operator.cc 74.7 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 "paddle/fluid/framework/operator.h"

17
#include <glog/logging.h>
P
peizhilin 已提交
18 19
#include <sstream>
#include <string>
20

21
#include "gflags/gflags.h"
Y
Yi Wang 已提交
22
#include "paddle/fluid/framework/data_transform.h"
23
#include "paddle/fluid/framework/data_type_transform.h"
W
WangXi 已提交
24
#include "paddle/fluid/framework/details/nan_inf_utils.h"
25
#include "paddle/fluid/framework/op_call_stack.h"
Y
Yi Wang 已提交
26
#include "paddle/fluid/framework/shape_inference.h"
27
#include "paddle/fluid/framework/transfer_scope_cache.h"
28
#include "paddle/fluid/framework/unused_var_check.h"
Y
Yi Wang 已提交
29
#include "paddle/fluid/framework/var_type.h"
L
Leo Chen 已提交
30
#include "paddle/fluid/platform/enforce.h"
31
#include "paddle/fluid/platform/profiler.h"
32
#include "paddle/pten/common/scalar.h"
33
#include "paddle/pten/common/scalar_array.h"
34 35 36 37 38 39

namespace paddle {
namespace framework {
class LoDTensor;
}  // namespace framework
}  // namespace paddle
40
#ifdef PADDLE_WITH_XPU
41 42
#include "paddle/fluid/platform/device/xpu/xpu_info.h"
#include "paddle/fluid/platform/device/xpu/xpu_op_list.h"
43
#endif
Q
Qiao Longfei 已提交
44

45 46 47 48
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif

D
dzhwinter 已提交
49
DECLARE_bool(benchmark);
50
DECLARE_bool(check_nan_inf);
51
DECLARE_bool(enable_unused_var_check);
52 53
PADDLE_DEFINE_EXPORTED_int32(inner_op_parallelism, 0,
                             "number of threads for inner op");
54
DECLARE_bool(run_pten_kernel);
D
dzhwinter 已提交
55

Q
Qiao Longfei 已提交
56 57 58
namespace paddle {
namespace framework {

59 60 61 62 63 64
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
65

66
static DDim GetDimsDebug(const ScopeBase& scope, const std::string& name,
67
                         bool get_actual_dim = false) {
68
  Variable* var = scope.FindVar(name);
Q
qiaolongfei 已提交
69 70
  if (var == nullptr) {
    return DDim({-1});
Q
Qiao Longfei 已提交
71 72
  }

M
minqiyang 已提交
73 74 75 76 77 78 79 80 81
  if (var->IsType<LoDTensor>()) {
    const LoDTensor& tensor = var->Get<LoDTensor>();
    return tensor.dims();
  } else if (var->IsType<SelectedRows>()) {
    if (get_actual_dim) {
      return var->Get<SelectedRows>().value().dims();
    } else {
      return var->Get<SelectedRows>().GetCompleteDims();
    }
S
Steffy-zxf 已提交
82 83
  } else if (var->IsType<Strings>()) {
    return DDim({static_cast<int64_t>(var->Get<Strings>().size())});
84 85 86 87 88
  } else {
    return DDim({-1});
  }
}

89
static bool VarInited(const ScopeBase& scope, const std::string& name) {
Q
Qiao Longfei 已提交
90 91 92 93 94
  Variable* var = scope.FindVar(name);
  if (var == nullptr) return false;
  return var->IsInitialized();
}

95
static std::string GetDtype(const ScopeBase& scope, const std::string& name) {
D
dzhwinter 已提交
96 97 98 99
  Variable* var = scope.FindVar(name);
  if (var == nullptr) {
    return "";
  }
100

M
minqiyang 已提交
101 102 103
  if (var->IsType<LoDTensor>()) {
    const LoDTensor& tensor = var->Get<LoDTensor>();
    if (UNLIKELY(!tensor.IsInitialized())) {
104 105
      return "";
    }
Y
Yu Yang 已提交
106
    return DataTypeToString(tensor.type());
M
minqiyang 已提交
107
  } else if (var->IsType<SelectedRows>()) {
Q
Qiao Longfei 已提交
108 109 110 111
    auto tensor = var->Get<SelectedRows>().value();
    if (UNLIKELY(!tensor.IsInitialized())) {
      return "uninited";
    } else {
Y
Yu Yang 已提交
112
      return DataTypeToString(tensor.type());
Q
Qiao Longfei 已提交
113
    }
S
Steffy-zxf 已提交
114 115
  } else if (var->IsType<Strings>()) {
    return "strings";
D
dzhwinter 已提交
116 117 118 119 120
  } else {
    return "";
  }
}

121
static std::string GetPlace(const ScopeBase& scope, const std::string& name) {
L
Leo Chen 已提交
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
  Variable* var = scope.FindVar(name);
  if (var == nullptr) {
    return "";
  }
  auto to_string = [](const platform::Place& p) {
    std::stringstream sstream;
    sstream << p;
    return sstream.str();
  };

  if (var->IsType<LoDTensor>()) {
    const LoDTensor& tensor = var->Get<LoDTensor>();
    if (UNLIKELY(!tensor.IsInitialized())) {
      return "";
    }
    return to_string(tensor.place());
  } else if (var->IsType<SelectedRows>()) {
    auto tensor = var->Get<SelectedRows>().value();
    if (UNLIKELY(!tensor.IsInitialized())) {
      return "uninited";
    } else {
      return to_string(tensor.place());
    }
  } else {
    return "";
  }
}

150
static int GetRowSize(const ScopeBase& scope, const std::string& name) {
151 152 153 154 155
  Variable* var = scope.FindVar(name);
  if (var == nullptr) {
    return -1;
  }

M
minqiyang 已提交
156 157
  if (var->IsType<SelectedRows>()) {
    return var->Get<SelectedRows>().rows().size();
158 159 160 161 162
  }

  return -1;
}

163
static LoD GetLoDDebug(const ScopeBase& scope, const std::string& name) {
Q
Qiao Longfei 已提交
164 165 166 167 168 169 170
  Variable* var = scope.FindVar(name);
  auto default_lod = LoD({{}});

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

M
minqiyang 已提交
171 172 173
  if (var->IsType<LoDTensor>()) {
    const LoDTensor& tensor = var->Get<LoDTensor>();
    return tensor.lod();
Q
Qiao Longfei 已提交
174 175 176 177 178
  } else {
    return default_lod;
  }
}

X
Xin Pan 已提交
179 180 181 182 183
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];
X
Xin Pan 已提交
184
    input_vars.reserve(var_name_item.second.size());
X
Xin Pan 已提交
185 186 187 188 189 190
    for (auto& var_name : var_name_item.second) {
      input_vars.push_back(scope.FindVar(var_name));
    }
  }
  for (auto& var_name_item : outnames) {
    std::vector<Variable*>& output_vars = outputs[var_name_item.first];
X
Xin Pan 已提交
191
    output_vars.reserve(var_name_item.second.size());
X
Xin Pan 已提交
192 193 194 195 196 197
    for (auto& var_name : var_name_item.second) {
      output_vars.push_back(scope.FindVar(var_name));
    }
  }
}

198
void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
P
peizhilin 已提交
199 200 201
  try {
    VLOG(4) << place << " " << DebugStringEx(&scope);
    if (platform::is_gpu_place(place)) {
202
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
203 204 205 206
      PADDLE_THROW(platform::errors::Unavailable(
          "Cannot run operator on place %s, please recompile paddle or "
          "reinstall Paddle with CUDA support.",
          place));
207
#else
208
      auto dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
P
peizhilin 已提交
209
      platform::SetDeviceId(dev_id);
210 211 212
#endif
    } else if (platform::is_xpu_place(place)) {
#ifndef PADDLE_WITH_XPU
213 214 215 216
      PADDLE_THROW(platform::errors::Unavailable(
          "Cannot run operator on place %s, please recompile paddle or "
          "reinstall Paddle with XPU support.",
          place));
217 218 219
#else
      auto dev_id = BOOST_GET_CONST(platform::XPUPlace, place).device;
      platform::SetXPUDeviceId(dev_id);
220 221 222 223 224 225 226 227 228 229
#endif
    } else if (platform::is_npu_place(place)) {
#ifndef PADDLE_WITH_ASCEND_CL
      PADDLE_THROW(platform::errors::Unavailable(
          "Cannot run operator on place %s, please recompile paddle or "
          "reinstall Paddle with NPU support.",
          place));
#else
      auto dev_id = BOOST_GET_CONST(platform::NPUPlace, place).device;
      platform::SetNPUDeviceId(dev_id);
230
#endif
P
peizhilin 已提交
231
    }
P
peizhilin 已提交
232

233
    {
234 235 236 237 238 239
      // TODO(wangchaochaohu) : refine code to use only one RecordEvent)
      // in order to record different op type cost time
      // and different op name cost time,we set two event.
      platform::RecordEvent op_type_record_event(Type());
      auto op_name = platform::OpName(outputs_, Type());
      platform::RecordEvent op_name_record_event(
240
          op_name, platform::EventRole::kUniqueOp);
P
peizhilin 已提交
241 242
      RunImpl(scope, place);
    }
243

Z
Zhang Ting 已提交
244
    VLOG(3) << GetExecutionPlace(place) << " " << DebugStringEx(&scope);
245
  } catch (platform::EnforceNotMet& exception) {
246
    framework::InsertCallStackInfo(Type(), Attrs(), &exception);
247
    throw std::move(exception);
248 249 250 251 252 253
  } catch (platform::EOFException&) {
    std::rethrow_exception(std::current_exception());
  } catch (std::exception& ex) {
    LOG(WARNING) << Type() << " raises an exception "
                 << platform::demangle(typeid(ex).name()) << ", " << ex.what();
    std::rethrow_exception(std::current_exception());
P
peizhilin 已提交
254
  } catch (...) {
255
    LOG(WARNING) << Type() << " raises an unknown exception";
P
peizhilin 已提交
256
    std::rethrow_exception(std::current_exception());
257
  }
258 259
}

260
bool OperatorBase::HasInputs(const std::string& name) const {
M
minqiyang 已提交
261
  return inputs_.find(name) != inputs_.end();
262 263
}

264
std::string OperatorBase::Input(const std::string& name) const {
Y
Yu Yang 已提交
265
  auto& ins = Inputs(name);
266 267
  PADDLE_ENFORCE_LE(
      ins.size(), 1UL,
268
      platform::errors::InvalidArgument(
269 270
          "Operator %s's input %s should contain only one variable.", type_,
          name));
Y
Yu Yang 已提交
271
  return ins.empty() ? kEmptyVarName : ins[0];
Y
Yan Chunwei 已提交
272 273
}

Y
Yu Yang 已提交
274 275
const std::vector<std::string>& OperatorBase::Inputs(
    const std::string& name) const {
Y
Yu Yang 已提交
276
  auto it = inputs_.find(name);
277 278 279 280
  PADDLE_ENFORCE_NE(
      it, inputs_.end(),
      platform::errors::NotFound("Operator %s does not have the input %s.",
                                 type_, name));
Y
Yu Yang 已提交
281
  return it->second;
Y
Yan Chunwei 已提交
282 283
}

