operator.cc 18.6 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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
#include <gflags/gflags.h>
D
dzhwinter 已提交
15
#include <glog/logging.h>
Q
Qiao Longfei 已提交
16

17
#include <algorithm>
D
dzhwinter 已提交
18

19
#include "paddle/framework/data_transform.h"
D
dzhwinter 已提交
20 21
#include "paddle/framework/executor.h"
#include "paddle/framework/operator.h"
22
#include "paddle/framework/shape_inference.h"
23
#include "paddle/framework/var_type.h"
Q
Qiao Longfei 已提交
24

D
dzhwinter 已提交
25 26 27 28
DEFINE_bool(op_sync, false,
            "Default cuda is asynchronous device, set to True will"
            "force op run in synchronous mode.");

Q
Qiao Longfei 已提交
29 30 31
namespace paddle {
namespace framework {

32 33 34 35 36 37
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
38

39 40
static DDim GetDims(const Scope& scope, const std::string& name) {
  Variable* var = scope.FindVar(name);
Q
qiaolongfei 已提交
41 42
  if (var == nullptr) {
    return DDim({-1});
Q
Qiao Longfei 已提交
43 44 45
  }

  if (var->IsType<LoDTensor>()) {
46 47 48 49 50 51 52 53
    return var->Get<LoDTensor>().dims();
  } else if (var->IsType<SelectedRows>()) {
    return var->Get<SelectedRows>().GetCompleteDims();
  } else {
    return DDim({-1});
  }
}

Q
Qiao Longfei 已提交
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
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;
  }

  if (var->IsType<LoDTensor>()) {
    return var->Get<LoDTensor>().lod();
  } else {
    return default_lod;
  }
}

69
std::string OperatorBase::Input(const std::string& name) const {
Y
Yu Yang 已提交
70
  auto& ins = Inputs(name);
Y
Yu Yang 已提交
71
  PADDLE_ENFORCE_LE(ins.size(), 1UL,
72 73
                    "Operator %s's input %s should contain only one variable.",
                    type_, name);
Y
Yu Yang 已提交
74
  return ins.empty() ? kEmptyVarName : ins[0];
Y
Yan Chunwei 已提交
75 76
}

Y
Yu Yang 已提交
77 78
const std::vector<std::string>& OperatorBase::Inputs(
    const std::string& name) const {
Y
Yu Yang 已提交
79
  auto it = inputs_.find(name);
80 81
  PADDLE_ENFORCE(it != inputs_.end(), "Operator %s does not have the input %s.",
                 type_, name);
Y
Yu Yang 已提交
82
  return it->second;
Y
Yan Chunwei 已提交
83 84
}

85
std::string OperatorBase::Output(const std::string& name) const {
Y
Yu Yang 已提交
86
  auto& outs = Outputs(name);
Y
Yu Yang 已提交
87
  PADDLE_ENFORCE_LE(outs.size(), 1UL,
88 89
                    "Operator %s's output %s should contain only one variable.",
                    type_, name);
Y
Yu Yang 已提交
90
  return outs.empty() ? kEmptyVarName : outs[0];
Y
Yan Chunwei 已提交
91 92
}

Y
Yu Yang 已提交
93 94
const std::vector<std::string>& OperatorBase::Outputs(
    const std::string& name) const {
Y
Yu Yang 已提交
95
  auto it = outputs_.find(name);
96 97
  PADDLE_ENFORCE(it != outputs_.end(),
                 "Operator %s does not have an output called %s.", type_, name);
Y
Yu Yang 已提交
98
  return it->second;
Y
Yan Chunwei 已提交
99 100
}

