {
Y
yuyang18 已提交
918 919 920 921
        new_scope = &scope.NewScope();
      }

      auto* trans_var = new_scope->Var(var_name);
X
fix  
Xin Pan 已提交
922
      input_vars[i] = trans_var;
923

Y
yuyang18 已提交
924
      Tensor out;
Y
yuyang18 已提交
925
      TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in, &out);
Y
yuyang18 已提交
926 927 928
      SetTensorToVariable(*var, out, trans_var);
    }
  }
X
Xin Pan 已提交
929 930 931 932 933
  for (auto& var_name_item : Outputs()) {
    std::vector<Variable*>& output_vars = ctx->outputs[var_name_item.first];
    for (size_t i = 0; i < var_name_item.second.size(); ++i) {
      auto& var_name = var_name_item.second[i];
      output_vars[i] = scope.FindVar(var_name);
X
Xin Pan 已提交
934
      LOG(ERROR) << "second out " << var_name_item.first << ":" << var_name;
X
Xin Pan 已提交
935 936
    }
  }
Y
yuyang18 已提交
937 938 939

  return new_scope;
}
Q
Qiao Longfei 已提交
940

941
proto::VarType::Type OperatorWithKernel::IndicateDataType(
Y
Yu Yang 已提交
942 943 944
    const ExecutionContext& ctx) const {
  auto& scope = ctx.scope();
  int data_type = -1;
945
  std::string last_input_name;
Y
Yu Yang 已提交
946 947 948 949 950 951 952 953 954 955 956 957 958
  for (auto& input : this->inputs_) {
    for (auto& ipt_name : input.second) {
      auto* var = scope.FindVar(ipt_name);
      if (var != nullptr) {
        const Tensor* t = nullptr;
        if (var->IsType<Tensor>()) {
          t = &var->Get<Tensor>();
        } else if (var->IsType<LoDTensor>()) {
          t = &var->Get<LoDTensor>();
        } else if (var->IsType<SelectedRows>()) {
          t = &(var->Get<SelectedRows>().value());
        }
        if (t != nullptr) {
S
fix bug  
sneaxiy 已提交
959 960
          PADDLE_ENFORCE(t->IsInitialized(), "Input %s is not initialized: %s",
                         ipt_name, DebugString());
Y
Yu Yang 已提交
961
          int tmp = static_cast<int>(t->type());
962 963
          PADDLE_ENFORCE(
              tmp == data_type || data_type == -1,
964 965
              "DataType of Paddle Op %s must be the same. Get %s(%d) != %s(%d)",
              Type(), last_input_name, data_type, ipt_name, tmp);
Y
Yu Yang 已提交
966
          data_type = tmp;
967
          last_input_name = ipt_name;
Y
Yu Yang 已提交
968 969 970 971 972
        }
      }
    }
  }
  PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
973
  return static_cast<proto::VarType::Type>(data_type);
Y
Yu Yang 已提交
974
}
975

976 977 978 979 980 981 982 983
OpKernelType OperatorWithKernel::GetExpectedKernelType(
    const ExecutionContext& ctx) const {
  return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
}

OpKernelType OperatorWithKernel::GetKernelTypeForVar(
    const std::string& var_name, const Tensor& tensor,
    const OpKernelType& expected_kernel_type) const {
M
mozga-intel 已提交
984 985
  return OpKernelType(expected_kernel_type.data_type_, tensor.place(),
                      tensor.layout());
986 987
}

Q
Qiao Longfei 已提交
988
}  // namespace framework
L
liaogang 已提交
989
}  // namespace paddle