284
bool OperatorBase::HasOutputs(const std::string& name) const {
285
  if (outputs_.find(name) != outputs_.end()) {
286 287 288 289 290 291
    return true;
  } else {
    return false;
  }
}

292
std::string OperatorBase::Output(const std::string& name) const {
Y
Yu Yang 已提交
293
  auto& outs = Outputs(name);
294 295 296 297 298
  PADDLE_ENFORCE_LE(
      outs.size(), 1UL,
      platform::errors::InvalidArgument(
          "Operator %s's output %s should contain only one variable.", type_,
          name));
Y
Yu Yang 已提交
299
  return outs.empty() ? kEmptyVarName : outs[0];
Y
Yan Chunwei 已提交
300 301
}

Y
Yu Yang 已提交
302 303
const std::vector<std::string>& OperatorBase::Outputs(
    const std::string& name) const {
Y
Yu Yang 已提交
304
  auto it = outputs_.find(name);
305 306 307 308
  PADDLE_ENFORCE_NE(
      it, outputs_.end(),
      platform::errors::NotFound(
          "Operator %s does not have an output called %s.", type_, name));
Y
Yu Yang 已提交
309
  return it->second;
Y
Yan Chunwei 已提交
310 311
}

312
std::string OperatorBase::DebugStringEx(const ScopeBase* scope) const {
Q
Qiao Longfei 已提交
313
  std::stringstream ss;
Y
Yu Yang 已提交
314
  ss << "Op(" << type_ << "), inputs:{";
315

316
  const std::unordered_set<std::string>* no_need_buffer_vars = nullptr;
317 318
  if (info_ && info_->NoNeedBufferVarsInferer()) {
    no_need_buffer_vars =
319 320
        &(Info().NoNeedBufferVarsInferer()(Inputs(), Outputs(), Attrs()));
    if (no_need_buffer_vars->empty()) no_need_buffer_vars = nullptr;
321 322
  }

Y
Yu Yang 已提交
323 324
  for (auto it = inputs_.begin(); it != inputs_.end();) {
    auto& input = *it;
325 326
    bool is_no_need_buffer_var =
        (no_need_buffer_vars && no_need_buffer_vars->count(input.first) > 0);
Y
Yu Yang 已提交
327 328
    ss << input.first << "[";
    for (size_t i = 0; i < input.second.size(); ++i) {
Q
Qiao Longfei 已提交
329 330
      auto var_name = input.second[i];
      ss << var_name;
331
      if (scope) {
Q
Qiao Longfei 已提交
332 333 334 335 336 337 338
        if (!VarInited(*scope, var_name)) {
          ss << "[uninited]";
        } else {
          int row_size = GetRowSize(*scope, var_name);
          if (row_size >= 0) {
            ss << "[row_size=" << row_size << "]";
          }
339 340 341
          std::string dtype = is_no_need_buffer_var
                                  ? "unknown_dtype"
                                  : GetDtype(*scope, var_name);
Q
Qiao Longfei 已提交
342
          ss << ":" << dtype;
343 344
          ss << "[" << GetDimsDebug(*scope, var_name, true) << "]";
          ss << "(" << GetLoDDebug(*scope, var_name) << ")";
L
Leo Chen 已提交
345
          ss << "(" << GetPlace(*scope, var_name) << ")";
346
        }
347
      }
Y
Yu Yang 已提交
348 349 350
      if (i != input.second.size() - 1) {
        ss << ", ";
      }
351
    }
Y
Yu Yang 已提交
352
    ss << "]";
Y
Yu Yang 已提交
353 354
    ++it;
    if (it != inputs_.end()) {
355 356
      ss << ", ";
    }
Q
Qiao Longfei 已提交
357
  }
Y
Yu Yang 已提交
358
  ss << "}, outputs:{";
Y
Yu Yang 已提交
359 360
  for (auto it = outputs_.begin(); it != outputs_.end();) {
    auto& output = *it;
Y
Yu Yang 已提交
361 362
    ss << output.first << "[";
    for (size_t i = 0; i < output.second.size(); ++i) {
Q
Qiao Longfei 已提交
363 364
      auto var_name = output.second[i];
      ss << var_name;
365
      if (scope) {
Q
Qiao Longfei 已提交
366 367 368 369 370 371 372
        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
373 374
          std::string dtype = GetDtype(*scope, output.second[i]);
          ss << ":" << dtype;
375 376
          ss << "[" << GetDimsDebug(*scope, var_name, true) << "]";
          ss << "(" << GetLoDDebug(*scope, var_name) << ")";
L
Leo Chen 已提交
377
          ss << "(" << GetPlace(*scope, var_name) << ")";
378
        }
379
      }
Y
Yu Yang 已提交
380 381 382
      if (i != output.second.size() - 1) {
        ss << ", ";
      }
383
    }
Y
Yu Yang 已提交
384
    ss << "]";
Y
Yu Yang 已提交
385 386
    ++it;
    if (it != outputs_.end()) {
387 388
      ss << ", ";
    }
Q
Qiao Longfei 已提交
389
  }
Y
Yu Yang 已提交
390
  ss << "}.";
Q
Qiao Longfei 已提交
391 392 393
  return ss.str();
}

Y
Yu Yang 已提交
394
OperatorBase::OperatorBase(const std::string& type,
Y
Yu Yang 已提交
395 396
                           const VariableNameMap& inputs,
                           const VariableNameMap& outputs,
Y
Yu Yang 已提交
397
                           const AttributeMap& attrs)
S
sneaxiy 已提交
398 399 400 401 402 403
    : type_(type),
      inputs_(inputs),
      outputs_(outputs),
      attrs_(attrs),
      // NOTE(zjl): why op_info may be nullptr?
      info_(OpInfoMap::Instance().GetNullable(type)) {
H
hong 已提交
404 405 406 407 408 409 410 411
  // In dygraph mode, all the OperatorBase will be constructed by function:
  // framework::OpRegistry::CreateOp(type, {}, {}, {}, false).
  // Inputs, outputs and attrs will be set to empty map
  // to improve the execution efficiency of dygraph.
  if (inputs_.size() > 0 || outputs_.size() > 0) {
    GenerateTemporaryNames();
    CheckAllInputOutputSet();
  }
Y
Yu Yang 已提交
412
}
413

Q
qijun 已提交
414 415
std::vector<std::string> OperatorBase::InputVars() const {
  std::vector<std::string> ret_val;
Y
Yu Yang 已提交
416
  for (auto& o : inputs_) {
Q
qijun 已提交
417 418 419 420 421 422
    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
423 424 425 426 427 428 429 430 431 432
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;
  }
S
sneaxiy 已提交
433
  auto& info = Info();
Y
Yu Yang 已提交
434 435

  // get all OpProto::Var for outputs
Y
Yu Yang 已提交
436
  for (auto& o : info.Proto().outputs()) {
Y
Yu Yang 已提交
437 438 439 440 441 442 443 444 445
    // 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
446 447
}

448
void OperatorBase::CheckAllInputOutputSet() const {
S
sneaxiy 已提交
449
  if (info_ == nullptr || info_->proto_ == nullptr) return;
450

S
sneaxiy 已提交
451
  for (auto& in : info_->Proto().inputs()) {
452
    if (!in.dispensable() && !in.extra()) {
453 454 455 456
      PADDLE_ENFORCE_NE(
          inputs_.find(in.name()), inputs_.end(),
          platform::errors::NotFound("Operator %s's input (%s) is not set.",
                                     Type(), in.name()));
457
    }
458 459
  }

S
sneaxiy 已提交
460
  for (auto& out : info_->Proto().outputs()) {
461
    if (!out.dispensable() && !out.extra()) {
462 463 464 465
      PADDLE_ENFORCE_NE(
          outputs_.find(out.name()), outputs_.end(),
          platform::errors::NotFound("Operator %s's output (%s) is not set.",
                                     Type(), out.name()));
466
    }
467 468 469 470 471 472 473 474 475 476 477 478 479 480 481
  }
}

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));
      }
    }
  }
}
482

C
chengduo 已提交
483
const Tensor* GetLoDTensorOrSelectedRowsValueFromVar(const Variable& var) {
C
chengduo 已提交
484 485 486 487
  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
488
  } else {
489 490 491
    PADDLE_THROW(platform::errors::InvalidArgument(
        "Variable type is %s, expect LoDTensor or SelectedRows.",
        ToTypeName(var.Type())));
Q
QI JUN 已提交
492 493 494
  }
}

C
chengduo 已提交
495
Tensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var) {
Q
QI JUN 已提交
496
  if (var->IsType<LoDTensor>()) {
497
    return var->GetMutable<LoDTensor>();
Q
QI JUN 已提交
498
  } else if (var->IsType<SelectedRows>()) {
499
    return var->GetMutable<SelectedRows>()->mutable_value();
Q
QI JUN 已提交
500
  } else {
501 502 503
    PADDLE_THROW(platform::errors::InvalidArgument(
        "Variable type is %s, expect LoDTensor or SelectedRows.",
        ToTypeName(var->Type())));
Q
QI JUN 已提交
504 505 506
  }
}

507
bool ExecutionContext::HasInput(const std::string& name) const {
508
  auto* var = InputVar(name);
509 510 511 512
  return var != nullptr;
}

bool ExecutionContext::HasOutput(const std::string& name) const {
513
  auto* var = OutputVar(name);
514 515 516
  return var != nullptr;
}

X
Xin Pan 已提交
517
const Variable* ExecutionContext::InputVar(const std::string& name) const {
518 519
  LogVarUsageIfUnusedVarCheckEnabled(name);

X
Xin Pan 已提交
520 521 522
  auto it = ctx_.inputs.find(name);
  if (it == ctx_.inputs.end()) return nullptr;

523 524
  PADDLE_ENFORCE_LE(
      it->second.size(), 1UL,
525
      platform::errors::InvalidArgument(
526 527
          "Operator %s's input %s should contain only one variable.",
          op_.Type(), name));
X
Xin Pan 已提交
528 529 530
  return it->second.empty() ? nullptr : it->second[0];
}

X
clean  
Xin Pan 已提交
531
Variable* ExecutionContext::OutputVar(const std::string& name) const {
X
Xin Pan 已提交
532 533 534
  auto it = ctx_.outputs.find(name);
  if (it == ctx_.outputs.end()) return nullptr;

535 536 537 538 539
  PADDLE_ENFORCE_LE(
      it->second.size(), 1UL,
      platform::errors::InvalidArgument(
          "Operator %s's output %s should contain only one variable.",
          op_.Type(), name));
X
Xin Pan 已提交
540 541 542
  return it->second.empty() ? nullptr : it->second[0];
}

543
template <>
544
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
C
chengduo 已提交
545
  return Input<LoDTensor>(name);
546 547 548
}

template <>
549
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
550
    const std::string& name) const {
551 552
  LogVarUsageIfUnusedVarCheckEnabled(name);