101
std::string OperatorBase::DebugStringEx(const Scope* scope) const {
Q
Qiao Longfei 已提交
102
  std::stringstream ss;
Y
Yu Yang 已提交
103
  ss << "Op(" << type_ << "), inputs:{";
Y
Yu Yang 已提交
104 105
  for (auto it = inputs_.begin(); it != inputs_.end();) {
    auto& input = *it;
Y
Yu Yang 已提交
106 107 108
    ss << input.first << "[";
    for (size_t i = 0; i < input.second.size(); ++i) {
      ss << input.second[i];
109
      if (scope) {
Q
Qiao Longfei 已提交
110 111
        ss << "[" << GetDims(*scope, input.second[i]) << "]";
        ss << "(" << GetLoD(*scope, input.second[i]) << ")";
112
      }
Y
Yu Yang 已提交
113 114 115
      if (i != input.second.size() - 1) {
        ss << ", ";
      }
116
    }
Y
Yu Yang 已提交
117
    ss << "]";
Y
Yu Yang 已提交
118 119
    ++it;
    if (it != inputs_.end()) {
120 121
      ss << ", ";
    }
Q
Qiao Longfei 已提交
122
  }
Y
Yu Yang 已提交
123
  ss << "}, outputs:{";
Y
Yu Yang 已提交
124 125
  for (auto it = outputs_.begin(); it != outputs_.end();) {
    auto& output = *it;
Y
Yu Yang 已提交
126 127 128
    ss << output.first << "[";
    for (size_t i = 0; i < output.second.size(); ++i) {
      ss << output.second[i];
129
      if (scope) {
Q
Qiao Longfei 已提交
130 131
        ss << "[" << GetDims(*scope, output.second[i]) << "]";
        ss << "(" << GetLoD(*scope, output.second[i]) << ")";
132
      }
Y
Yu Yang 已提交
133 134 135
      if (i != output.second.size() - 1) {
        ss << ", ";
      }
136
    }
Y
Yu Yang 已提交
137
    ss << "]";
Y
Yu Yang 已提交
138 139
    ++it;
    if (it != outputs_.end()) {
140 141
      ss << ", ";
    }
Q
Qiao Longfei 已提交
142
  }
Y
Yu Yang 已提交
143
  ss << "}.";
Q
Qiao Longfei 已提交
144 145 146
  return ss.str();
}

D
dongzhihong 已提交
147 148
void OperatorBase::Rename(const std::string& old_name,
                          const std::string& new_name) {
Y
Yu Yang 已提交
149 150 151 152 153 154 155
  for (auto& input : inputs_) {
    std::replace(input.second.begin(), input.second.end(), old_name, new_name);
  }
  for (auto& output : outputs_) {
    std::replace(output.second.begin(), output.second.end(), old_name,
                 new_name);
  }
D
dongzhihong 已提交
156 157
}

Y
Yu Yang 已提交
158
OperatorBase::OperatorBase(const std::string& type,
Y
Yu Yang 已提交
159 160
                           const VariableNameMap& inputs,
                           const VariableNameMap& outputs,
Y
Yu Yang 已提交
161 162
                           const AttributeMap& attrs)
    : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
163 164
  GenerateTemporaryNames();
  CheckAllInputOutputSet();
Y
Yu Yang 已提交
165
}
166

Q
qijun 已提交
167 168
std::vector<std::string> OperatorBase::InputVars() const {
  std::vector<std::string> ret_val;
Y
Yu Yang 已提交
169
  for (auto& o : inputs_) {
Q
qijun 已提交
170 171 172 173 174 175
    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
176 177 178 179 180 181 182 183 184 185
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
186
  auto& info = OpInfoMap::Instance().Get(Type());
Y
Yu Yang 已提交
187 188

  // get all OpProto::Var for outputs
Y
Yu Yang 已提交
189
  for (auto& o : info.Proto().outputs()) {
Y
Yu Yang 已提交
190 191 192 193 194 195 196 197 198
    // 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
199 200
}

201 202 203
void OperatorBase::CheckAllInputOutputSet() const {
  auto& info_map = OpInfoMap::Instance();
  auto* op_info = info_map.GetNullable(Type());
Y
Yu Yang 已提交
204
  if (op_info == nullptr || op_info->proto_ == nullptr) return;
205 206 207

  for (auto& in : op_info->Proto().inputs()) {
    PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(),
Y
Yu Yang 已提交
208
                   "Type %s's input %s is not set", Type(), in.name());
209 210 211 212
  }

  for (auto& out : op_info->Proto().outputs()) {
    PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(),
Y
Yu Yang 已提交
213
                   "Type %s's output %s is not set", Type(), out.name());
214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229
  }
}