H
hong 已提交
553 554
  auto vars = MultiInputVar(name);
  if (vars.size() == 0) {
X
Xin Pan 已提交
555 556 557 558 559
    return {};
  }
  std::vector<const Tensor*> res;
  res.reserve(vars.size());
  std::transform(vars.begin(), vars.end(), std::back_inserter(res),
H
hong 已提交
560
                 [&](const Variable* var) -> const Tensor* {
X
Xin Pan 已提交
561
                   if (var == nullptr) return nullptr;
562 563 564 565 566
                   PADDLE_ENFORCE_EQ(var->IsType<LoDTensor>(), true,
                                     platform::errors::InvalidArgument(
                                         "Input variable should be LoDTensor, "
                                         "but the received type is %s.",
                                         ToTypeName(var->Type())));
X
Xin Pan 已提交
567 568 569 570 571
                   return &(var->Get<LoDTensor>());
                 });
  return res;
}

572
template <>
573
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
C
chengduo 已提交
574
  return Output<LoDTensor>(name);
575 576 577
}

template <>
578
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
579
    const std::string& name) const {
H
hong 已提交
580 581 582
  auto vars = MultiOutputVar(name);

  if (vars.size() == 0) {
583 584
    return {};
  }
585
  std::vector<Tensor*> res;
586 587 588 589 590
  res.reserve(vars.size());
  std::transform(vars.begin(), vars.end(), std::back_inserter(res),
                 [&](Variable* var) -> Tensor* {
                   return var == nullptr ? nullptr
                                         : var->GetMutable<LoDTensor>();
591
                 });
592 593 594
  return res;
}

Y
Yu Yang 已提交
595 596 597 598 599 600 601 602 603 604 605 606 607 608 609
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;
}

610 611
class RuntimeInferShapeContext : public InferShapeContext {
 public:
612
  RuntimeInferShapeContext(const OperatorBase& op, const RuntimeContext& ctx)
G
Gabor Buella 已提交
613
      : op_(op), ctx_(ctx) {}
614 615

  bool HasInput(const std::string& name) const override {
616
    // has only one input
X
Xin Pan 已提交
617
    const auto& ins = ctx_.inputs;
618 619
    auto it = ins.find(name);
    if (it == ins.end()) {
620 621
      return false;
    }
622
    const auto& in = it->second;
X
Xin Pan 已提交
623
    if (in.size() == 0) return false;
624 625 626 627
    PADDLE_ENFORCE_EQ(
        in.size(), 1UL,
        platform::errors::InvalidArgument(
            "Input %s should not contain more than one inputs.", name));
X
Xin Pan 已提交
628
    return in[0] != nullptr;
629 630 631
  }

  bool HasOutput(const std::string& name) const override {
632
    // has only one output
X
Xin Pan 已提交
633
    const auto& outs = ctx_.outputs;
634 635
    auto it = outs.find(name);
    if (it == outs.end()) {
636 637
      return false;
    }
638
    const auto& out = it->second;
X
Xin Pan 已提交
639
    if (out.size() == 0) {
640 641
      return false;
    }
642 643 644 645
    PADDLE_ENFORCE_EQ(
        out.size(), 1UL,
        platform::errors::InvalidArgument(
            "Output %s should not contain more than one outputs.", name));
X
Xin Pan 已提交
646
    return out[0] != nullptr;
647 648 649
  }

  bool HasInputs(const std::string& name) const override {
X
Xin Pan 已提交
650 651
    const auto& ins = ctx_.inputs;
    auto it = ins.find(name);
X
fix  
Xin Pan 已提交
652
    if (it == ins.end() || it->second.empty()) {
653 654
      return false;
    }
X
Xin Pan 已提交
655 656
    for (auto& input : it->second) {
      if (input == nullptr) {
657 658 659 660 661 662 663
        return false;
      }
    }
    return true;
  }

  bool HasOutputs(const std::string& name) const override {
X
Xin Pan 已提交
664 665
    const auto& outs = ctx_.outputs;
    auto it = outs.find(name);
X
fix  
Xin Pan 已提交
666
    if (it == outs.end() || it->second.empty()) {
667 668
      return false;
    }
X
Xin Pan 已提交
669 670
    for (auto& output : it->second) {
      if (output == nullptr) {
671 672 673 674 675 676 677 678
        return false;
      }
    }
    return true;
  }

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

H
hong 已提交
679
  std::vector<std::string> Inputs(const std::string& name) const override {
680 681 682
    return op_.Inputs(name);
  }

H
hong 已提交
683
  std::vector<std::string> Outputs(const std::string& name) const override {
684 685 686
    return op_.Outputs(name);
  }

687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
  std::string GetInputNameByIdx(size_t idx) const override {
    auto& op_proto =
        paddle::framework::OpInfoMap::Instance().Get(op_.Type()).proto_;
    PADDLE_ENFORCE_LT(idx, op_proto->inputs().size(),
                      platform::errors::OutOfRange(
                          "The index should be less than the size of inputs of "
                          "operator %s, but got index is %d and size is %d",
                          op_.Type(), idx, op_proto->inputs().size()));
    return op_proto->inputs()[idx].name();
  }

  std::string GetOutputNameByIdx(size_t idx) const override {
    auto& op_proto =
        paddle::framework::OpInfoMap::Instance().Get(op_.Type()).proto_;
    PADDLE_ENFORCE_LT(
        idx, op_proto->outputs().size(),
        platform::errors::OutOfRange(
            "The index should be less than the size of outputs of "
            "operator %s, but got index is %d and size is %d",
            op_.Type(), idx, op_proto->outputs().size()));
    return op_proto->outputs()[idx].name();
  }

710 711
  void ShareDim(const std::string& in, const std::string& out, size_t i = 0,
                size_t j = 0) override {
X
Xin Pan 已提交
712 713
    auto in_it = ctx_.inputs.find(in);
    auto out_it = ctx_.outputs.find(out);
714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729
    PADDLE_ENFORCE_NE(
        in_it, ctx_.inputs.end(),
        platform::errors::NotFound("Input %s does not exist.", in));
    PADDLE_ENFORCE_NE(
        out_it, ctx_.outputs.end(),
        platform::errors::NotFound("Output %s does not exist.", out));
    PADDLE_ENFORCE_LT(i, in_it->second.size(),
                      platform::errors::InvalidArgument(
                          "The index of input dimension is out of range, "
                          "excepted index less than %zu, but received %zu.",
                          in_it->second.size(), i));
    PADDLE_ENFORCE_LT(j, out_it->second.size(),
                      platform::errors::InvalidArgument(
                          "The index of output dimension is out of range, "
                          "excepted index less than %zu, but received %zu.",
                          out_it->second.size(), j));
X
Xin Pan 已提交
730 731 732

    Variable* in_var = in_it->second[i];
    Variable* out_var = out_it->second[j];
733

734 735 736 737 738
    PADDLE_ENFORCE_EQ(
        in_var->Type(), out_var->Type(),
        platform::errors::InvalidArgument(
            "The type of input (%s) and output (%s) are inconsistent.", in,
            out));
739 740 741 742 743 744 745 746 747 748 749 750

    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 {
751
      PADDLE_THROW(platform::errors::Unimplemented(
752
          "Currently, the input type of ShareDim only can be LoDTensor "
753
          "or SelectedRows."));
754 755 756
    }
  }

H
hong 已提交
757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774
  void ShareAllLoD(const std::string& in,
                   const std::string& out) const override {
    auto in_it = ctx_.inputs.find(in);
    auto out_it = ctx_.outputs.find(out);
    PADDLE_ENFORCE_NE(in_it, ctx_.inputs.end(),
                      platform::errors::NotFound(
                          "Input [%s] found error in Op [%s]", in, op_.Type()));
    PADDLE_ENFORCE_NE(
        out_it, ctx_.outputs.end(),
        platform::errors::NotFound("Output [%s] found error in Op [%s]", out,
                                   op_.Type()));

    auto& in_var_list = in_it->second;
    auto& out_var_list = out_it->second;

    PADDLE_ENFORCE_EQ(
        in_var_list.size(), out_var_list.size(),
        platform::errors::PreconditionNotMet(
T
tianshuo78520a 已提交
775
            "Op [%s]: Input var size should be equal with output var size",
H
hong 已提交
776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801
            op_.Type()));

    auto& out_var_names = op_.Outputs(out);

    for (size_t i = 0; i < in_var_list.size(); ++i) {
      if (out_var_names[i] == framework::kEmptyVarName) {
        continue;
      }

      Variable* in_var = in_var_list[i];
      if (!in_var->IsType<LoDTensor>()) return;
      Variable* out_var = out_var_list[i];
      PADDLE_ENFORCE_EQ(out_var->IsType<LoDTensor>(), true,
                        platform::errors::PreconditionNotMet(
                            "The %d-th output of Output(%s) must be LoDTensor.",
                            i, out_var_names[i]));
      auto& in_tensor = in_var->Get<LoDTensor>();
      auto* out_tensor = out_var->GetMutable<LoDTensor>();
      out_tensor->set_lod(in_tensor.lod());
#ifdef PADDLE_WITH_MKLDNN
      if (in_tensor.layout() != DataLayout::kMKLDNN)
#endif
        out_tensor->set_layout(in_tensor.layout());
    }
  }

Q
Qiao Longfei 已提交
802 803
  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
804 805
    auto in_it = ctx_.inputs.find(in);
    auto out_it = ctx_.outputs.find(out);
806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821
    PADDLE_ENFORCE_NE(
        in_it, ctx_.inputs.end(),
        platform::errors::NotFound("Input %s does not exist.", in));
    PADDLE_ENFORCE_NE(
        out_it, ctx_.outputs.end(),
        platform::errors::NotFound("Output %s does not exist.", out));
    PADDLE_ENFORCE_LT(i, in_it->second.size(),
                      platform::errors::InvalidArgument(
                          "The index of input dimension is out of range, "
                          "excepted index less than %zu, but received %zu.",
                          in_it->second.size(), i));
    PADDLE_ENFORCE_LT(j, out_it->second.size(),
                      platform::errors::InvalidArgument(
                          "The index of output dimension is out of range, "
                          "excepted index less than %zu, but received %zu.",
                          out_it->second.size(), j));
X
Xin Pan 已提交
822 823

    Variable* in_var = in_it->second.at(i);
Q
Qiao Longfei 已提交
824
    if (!in_var->IsType<LoDTensor>()) return;
X
Xin Pan 已提交
825
    Variable* out_var = out_it->second.at(j);
826 827 828 829
    PADDLE_ENFORCE_EQ(
        out_var->IsType<LoDTensor>(), true,
        platform::errors::InvalidArgument(
            "The %zu-th output of Output(%s) must be LoDTensor.", j, out));
830
    auto& in_tensor = in_var->Get<LoDTensor>();
Q
Qiao Longfei 已提交
831 832
    auto* out_tensor = out_var->GetMutable<LoDTensor>();
    out_tensor->set_lod(in_tensor.lod());
D
dzhwinter 已提交
833

M
mozga-intel 已提交
834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852
// 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
853 854
  }

855
  int32_t GetLoDLevel(const std::string& in, size_t i = 0) const override {
856
    PADDLE_THROW(platform::errors::PreconditionNotMet(
857
        "GetLoDLevel is only used in compile time. The calculation of "
858
        "output's actual lod is different among operators so that should be "
859
        "set in the runtime kernel."));
860 861
  }

862 863
  void SetLoDLevel(const std::string& out, int32_t lod_level,
                   size_t j = 0) const override {
864
    PADDLE_THROW(platform::errors::PreconditionNotMet(
865
        "SetLoDLevel is only used in compile time. The calculation of "
866
        "output's actual lod is different among operators so that should be "
867
        "set in the runtime kernel."));
C
chengduo 已提交
868 869
  }

870 871
  bool IsRuntime() const override { return true; }

872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890
  // TODO(paddle-dev): Can this be template?
  std::vector<InferShapeVarPtr> GetInputVarPtrs(
      const std::string& name) override {
    const std::vector<Variable*>& vars = InputVars(name);
    std::vector<InferShapeVarPtr> res;
    res.reserve(vars.size());
    res.insert(res.begin(), vars.begin(), vars.end());
    return res;
  }

  std::vector<InferShapeVarPtr> GetOutputVarPtrs(
      const std::string& name) override {
    const std::vector<Variable*>& vars = OutputVars(name);
    std::vector<InferShapeVarPtr> res;
    res.reserve(vars.size());
    res.insert(res.begin(), vars.begin(), vars.end());
    return res;
  }

X
Xin Pan 已提交
891 892
  DDim GetInputDim(const std::string& name) const override {
    const std::vector<Variable*>& vars = InputVars(name);
893 894 895 896 897
    PADDLE_ENFORCE_EQ(
        vars.size(), 1UL,
        platform::errors::InvalidArgument(
            "Input(%s) should hold one element, but now it holds %zu elements.",
            name, vars.size()));
X
Xin Pan 已提交
898 899 900 901 902 903 904 905
    return this->GetDim(vars[0]);
  }

  std::vector<DDim> GetInputsDim(const std::string& name) const override {
    const std::vector<Variable*>& vars = InputVars(name);
    return GetDims(vars);
  }

X
Xin Pan 已提交
906 907 908 909 910 911 912 913 914 915
  std::vector<proto::VarType::Type> GetInputsVarType(
      const std::string& name) const override {
    return GetVarTypes(InputVars(name));
  }

  std::vector<proto::VarType::Type> GetOutputsVarType(
      const std::string& name) const override {
    return GetVarTypes(OutputVars(name));
  }

X
Xin Pan 已提交
916 917
  void SetOutputDim(const std::string& name, const DDim& dim) override {
    auto& vars = OutputVars(name);
918 919 920 921 922
    PADDLE_ENFORCE_EQ(
        vars.size(), 1UL,
        platform::errors::InvalidArgument("Output(%s) should hold one element, "
                                          "but now it holds %zu elements.",
                                          name, vars.size()));
X
Xin Pan 已提交
923 924 925 926 927 928 929 930 931
    SetDim(vars[0], dim);
  }

  void SetOutputsDim(const std::string& name,
                     const std::vector<DDim>& dims) override {
    auto& vars = OutputVars(name);
    SetDims(vars, dims);
  }

932
 protected:
X
Xin Pan 已提交
933
  DDim GetDim(Variable* var) const {
934 935
    PADDLE_ENFORCE_NOT_NULL(
        var, platform::errors::InvalidArgument("Input variable is nullptr."));
936 937 938 939 940
    if (var->IsType<LoDTensor>()) {
      return var->Get<LoDTensor>().dims();
    } else if (var->IsType<SelectedRows>()) {
      return var->Get<SelectedRows>().GetCompleteDims();
    } else {
941 942 943 944
      PADDLE_THROW(platform::errors::InvalidArgument(
          "Only LoDTensor or SelectedRows support 'GetDim', but input "
          "Variable's type is %s.",
          ToTypeName(var->Type())));
F
fengjiayi 已提交
945 946 947
    }
  }

X
Xin Pan 已提交
948 949 950 951 952 953 954 955
  std::vector<DDim> GetDims(const std::vector<Variable*>& vars) const {
    std::vector<DDim> ret;
    ret.reserve(vars.size());
    std::transform(vars.begin(), vars.end(), std::back_inserter(ret),
                   [this](Variable* var) { return this->GetDim(var); });
    return ret;
  }

F
fengjiayi 已提交
956
  std::vector<DDim> GetRepeatedDims(const std::string& name) const override {
957 958
    PADDLE_THROW(platform::errors::PreconditionNotMet(
        "GetRepeatedDims method only ban be used in compile time."));
959 960
  }

X
Xin Pan 已提交
961
  void SetDim(Variable* var, const DDim& dim) {
962 963 964 965 966
    if (var->IsType<LoDTensor>()) {
      var->GetMutable<LoDTensor>()->Resize(dim);
    } else if (var->IsType<SelectedRows>()) {
      var->GetMutable<SelectedRows>()->set_height(dim[0]);
    } else {
967 968 969 970
      PADDLE_THROW(platform::errors::Unimplemented(
          "Variable type error, expect LoDTensor or SelectedRows, but received "
          "(%s).",
          ToTypeName(var->Type())));
X
Xin Pan 已提交
971 972 973 974 975 976
    }
  }

  void SetDims(const std::vector<Variable*>& vars,
               const std::vector<DDim>& dims) {
    size_t length = vars.size();
977 978 979 980 981 982
    PADDLE_ENFORCE_EQ(length, dims.size(),
                      platform::errors::InvalidArgument(
                          "The number of input variables do not match the "
                          "number of input dimensions, the number of variables "
                          "is %zu, the number of dimensions is %zu.",
                          length, dims.size()));
X
Xin Pan 已提交
983 984 985 986 987
    for (size_t i = 0; i < length; ++i) {
      if (vars[i] == nullptr) {
        continue;
      }
      SetDim(vars[i], dims[i]);
988 989 990
    }
  }

F
fengjiayi 已提交
991 992
  void SetRepeatedDims(const std::string& name,
                       const std::vector<DDim>& dims) override {
993 994
    PADDLE_THROW(platform::errors::PreconditionNotMet(
        "SetRepeatedDims method only can be used in compile time."));
F
fengjiayi 已提交
995 996
  }

X
Xin Pan 已提交
997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
  std::vector<proto::VarType::Type> GetVarTypes(
      const std::vector<Variable*>& vars) const {
    std::vector<proto::VarType::Type> retv;
    retv.resize(vars.size());
    std::transform(vars.begin(), vars.end(), retv.begin(),
                   std::bind(std::mem_fn(&RuntimeInferShapeContext::GetVarType),
                             this, std::placeholders::_1));
    return retv;
  }

  proto::VarType::Type GetVarType(Variable* var) const {
1008 1009 1010
    return ToVarType(var->Type());
  }

1011 1012 1013
 private:
  const std::vector<Variable*>& InputVars(const std::string& name) const {
    auto it = ctx_.inputs.find(name);
1014 1015 1016 1017
    PADDLE_ENFORCE_NE(
        it, ctx_.inputs.end(),
        platform::errors::NotFound(
            "Operator (%s) does not have the input (%s).", op_.Type(), name));
1018 1019 1020 1021 1022
    return it->second;
  }

  const std::vector<Variable*>& OutputVars(const std::string& name) const {
    auto it = ctx_.outputs.find(name);
1023 1024 1025 1026
    PADDLE_ENFORCE_NE(
        it, ctx_.outputs.end(),
        platform::errors::NotFound(
            "Operator (%s) does not have the outputs (%s).", op_.Type(), name));
1027
    return it->second;
F
fengjiayi 已提交
1028 1029
  }

1030
  const OperatorBase& op_;
X
Xin Pan 已提交
1031
  const RuntimeContext& ctx_;
1032 1033
};

1034 1035
static void CheckTensorNANOrInf(const std::string& op_type,
                                const std::string& name,
C
chengduoZH 已提交
1036 1037 1038 1039
                                const framework::Tensor& tensor) {
  if (tensor.memory_size() == 0) {
    return;
  }
Y
Yu Yang 已提交
1040 1041
  if (tensor.type() != proto::VarType::FP32 &&
      tensor.type() != proto::VarType::FP64) {
C
chengduoZH 已提交
1042 1043
    return;
  }
1044 1045 1046 1047 1048 1049 1050 1051
  PADDLE_ENFORCE_NE(
      framework::TensorContainsInf(tensor), true,
      platform::errors::Fatal("Operator %s output Tensor %s contains Inf.",
                              op_type, name));
  PADDLE_ENFORCE_NE(
      framework::TensorContainsNAN(tensor), true,
      platform::errors::Fatal("Operator %s output Tensor %s contains NAN.",
                              op_type, name));
C
chengduoZH 已提交
1052 1053
}

1054 1055
bool OperatorWithKernel::SupportsMKLDNN(
    const proto::VarType::Type data_type) const {
1056 1057
  auto& op_kernels = OperatorWithKernel::AllOpKernels().at(type_);
  return std::any_of(op_kernels.begin(), op_kernels.end(),
1058
                     [data_type](OpKernelMap::const_reference kern_pair) {
1059 1060
                       return platform::is_cpu_place(kern_pair.first.place_) &&
                              kern_pair.first.library_type_ ==
1061 1062
                                  LibraryType::kMKLDNN &&
                              kern_pair.first.data_type_ == data_type;
1063 1064 1065
                     });
}

1066 1067
bool OperatorWithKernel::CanMKLDNNBeUsed(const framework::ExecutionContext& ctx,
                                         proto::VarType::Type data_type) const {
1068 1069
  bool use_mkldnn_ctx =
      ctx.Attr<bool>("use_mkldnn") && platform::is_cpu_place(ctx.GetPlace());
1070
  return use_mkldnn_ctx && this->SupportsMKLDNN(data_type);
1071 1072
}

B
baojun-nervana 已提交
1073
void OperatorWithKernel::RuntimeInferShape(const Scope& scope,
X
Xin Pan 已提交
1074 1075
                                           const platform::Place& place,
                                           const RuntimeContext& ctx) const {
1076
  RuntimeInferShapeContext infer_shape_ctx(*this, ctx);
B
baojun-nervana 已提交
1077 1078 1079
  this->InferShape(&infer_shape_ctx);
}

L
luotao1 已提交
1080 1081
void OperatorWithKernel::RunImpl(const Scope& scope,
                                 const platform::Place& place) const {
L
luotao1 已提交
1082 1083
  // To reduce the elapsed time of HasAttr, we use bool variable to record the
  // result of HasAttr.
1084 1085 1086
  if (!enable_cache_runtime_context_ && HasAttr(kEnableCacheRuntimeContext))
    enable_cache_runtime_context_ = true;
  if (!all_kernels_must_compute_runtime_shape_ &&
L
luotao1 已提交
1087
      HasAttr(kAllKernelsMustComputeRuntimeShape))
1088
    all_kernels_must_compute_runtime_shape_ = true;
1089
  const Scope* cur_scope = &scope;
1090
  if (!enable_cache_runtime_context_) {
L
luotao1 已提交
1091 1092
    RuntimeContext ctx(Inputs(), Outputs(), scope);
    RunImpl(scope, place, &ctx);
1093
    pre_scope_ = cur_scope;
L
luotao1 已提交
1094
  } else {
1095 1096 1097 1098 1099 1100
    if (runtime_ctx_.get() == nullptr || pre_scope_ != cur_scope) {
      std::lock_guard<std::mutex> lock(cache_update_mutex_);
      if (runtime_ctx_.get() == nullptr || pre_scope_ != cur_scope) {
        runtime_ctx_.reset(new RuntimeContext(Inputs(), Outputs(), scope));
        pre_scope_ = cur_scope;
      }
L
luotao1 已提交
1101 1102 1103 1104 1105 1106 1107 1108
    }
    RunImpl(scope, place, runtime_ctx_.get());
  }
}

void OperatorWithKernel::RunImpl(const Scope& scope,
                                 const platform::Place& place,
                                 RuntimeContext* runtime_ctx) const {
Y
Yu Yang 已提交
1109
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
1110
  auto* dev_ctx = pool.Get(place);
1111

1112 1113 1114 1115 1116 1117 1118 1119 1120 1121
#ifdef PADDLE_WITH_ASCEND_CL
  // NOTE(wangxi): nan/inf cannot be detected on NPU by checking the variable
  // values, but only through special `float_status` to checks whether
  // the operation is overflow. More about `float_status`, see:
  // https://gitee.com/ascend/modelzoo/issues/I3NF8V?from=project-issue
  if (FLAGS_check_nan_inf) {
    framework::details::NPUAllocAndClearFloatStatus(*this, scope, place);
  }
#endif

1122 1123 1124 1125 1126 1127 1128 1129 1130
  auto exe_ctx = ExecutionContext(*this, scope, *dev_ctx, *runtime_ctx);

  // TODO(chenweihang): Now we are still reusing a lot of the original fluid
  // implementation, this is a gradual replacement process
  // TODO(chenweihang): in the first phase of project, we only support CPU, CUDA
  // and RCOM backend, the XPU, NPU and MKLDNN will be supported in the second
  // phase
  if (FLAGS_run_pten_kernel &&
      pten::KernelFactory::Instance().HasCompatiblePtenKernel(type_)) {
1131
    if (pt_kernel_signature_ == nullptr || pt_kernel_ == nullptr) {
1132 1133 1134 1135 1136 1137 1138 1139
      ChoosePtenKernel(exe_ctx);
    }
    run_pten_kernel_ = pt_kernel_->IsValid();
  }
  if (!run_pten_kernel_) {
    if (kernel_type_.get() == nullptr || kernel_func_.get() == nullptr) {
      ChooseKernel(exe_ctx);
    }
1140 1141
  }

Y
yuyang18 已提交
1142 1143
  // do data transformScope &transfer_scope;
  std::vector<std::string> transfered_inplace_vars;
1144 1145
  Scope* transfer_scope = nullptr;
  {
1146
    platform::RecordEvent record_event("prepare_data",
1147
                                       platform::EventRole::kInnerOp);
1148 1149 1150 1151
    if (need_prepare_data_) {
      transfer_scope = PrepareData(scope, *kernel_type_,
                                   &transfered_inplace_vars, runtime_ctx);
    }
1152
  }
Y
yuyang18 已提交
1153 1154 1155 1156
  // exec scope is the scope that kernel actually executed on.
  const Scope& exec_scope =
      (transfer_scope == nullptr ? scope : *transfer_scope);

1157 1158
  if (!(kernel_type_->place_ == dev_ctx->GetPlace())) {
    dev_ctx = pool.Get(kernel_type_->place_);
1159
  }
Q
QI JUN 已提交
1160

1161
  if (!all_kernels_must_compute_runtime_shape_) {
1162
    platform::RecordEvent record_event("infer_shape",
1163
                                       platform::EventRole::kInnerOp);
1164
    RuntimeInferShapeContext infer_shape_ctx(*this, *runtime_ctx);
1165 1166
    this->InferShape(&infer_shape_ctx);
  }
1167 1168 1169 1170 1171

  if (FLAGS_enable_unused_var_check) {
    GetThreadLocalUsedVarNameSet()->clear();
  }

X
clean  
Xin Pan 已提交
1172 1173
  // TODO(panyx0718): ExecutionContext should only depend on RuntimeContext
  // not Scope. Imperative mode only pass inputs and get outputs.
1174
  {
1175
    platform::RecordEvent record_event("compute",
1176
                                       platform::EventRole::kInnerOp);
1177
    if (run_pten_kernel_) {
1178 1179 1180 1181 1182
      if (pt_kernel_context_ == nullptr) {
        pt_kernel_context_.reset(new pten::KernelContext());
      }
      BuildPtenKernelContext(*runtime_ctx, dev_ctx);
      (*pt_kernel_)(pt_kernel_context_.get());
1183 1184 1185

      WriteBackToOutputs(runtime_ctx);

1186
      pt_kernel_context_->ClearData();
1187 1188 1189 1190
    } else {
      (*kernel_func_)(
          ExecutionContext(*this, exec_scope, *dev_ctx, *runtime_ctx));
    }
1191
  }
D
dzhwinter 已提交
1192

Y
yuyang18 已提交
1193
  if (!transfered_inplace_vars.empty()) {
T
tianshuo78520a 已提交
1194
    // there is inplace variable has been transferred.
Y
yuyang18 已提交
1195
    TransferInplaceVarsBack(scope, transfered_inplace_vars, *transfer_scope);
1196
  }
1197 1198 1199 1200 1201 1202 1203

  // See [ Why need handle complex gradient to real gradient? ]
  // Only handle the case where the current kernel data type is complex
  if (framework::IsComplexType(kernel_type_->data_type_)) {
    HandleComplexGradToRealGrad(scope, runtime_ctx);
  }

1204 1205 1206 1207 1208 1209 1210 1211
  if (FLAGS_enable_unused_var_check) {
    // skip op that uses mkldnn because it has different memory reuse strategy.
    // use attr here because some GradMakers (like ActivationGradOpMaker) add
    // input when use_mkldnn=true;
    if (!(HasAttr("use_mkldnn") && Attr<bool>("use_mkldnn"))) {
      CheckUnusedVar(*this, scope);
    }
  }
1212

D
dzhwinter 已提交
1213
  /*For profiling/benchmark only*/
D
dzhwinter 已提交
1214
  if (FLAGS_benchmark) {
Y
yuyang18 已提交
1215
    dev_ctx->Wait();
L
Leo Chen 已提交
1216 1217 1218
#if defined(PADDLE_WITH_CUDA)
    PADDLE_ENFORCE_CUDA_SUCCESS(cudaGetLastError());
    VLOG(4) << "Operator(" << Type() << "): context wait and get last error";
1219 1220 1221 1222
#endif
#if defined(PADDLE_WITH_HIP)
    PADDLE_ENFORCE_CUDA_SUCCESS(hipGetLastError());
    VLOG(4) << "Operator(" << Type() << "): context wait and get last error";
L
Leo Chen 已提交
1223
#endif
D
dzhwinter 已提交
1224
  }
C
chengduoZH 已提交
1225 1226

  if (FLAGS_check_nan_inf) {
W
WangXi 已提交
1227
    framework::details::CheckOpHasNanOrInf(*this, exec_scope, place);
C
chengduoZH 已提交
1228
  }
1229 1230 1231 1232 1233 1234 1235

  // To solve issue #15032, have a discussion with @Luotao for cpu inference,
  // do not cache transfer scope, hence in this case delete transfer scope
  // after run to avoid memory leak
  if (transfer_scope && !run_by_executor_ && !enable_cache_transfer_scope_) {
    scope.DeleteScope(transfer_scope);
  }
Q
Qiao Longfei 已提交
1236
}
X
Xin Pan 已提交
1237

1238 1239 1240
OpKernelType OperatorWithKernel::InnerGetExpectedKernelType(
    const ExecutionContext& ctx) const {
  auto& dev_ctx = ctx.device_context();
L
Liu Yiqun 已提交
1241

1242
  auto expected_kernel_key = this->GetExpectedKernelType(ctx);
1243 1244 1245
  if (HasAttr("op_device")) {
    if (Attr<std::string>("op_device") == "cpu") {
      expected_kernel_key.place_ = platform::CPUPlace();
1246 1247 1248 1249 1250 1251 1252 1253 1254 1255
    } else if (Attr<std::string>("op_device").find("gpu") !=
               std::string::npos) {
      auto device = Attr<std::string>("op_device");
      size_t pos = device.find(':');
      if (pos != std::string::npos) {
        device = device.substr(0, pos);
        LOG_FIRST_N(WARNING, 1)
            << "Device index is only supported under pipeline parallelism, "
            << "so it will be ignored.";
      }
1256 1257 1258
      // when the Op that only has CPUKernel is assigned to GPU, the CPUKernel
      // will be executed and a warning will be given at the same time.
      if (SupportGPU()) {
1259
        expected_kernel_key.place_ = dev_ctx.GetPlace();
B
Baibaifan 已提交
1260
      } else if (SupportNPU()) {
1261
        expected_kernel_key.place_ = dev_ctx.GetPlace();
1262 1263 1264 1265 1266 1267 1268 1269
      } else {
        expected_kernel_key.place_ = platform::CPUPlace();
        LOG_FIRST_N(WARNING, 1)
            << "Op(" << type_
            << ") has no CUDA implementation. It will be assigned to CPUPlace.";
      }
    }
  }
C
cc 已提交
1270 1271
  VLOG(3) << "op type:" << type_
          << ", expected_kernel_key:" << expected_kernel_key;
1272 1273 1274 1275 1276 1277
  return expected_kernel_key;
}

void OperatorWithKernel::ChoosePtenKernel(const ExecutionContext& ctx) const {
  pt_kernel_signature_.reset(
      new KernelSignature(std::move(this->GetExpectedPtenKernelArgs(ctx))));
C
Chen Weihang 已提交
1278
  VLOG(6) << KernelSignatureToString(*pt_kernel_signature_.get());
1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289

  kernel_type_.reset(
      new OpKernelType(std::move(InnerGetExpectedKernelType(ctx))));

  auto pt_kernel_name = pten::KernelName(pt_kernel_signature_->name);
  auto pt_kernel_key = TransOpKernelTypeToPtenKernelKey(*kernel_type_.get());
  pt_kernel_.reset(
      new pten::Kernel(pten::KernelFactory::Instance().SelectKernel(
          pt_kernel_name, pt_kernel_key)));

  if (pt_kernel_->IsValid()) {
C
Chen Weihang 已提交
1290
    VLOG(6) << "Static mode ChoosePtenKernel - kernel name: " << pt_kernel_name
1291 1292 1293
            << " | kernel key: " << pt_kernel_key
            << " | kernel: " << *pt_kernel_;
  } else {
C
Chen Weihang 已提交
1294
    VLOG(6) << "Static mode ChoosePtenKernel - kernel `" << pt_kernel_name
1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
            << "` not found.";
  }
}

void OperatorWithKernel::ChooseKernel(const ExecutionContext& ctx) const {
  // check if op[type] has kernel registered.
  auto& all_op_kernels = AllOpKernels();
  auto kernels_iter = all_op_kernels.find(type_);
  PADDLE_ENFORCE_NE(
      kernels_iter, all_op_kernels.end(),
      platform::errors::Unavailable(
          "There are no kernels which are registered in the %s operator.",
          type_));

  OpKernelMap& kernels = kernels_iter->second;

  auto expected_kernel_key = InnerGetExpectedKernelType(ctx);
L
Liu Yiqun 已提交
1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322

  auto kernel_iter = kernels.find(expected_kernel_key);
#ifdef PADDLE_WITH_MKLDNN
  // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
  if (kernel_iter == kernels.end() &&
      expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
    VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
    expected_kernel_key.library_type_ = LibraryType::kPlain;
    expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
    kernel_iter = kernels.find(expected_kernel_key);
  }
1323 1324
#endif
#ifdef PADDLE_WITH_XPU
Q
QingshuChen 已提交
1325 1326 1327 1328
  if (is_xpu_place(expected_kernel_key.place_) &&
      (kernel_iter == kernels.end() ||
       !paddle::platform::is_xpu_support_op(type_, expected_kernel_key) ||
       paddle::platform::is_in_xpu_black_list(type_))) {
1329 1330 1331 1332 1333 1334
    VLOG(3) << "missing XPU kernel: " << type_
            << ", expected_kernel_key:" << expected_kernel_key
            << ", fallbacking to CPU one!";
    expected_kernel_key.place_ = platform::CPUPlace();
    kernel_iter = kernels.find(expected_kernel_key);
  }
1335 1336 1337 1338 1339 1340 1341 1342 1343 1344
#endif
#ifdef PADDLE_WITH_ASCEND_CL
  if (kernel_iter == kernels.end() &&
      is_npu_place(expected_kernel_key.place_)) {
    VLOG(3) << "missing NPU kernel: " << type_
            << ", expected_kernel_key:" << expected_kernel_key
            << ", fallbacking to CPU one!";
    expected_kernel_key.place_ = platform::CPUPlace();
    kernel_iter = kernels.find(expected_kernel_key);
  }
L
Liu Yiqun 已提交
1345
#endif
1346 1347 1348 1349
  PADDLE_ENFORCE_NE(kernel_iter, kernels.end(),
                    platform::errors::NotFound(
                        "Operator (%s) does not have kernel for %s.", type_,
                        KernelTypeToString(expected_kernel_key)));
L
Liu Yiqun 已提交
1350

1351 1352 1353 1354 1355
  std::lock_guard<std::mutex> lock(cache_update_mutex_);
  if (kernel_type_.get() == nullptr || kernel_func_.get() == nullptr) {
    kernel_type_.reset(new OpKernelType(expected_kernel_key));
    kernel_func_.reset(new OpKernelFunc(kernel_iter->second));
  }
L
Liu Yiqun 已提交
1356 1357
}

Y
yuyang18 已提交
1358 1359 1360 1361
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
1362
    VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
C
chengduo 已提交
1363
    auto* origin_var = scope.FindVar(var_name);
1364 1365 1366
    PADDLE_ENFORCE_NOT_NULL(origin_var,
                            platform::errors::InvalidArgument(
                                "The variable[%s] is nullptr.", var_name));
C
chengduo 已提交
1367
    auto* original_tensor =
C
chengduo 已提交
1368
        GetMutableLoDTensorOrSelectedRowsValueFromVar(origin_var);
C
chengduo 已提交
1369
    auto* var = transfer_scope.FindVar(var_name);
1370 1371
    PADDLE_ENFORCE_NOT_NULL(var, platform::errors::InvalidArgument(
                                     "The variable[%s] is nullptr.", var_name));
C
chengduo 已提交
1372
    auto* transformed_tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var);
1373
    auto original_dims = original_tensor->dims();
Y
yuyang18 已提交
1374
    original_tensor->ShareDataWith(*transformed_tensor);
B
Baibaifan 已提交
1375 1376 1377 1378 1379
    // In order to solve the problem that the output latitude of NPU reshape
    // operator is not changed when inplace.
    if (type_ != "reshape2" && type_ != "reshape2_grad") {
      original_tensor->Resize(original_dims);
    }
Y
yuyang18 已提交
1380 1381 1382
  }
}

1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449
void OperatorWithKernel::HandleComplexGradToRealGrad(
    const Scope& scope, RuntimeContext* ctx) const {
  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) {
      // 1. find grad_var & check whether is complex tensor
      auto var_name = var_name_item.second[i];
      auto orig_var_name = GradOriginalVarName(var_name);
      // only focus on gradient var
      if (var_name == orig_var_name) {
        continue;
      }
      auto* grad_var = output_vars[i];
      // skip nullptr var
      if (grad_var == nullptr) {
        continue;
      }
      // don't process LoDTensorArray temporarily,
      // add support if necessary for complex number calculations in the future
      if (!VarIsTensor(*grad_var)) {
        continue;
      }
      auto* grad_tensor =
          GetMutableLoDTensorOrSelectedRowsValueFromVar(grad_var);
      // skip nullptr tensor
      if (grad_tensor == nullptr || !grad_tensor->IsInitialized()) {
        continue;
      }
      // only focus on complex dtype now
      auto src_type = grad_tensor->type();
      if (!IsComplexType(src_type)) {
        continue;
      }

      // 2. find forward var & check whether need to cast
      auto* var = scope.FindVar(orig_var_name);
      // if forward var not exists, do nothing
      if (var == nullptr) {
        continue;
      }
      if (!VarIsTensor(*var)) {
        continue;
      }
      const auto* tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var);
      PADDLE_ENFORCE_NOT_NULL(
          tensor,
          platform::errors::Unavailable(
              "Forward tensor is nullptr when handle complex data to real."));
      // only need record type, the allocation may have been released
      auto dst_type = tensor->saved_type();
      // only focus on real dtype and need casting
      if (IsComplexType(dst_type)) {
        continue;
      }

      // 3. cast complex grad to real grad
      VLOG(6) << "Transform " << framework::DataTypeToString(src_type)
              << " var `" << var_name << "` to "
              << framework::DataTypeToString(dst_type)
              << " real var in static graph.";
      Tensor out;
      TransComplexToReal(dst_type, src_type, *grad_tensor, &out);
      SetTensorToVariable(*grad_var, out, grad_var);
    }
  }
}

X
Xin Pan 已提交
1450
Scope* OperatorWithKernel::PrepareData(
Y
yuyang18 已提交
1451
    const Scope& scope, const OpKernelType& expected_kernel_key,
X
Xin Pan 已提交
1452 1453
    std::vector<std::string>* transfered_inplace_vars,
    RuntimeContext* ctx) const {
Y
yuyang18 已提交
1454
  Scope* new_scope = nullptr;
S
sneaxiy 已提交
1455

1456
  const std::unordered_set<std::string>* no_buffer_ins = nullptr;
S
sneaxiy 已提交
1457 1458 1459 1460
  if (info_) {
    auto& no_buffer_inferer = info_->NoNeedBufferVarsInferer();
    // Some op may not register NoNeedBufferVarsInferer
    if (no_buffer_inferer) {
1461 1462
      no_buffer_ins = &(no_buffer_inferer(Inputs(), Outputs(), Attrs()));
      if (no_buffer_ins->empty()) no_buffer_ins = nullptr;
S
sneaxiy 已提交
1463 1464 1465
    }
  }

Y
yuyang18 已提交
1466
  for (auto& var_name_item : Inputs()) {
1467 1468
    bool should_skip_input =
        no_buffer_ins && no_buffer_ins->count(var_name_item.first) > 0;
S
sneaxiy 已提交
1469

X
Xin Pan 已提交
1470 1471 1472 1473
    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];
X
Xin Pan 已提交
1474
      auto* var = input_vars[i];
X
Xin Pan 已提交
1475

Y
yuyang18 已提交
1476
      // Only tensor can be tranfer to another device.
C
chengduo 已提交
1477
      if (var == nullptr || !VarIsTensor(*var)) {
Y
yuyang18 已提交
1478 1479 1480
        continue;
      }

C
chengduo 已提交
1481
      auto* tensor_in = GetLoDTensorOrSelectedRowsValueFromVar(*var);
1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496

      // When no_buffer_ins then checking of Tensor::holder_ is
      // not a thread safe. And for infershape scenario checks
      // to be omitted are not really needed
      if (should_skip_input == true) {
#ifdef PADDLE_WITH_MKLDNN
        // Var without buffer may be needed
        // for some situation like InferShape().
        // In this situation We cannot skip Var analysis, as
        // MKL-DNN shape of Var may differ from kNHWC Var
        // In such situation corressponding resized Var
        // has to be created and registered
        if ((tensor_in->layout() == DataLayout::kMKLDNN) &&
            (var->IsType<LoDTensor>() == true) &&
            (expected_kernel_key.data_layout_ != DataLayout::kMKLDNN) &&
1497 1498
            (paddle::platform::MKLDNNDeviceContext::tls()
                 .get_cur_paddle_data_layout() == DataLayout::kNHWC)) {
1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519
          // Mixed execution : MKL-DNN and GPU is not supported!
          if (!new_scope) {
            new_scope = &scope.NewScope();
          }
          auto* trans_var = new_scope->Var(var_name);
          input_vars[i] = trans_var;
          auto out = trans_var->GetMutable<LoDTensor>();
          out->Resize(tensor_in->dims());
          platform::MatchShapeToLayout(out, tensor_in->layout(),
                                       DataLayout::kNHWC);
          VLOG(7) << "Created reshaped dummy input based on MKL-DNN Tensor , "
                     "but kNHWC layout"
                  << var_name_item.first << " in Operator " << type_;
        } else {
          VLOG(7) << "Skip scanning input " << var_name_item.first
                  << " in Operator " << type_;
        }
#endif
        continue;
      }

Y
yuyang18 已提交
1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530
      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;
      }

M
minqiyang 已提交
1531 1532
      VLOG(3) << "Transform Variable " << var_name << " from "
              << kernel_type_for_var << " to " << expected_kernel_key;
Y
yuyang18 已提交
1533

1534 1535 1536
      // 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.
1537
      // We use a thread_local cache to fix that issue, the key in the cache is
1538 1539 1540 1541 1542
      // 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.
1543 1544
      // If this op is not called by an Executor or ParallelExecutor, it should
      // called by a NaiveExecutor, the NaiveExecutor will cache the scopes and
1545
      // variables, that behavior a lot different.
1546 1547 1548 1549 1550 1551 1552 1553 1554
      //
      // To solve issue #15032, have a discussion with @Luotao for cpu
      // inference, for all cpu kernels cases without GPU participation, here
      // not do transfer scope caching, and cpu inference performance is not
      // impacted by test.
      enable_cache_transfer_scope_ = false;
      if (!run_by_executor_ &&
          (platform::is_gpu_place(kernel_type_for_var.place_) ||
           platform::is_gpu_place(expected_kernel_key.place_))) {
1555 1556
        new_scope = TryCreateTransferScope(kernel_type_for_var,
                                           expected_kernel_key, &scope);
1557
        enable_cache_transfer_scope_ = true;
1558
      }
1559
      if (!new_scope) {
Y
yuyang18 已提交
1560 1561
        new_scope = &scope.NewScope();
      }
1562 1563 1564 1565
      // For inference, if a gpu model has an op which could only run on CPU,
      // each result of different input will be the same with the first one.
      // The reason is that if a gpu tensor is the input of a cpu kernel,
      // we will create a new cpu tensor in new scope.
1566
      // However, if enable_cache_runtime_context_, we get the cpu tensor each
1567 1568
      // time, not the gpu tensor. Thus, we set pre_scope_ = nullptr
      // to trigger `new RuntimeContext()` in RunImpl().
1569
      if (enable_cache_runtime_context_) {
1570 1571
        pre_scope_ = nullptr;
      }
L
Leo Chen 已提交
1572 1573

      // Create new var with the same name in transfer scopes
Y
yuyang18 已提交
1574
      auto* trans_var = new_scope->Var(var_name);
X
fix  
Xin Pan 已提交
1575
      input_vars[i] = trans_var;
L
Leo Chen 已提交
1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592

      // Find if inplace exists between input and output
      // If inplace exists, set the new created var to inplaced output, and
      // record its name in transfered_inplace_vars.
      for (auto& pair : Outputs()) {
        for (size_t j = 0; j < pair.second.size(); ++j) {
          if (pair.second[j] == var_name) {
            VLOG(4) << "Found inplace between input(" << var_name_item.first
                    << ") and output(" << pair.first
                    << "), the variable name is " << var_name;
            ctx->outputs[pair.first][j] = trans_var;
            transfered_inplace_vars->emplace_back(var_name);
          }
        }
      }

      // Do transfer
Y
yuyang18 已提交
1593
      Tensor out;
Y
yuyang18 已提交
1594
      TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in, &out);
Y
yuyang18 已提交
1595 1596 1597
      SetTensorToVariable(*var, out, trans_var);
    }
  }
L
Leo Chen 已提交
1598

1599 1600 1601 1602 1603 1604
  // If pre_scope = &scope, it means that scope is cached and the op is not in
  // while block. If new_scope = nullptr, it means that for each input of this
  // Op, there is no need to do PrepareData. So PrepareData could be skipped at
  // the rest iterations to save the elapsed time.
  // We do not support skipping PrepareData in while block, because the Op's
  // input may be changed by subsequent Ops, which may cause an error.
W
wenbin 已提交
1605 1606 1607 1608 1609 1610

  // For inference, ops that behind conditional branch aren't supported well,
  // so disable prepare optimization conservatively.
  bool force_prepare_data = HasAttr("inference_force_prepare_data") &&
                            Attr<bool>("inference_force_prepare_data");
  if (pre_scope_ == &scope && new_scope == nullptr && !force_prepare_data) {
1611 1612
    need_prepare_data_ = false;
  }
Y
yuyang18 已提交
1613 1614 1615

  return new_scope;
}
Q
Qiao Longfei 已提交
1616

1617
void OperatorWithKernel::ParseInputDataType(
1618
    const std::vector<Variable*>& vars, const std::string& name,
1619
    proto::VarType::Type* data_type) const {
1620
  proto::VarType::Type default_data_type =
1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631
      static_cast<proto::VarType::Type>(-1);
  for (size_t i = 0; i < vars.size(); ++i) {
    const Variable* var = vars[i];
    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());
1632
      } else if (var->IsType<LoDTensorArray>()) {
1633 1634 1635 1636
        auto t_arr = &var->Get<LoDTensorArray>();
        for (size_t j = 0; j < t_arr->size(); j++) {
          if (t_arr->at(j).IsInitialized()) {
            t = &(t_arr->at(j));
1637 1638
          }
        }
1639 1640
      }
      if (t != nullptr) {
1641 1642
        PADDLE_ENFORCE_EQ(
            t->IsInitialized(), true,
1643 1644 1645
            platform::errors::InvalidArgument("The %s Op's Input Variable `%s` "
                                              "contains uninitialized Tensor.",
                                              Type(), name));
1646
        proto::VarType::Type tmp = t->type();
1647 1648 1649 1650 1651 1652 1653 1654 1655
        PADDLE_ENFORCE(tmp == *data_type || *data_type == default_data_type,
                       platform::errors::InvalidArgument(
                           "The DataType of %s Op's duplicable or different "
                           "slot Variable %s must be "
                           "consistent or reigster GetExpectedKernelType. The "
                           "current variable type is (%s), but the "
                           "previous variable type is (%s).",
                           Type(), name, DataTypeToString(tmp),
                           DataTypeToString(*data_type)));
1656 1657 1658 1659 1660 1661
        *data_type = tmp;
      }
    }
  }
}

1662
proto::VarType::Type OperatorWithKernel::IndicateDataType(
Y
Yu Yang 已提交
1663
    const ExecutionContext& ctx) const {
1664 1665 1666
  proto::VarType::Type dafault_data_type =
      static_cast<proto::VarType::Type>(-1);
  proto::VarType::Type data_type = dafault_data_type;
H
hong 已提交
1667
  for (auto& input : ctx.InNameList()) {
1668 1669
    const std::vector<Variable*> vars = ctx.MultiInputVar(input);
    ParseInputDataType(vars, input, &data_type);
Y
Yu Yang 已提交
1670
  }
1671 1672 1673 1674
  PADDLE_ENFORCE_NE(
      data_type, dafault_data_type,
      platform::errors::NotFound(
          "DataType should be indicated by input Variable at %s.", Type()));
1675 1676 1677 1678 1679 1680 1681 1682
  return data_type;
}

proto::VarType::Type OperatorWithKernel::IndicateVarDataType(
    const ExecutionContext& ctx, const std::string& name) const {
  proto::VarType::Type dafault_data_type =
      static_cast<proto::VarType::Type>(-1);
  proto::VarType::Type data_type = dafault_data_type;
1683
  ParseInputDataType(ctx.MultiInputVar(name), name, &data_type);
1684 1685
  PADDLE_ENFORCE_NE(
      data_type, dafault_data_type,
1686 1687 1688 1689 1690
      platform::errors::InvalidArgument(
          "The Input Variable(%s) of (%s) Operator used to determine kernel "
          "data type is empty or not LoDTensor or SelectedRows or "
          "LoDTensorArray.",
          name, Type()));
1691
  return data_type;
Y
Yu Yang 已提交
1692
}
1693

1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753
Tensor* OperatorWithKernel::GetTensorFormInputSafely(
    const ExecutionContext& ctx, const std::string& name) const {
  // 1. get variable and check
  // NOTE: only supports signal input var now
  // NOTE: using const_cast is because we don't have method
  // can get single mutable var, and here will not change
  // the var's data, only use some attribute
  Variable* var = const_cast<Variable*>(ctx.InputVar(name));
  PADDLE_ENFORCE_NOT_NULL(
      var,
      platform::errors::NotFound(
          "The variable %s is not found when promote complex types.", name));
  // 2. get tensor and check
  Tensor* t = nullptr;
  if (var->IsType<Tensor>()) {
    t = var->GetMutable<Tensor>();
  } else if (var->IsType<LoDTensor>()) {
    t = var->GetMutable<LoDTensor>();
  } else if (var->IsType<SelectedRows>()) {
    t = var->GetMutable<SelectedRows>()->mutable_value();
  } else {
    PADDLE_THROW(platform::errors::Unimplemented(
        "Unsupported input variable type in complex type promotion."));
  }
  PADDLE_ENFORCE_NOT_NULL(
      t,
      platform::errors::InvalidArgument(
          "The Tensor of variable %s is nullptr when promote complex types."));
  PADDLE_ENFORCE_EQ(t->IsInitialized(), true,
                    platform::errors::InvalidArgument(
                        "The Tensor in the %s Op's Input Variable %s(%s) is "
                        "not initialized.",
                        Type(), name, ctx.InputName(name)));
  return t;
}

/** NOTE(chenweihang): For safety reasons, we now only
 * perform type promotes for binary operations with
 * complex type inputs, which is used to support the
 * paddle quantum function.
 * In other cases, the first input data type is used as
 * the kernel data type.
 */
proto::VarType::Type OperatorWithKernel::IndicateOrPromoteVarDataTypes(
    const ExecutionContext& ctx, const std::string& name1,
    const std::string& name2) const {
  // 1. Get tensor
  auto* tensor_a = GetTensorFormInputSafely(ctx, name1);
  auto* tensor_b = GetTensorFormInputSafely(ctx, name2);

  // 2. Get two input types
  auto type_a = tensor_a->type();
  auto type_b = tensor_b->type();

  // 3. Get first input type or promote complex types
  auto target_type = PromoteTypesIfComplexExists(type_a, type_b);

  return target_type;
}

1754 1755 1756 1757 1758 1759 1760 1761
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
1762 1763
  return OpKernelType(expected_kernel_type.data_type_, tensor.place(),
                      tensor.layout());
1764 1765
}

1766 1767 1768 1769 1770
KernelSignature OperatorWithKernel::GetExpectedPtenKernelArgs(
    const ExecutionContext& ctx) const {
  return KernelSignatureMap::Instance().Get(Type());
}

1771 1772
void OperatorWithKernel::BuildPtenKernelContext(
    const RuntimeContext& ctx, platform::DeviceContext* dev_ctx) const {
1773 1774 1775 1776 1777 1778 1779
  // TODO(chenweihang): now only work for very simple case,
  // many cases need to be deal with later:
  // 1. the input and output are not tensor
  // 2. the dispensbale, duplicable input and output
  // 3. needless attributes remove
  // 4. use pt Tensor directly
  // 5. kernel input is not DenseTensor
1780
  pt_kernel_context_->SetDeviceContext(dev_ctx);
1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808

  auto& input_names = std::get<0>(pt_kernel_signature_->args);
  auto& attr_names = std::get<1>(pt_kernel_signature_->args);
  auto& output_names = std::get<2>(pt_kernel_signature_->args);

  auto input_defs = pt_kernel_->args_def().input_defs();
  auto attr_defs = pt_kernel_->args_def().attribute_defs();
  auto output_defs = pt_kernel_->args_def().output_defs();

  PADDLE_ENFORCE_EQ(input_names.size(), input_defs.size(),
                    platform::errors::InvalidArgument(
                        "The size of inputs_args names (%d) must be equal to "
                        "the size of kernel input_defs (%d).",
                        input_names.size(), input_defs.size()));

  PADDLE_ENFORCE_EQ(output_names.size(), output_defs.size(),
                    platform::errors::InvalidArgument(
                        "The size of outputs_args names (%d) must be equal to "
                        "the size of kernel output_defs (%d).",
                        output_names.size(), output_defs.size()));

  PADDLE_ENFORCE_EQ(attr_names.size(), attr_defs.size(),
                    platform::errors::InvalidArgument(
                        "The size of attribute_args names (%d) must be equal "
                        "to the size of kernel attribute_defs (%d).",
                        attr_names.size(), attr_defs.size()));

  for (size_t i = 0; i < input_names.size(); ++i) {
1809 1810
    auto& in_def = input_defs.at(i);
    auto& ins_vector = ctx.inputs.at(input_names[i]);
1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822

    // calcute the start and end index of the input tensors
    size_t start_idx =
        (i == 0 ? 0 : pt_kernel_context_->InputRangeAt(i - 1).second);
    size_t end_idx = start_idx + ins_vector.size();

    // The current size of input/output in pt_kernel_context_ is at least equal
    // the start_idx. For the reason of reusing the allocted of inputs or
    // outputs in pt_kernel_context_, the current size of input/output can be
    // greater then the index of which the tensort wanted to set to, so it will
    // use ReMakePtenDenseTensorFromVar to make pten tensor.
    if (pt_kernel_context_->InputsSize() == start_idx) {
1823 1824 1825 1826 1827 1828
      paddle::SmallVector<std::shared_ptr<pten::TensorBase>> tmp_inputs;
      for (auto* var : ins_vector) {
        tmp_inputs.emplace_back(
            experimental::MakePtenTensorBaseFromVar(*var, in_def));
      }
      pt_kernel_context_->EmplaceBackInputs(std::move(tmp_inputs));
1829
    } else if (pt_kernel_context_->InputsSize() > start_idx) {
1830 1831
      size_t input_size = pt_kernel_context_->InputsSize();
      for (size_t j = 0; j < ins_vector.size(); ++j) {
1832
        if (input_size > start_idx + j) {
1833 1834
          experimental::ReMakePtenDenseTensorFromVar(
              *ins_vector[j], in_def,
1835 1836 1837 1838 1839 1840 1841 1842 1843
              pt_kernel_context_->MutableInputAt<pten::DenseTensor>(start_idx +
                                                                    j));
          // TODO(chentianyu03): When multi input kernel, open this code
          /*
          } else {
            pt_kernel_context_->EmplaceBackInputWithoutSetRange(
                experimental::MakePtenTensorBaseFromVar(*ins_vector[j],
          in_def));
          */
1844 1845 1846
        }
      }
      pt_kernel_context_->MutableInputRangeAt(i) =
1847 1848 1849 1850 1851 1852 1853
          std::make_pair(start_idx, end_idx);
    } else {
      PADDLE_THROW(platform::errors::PreconditionNotMet(
          "Error start index when trying to set new tensor to inputs, start "
          "index is `%d`, but current pt_kernel_context_.inputs.size() is "
          "`%d`.",
          start_idx, pt_kernel_context_->InputsSize()));
1854 1855 1856 1857
    }
  }

  for (size_t i = 0; i < output_names.size(); ++i) {
1858 1859
    auto& out_def = output_defs.at(i);
    auto& outs_vector = ctx.outputs.at(output_names[i]);
1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870

    size_t start_idx =
        (i == 0 ? 0 : pt_kernel_context_->OutputRangeAt(i - 1).second);
    size_t end_idx = start_idx + outs_vector.size();

    // The current size of input/output in pt_kernel_context_ is at least equal
    // the start_idx. For the reason of reusing the allocted of inputs or
    // outputs in pt_kernel_context_, the current size of input/output can be
    // greater then the index of which the tensort wanted to set to, so it will
    // use ReMakePtenDenseTensorFromVar to make pten tensor.
    if (pt_kernel_context_->OutputsSize() == start_idx) {
1871 1872 1873 1874 1875 1876
      paddle::SmallVector<std::shared_ptr<pten::TensorBase>> tmp_outputs;
      for (auto* var : outs_vector) {
        tmp_outputs.emplace_back(
            experimental::MakePtenTensorBaseFromVar(var, out_def));
      }
      pt_kernel_context_->EmplaceBackOutputs(std::move(tmp_outputs));
1877
    } else if (pt_kernel_context_->OutputsSize() > start_idx) {
1878 1879
      size_t output_size = pt_kernel_context_->OutputsSize();
      for (size_t j = 0; j < outs_vector.size(); ++j) {
1880
        if (output_size > start_idx + j) {
1881 1882
          experimental::ReMakePtenDenseTensorFromVar(
              outs_vector[j], out_def,
1883 1884 1885 1886 1887 1888 1889 1890 1891 1892
              pt_kernel_context_->MutableOutputAt<pten::DenseTensor>(start_idx +
                                                                     j));

          // TODO(chentianyu03): When multi output kernel, open this code
          /*
          } else {
            pt_kernel_context_->EmplaceBackOutputWithoutSetRange(
                experimental::MakePtenTensorBaseFromVar(outs_vector[j],
          out_def));
              */
1893 1894 1895
        }
      }
      pt_kernel_context_->MutableOutputRangeAt(i) =
1896 1897 1898 1899 1900 1901 1902
          std::make_pair(start_idx, end_idx);
    } else {
      PADDLE_THROW(platform::errors::PreconditionNotMet(
          "Error start index when trying to set new tensor to inputs, start "
          "index is `%d`, but current pt_kernel_context_.outputs.size() is "
          "`%d`.",
          start_idx, pt_kernel_context_->OutputsSize()));
1903 1904 1905 1906
    }
  }

  for (size_t i = 0; i < attr_names.size(); ++i) {
1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931
    if (attr_defs[i].type_index == std::type_index(typeid(pten::ScalarArray))) {
      auto attr_iter = Attrs().find(attr_names[i]);
      if (attr_iter != Attrs().end()) {  // shape is in the attribute
        if (std::type_index(attr_iter->second.type()) ==
            std::type_index(typeid(std::vector<int64_t>))) {
          pt_kernel_context_->EmplaceBackAttr(std::move(pten::ScalarArray(
              BOOST_GET_CONST(std::vector<int64_t>, attr_iter->second))));
        } else {
          PADDLE_THROW(platform::errors::Unimplemented(
              "Unsupported cast op attribute `%s` to ScalarArray when "
              "construct KernelContext.",
              attr_names[i]));
        }
      } else {  // shape is in the input
        auto& ins_vector = ctx.inputs.at(attr_names[i]);
        if (ins_vector.size() == 1) {  // ShapeTensor
          pt_kernel_context_->EmplaceBackAttr(std::move(
              experimental::MakePtenScalarArrayFromVar(*ins_vector.front())));
        } else {  // ShapeTensorList
          pt_kernel_context_->EmplaceBackAttr(std::move(
              experimental::MakePtenScalarArrayFromVarList(ins_vector)));
        }
      }
    } else if (attr_defs[i].type_index ==
               std::type_index(typeid(pten::Scalar))) {
1932 1933 1934
      // TODO(chenweihang): support other attrs later
      // TODO(zhangyunfei): Scalar should hold scaler type, and we should check
      // attribtue type by attr_defs
1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950
      auto attr_iter = Attrs().find(attr_names[i]);
      if (attr_iter != Attrs().end()) {  // scalar is in the attribute
        auto& attr = Attrs().at(attr_names[i]);
        if (std::type_index(attr.type()) == std::type_index(typeid(float))) {
          pt_kernel_context_->EmplaceBackAttr(
              std::move(pten::Scalar(BOOST_GET_CONST(float, attr))));
        } else if (std::type_index(attr.type()) ==
                   std::type_index(typeid(std::string))) {
          pt_kernel_context_->EmplaceBackAttr(
              std::move(pten::Scalar(BOOST_GET_CONST(std::string, attr))));
        } else {
          PADDLE_THROW(platform::errors::Unimplemented(
              "Unsupported cast op attribute `%s` to Scalar when construct "
              "KernelContext.",
              attr_names[i]));
        }
1951
      } else {
1952 1953 1954
        auto& ins_vector = ctx.inputs.at(attr_names[i]);
        pt_kernel_context_->EmplaceBackAttr(std::move(
            experimental::MakePtenScalarFromVar(*ins_vector.front())));
1955
      }
1956

1957 1958
    } else {
      // TODO(chenweihang): support other attrs later
1959
      auto& attr = Attrs().at(attr_names[i]);
1960
      if (attr_defs[i].type_index == std::type_index(typeid(int))) {
1961
        pt_kernel_context_->EmplaceBackAttr(BOOST_GET_CONST(int, attr));
1962
      } else if (attr_defs[i].type_index == std::type_index(typeid(float))) {
1963
        pt_kernel_context_->EmplaceBackAttr(BOOST_GET_CONST(float, attr));
1964
      } else if (attr_defs[i].type_index == std::type_index(typeid(bool))) {
1965
        pt_kernel_context_->EmplaceBackAttr(BOOST_GET_CONST(bool, attr));
1966
      } else if (attr_defs[i].type_index ==
1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983
                 std::type_index(typeid(pten::DataType))) {
        auto data_type = pten::TransToPtenDataType(
            static_cast<framework::proto::VarType::Type>(
                BOOST_GET_CONST(int, attr)));
        pt_kernel_context_->EmplaceBackAttr(data_type);
      } else if (attr_defs[i].type_index ==
                 std::type_index(typeid(std::vector<int64_t>))) {
        if (std::type_index(attr.type()) ==
            std::type_index(typeid(std::vector<int>))) {
          // Emplace Back Attr according to the type of Pten_Kernel args.
          const auto& vector_int_attr = BOOST_GET_CONST(std::vector<int>, attr);
          const std::vector<int64_t> vector_int64_attr(vector_int_attr.begin(),
                                                       vector_int_attr.end());
          pt_kernel_context_->EmplaceBackAttr(vector_int64_attr);
        }
        // TODO(YuanRisheng) Need support vector<int64_t> attr

1984 1985
      } else {
        PADDLE_THROW(platform::errors::Unimplemented(
1986
            "Unsupported cast op attribute `%s` when construct "
1987 1988 1989 1990 1991 1992 1993
            "KernelContext.",
            attr_names[i]));
      }
    }
  }
}

1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
void OperatorWithKernel::WriteBackToOutputs(RuntimeContext* ctx) const {
  // auto& input_names = std::get<0>(pt_kernel_signature_->args);
  // auto& attr_names = std::get<1>(pt_kernel_signature_->args);
  auto& output_names = std::get<2>(pt_kernel_signature_->args);

  // pt_kernel_context_

  for (size_t i = 0; i < output_names.size(); ++i) {
    auto& outs_vector = ctx->outputs.at(output_names[i]);

    auto& range_pair = pt_kernel_context_->OutputRangeAt(i);
    auto pten_outs =
        pt_kernel_context_->MutableOutputBetween<pten::DenseTensor>(
            range_pair.first, range_pair.second);

    for (size_t j = 0; j < pten_outs.size(); ++j) {
      experimental::MakeVariableFromPtenTensor(pten_outs[j], outs_vector[j]);
    }
  }
}

Q
Qiao Longfei 已提交
2015
}  // namespace framework
L
liaogang 已提交
2016
}  // namespace paddle