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));
      }
    }
  }
}

230 231 232 233
static bool VarIsTensor(const Variable* var) {
  return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
}

234
static const Tensor* GetTensorFromVar(Variable* var) {
Q
QI JUN 已提交
235
  if (var->IsType<LoDTensor>()) {
236
    return var->GetMutable<LoDTensor>();
Q
QI JUN 已提交
237
  } else if (var->IsType<SelectedRows>()) {
238
    return var->GetMutable<SelectedRows>()->mutable_value();
Q
QI JUN 已提交
239
  } else {
Y
Yang Yang 已提交
240 241
    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
                 var->Type().name());
Q
QI JUN 已提交
242 243 244 245 246
  }
}

static Tensor* GetMutableTensorFromVar(Variable* var) {
  if (var->IsType<LoDTensor>()) {
247
    return var->GetMutable<LoDTensor>();
Q
QI JUN 已提交
248
  } else if (var->IsType<SelectedRows>()) {
249
    return var->GetMutable<SelectedRows>()->mutable_value();
Q
QI JUN 已提交
250
  } else {
Y
Yang Yang 已提交
251 252
    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
                 var->Type().name());
Q
QI JUN 已提交
253 254 255
  }
}

256
template <>
257
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
258
  auto* var = InputVar(name);
259 260
  return var == nullptr ? nullptr
                        : GetTensorFromVar(const_cast<Variable*>(var));
261 262 263
}

template <>
264
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
265 266 267 268
    const std::string& name) const {
  auto names = op().Inputs(name);
  std::vector<const Tensor*> res;
  res.reserve(names.size());
269 270 271 272 273
  std::transform(names.begin(), names.end(), std::back_inserter(res),
                 [&](const std::string& sub_name) {
                   auto var = scope_.FindVar(sub_name);
                   return var == nullptr ? nullptr : GetTensorFromVar(var);
                 });
274 275 276 277
  return res;
}

template <>
278
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
279
  auto var = OutputVar(name);
Q
QI JUN 已提交
280
  return var == nullptr ? nullptr : GetMutableTensorFromVar(var);
281 282 283
}

template <>
284
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
285 286 287 288
    const std::string& name) const {
  auto names = op().Outputs(name);
  std::vector<Tensor*> res;
  res.reserve(names.size());
289 290
  std::transform(names.begin(), names.end(), std::back_inserter(res),
                 [&](const std::string& sub_name) {
291 292
                   auto var = scope_.FindVar(sub_name);
                   return var == nullptr ? nullptr
Q
QI JUN 已提交
293
                                         : GetMutableTensorFromVar(var);
294
                 });
295 296 297
  return res;
}

Y
Yu Yang 已提交
298 299 300 301 302
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
303

Y
Yu Yang 已提交
304 305 306 307 308 309 310 311 312 313
    return true;
  }
  for (auto& kern_pair : it->second) {
    if (platform::is_gpu_place(kern_pair.first.place_)) {
      return true;
    }
  }
  return false;
}

314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390
class RuntimeInferShapeContext : public InferShapeContext {
 public:
  RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope)
      : op_(op), scope_(scope) {}

  bool HasInput(const std::string& name) const override {
    auto& ins = Inputs(name);
    size_t length = ins.size();
    if (length == 0) {
      return false;
    }
    PADDLE_ENFORCE_EQ(length, 1UL, "Input %s should have more than one inputs",
                      name);
    auto ipt = ins[0];
    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
    return var != nullptr;
  }

  bool HasOutput(const std::string& name) const override {
    auto& outs = Outputs(name);
    size_t length = outs.size();
    if (length == 0) {
      return false;
    }
    PADDLE_ENFORCE_EQ(length, 1UL, "Output %s should have more than one inputs",
                      name);
    auto ipt = outs[0];
    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
    return var != nullptr;
  }

  bool HasInputs(const std::string& name) const override {
    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 {
    auto outputs = op_.Outputs(name);
    if (outputs.empty()) {
      return false;
    }
    for (auto& output : outputs) {
      if (scope_.FindVar(output) == nullptr) {
        return false;
      }
    }
    return true;
  }

  DDim GetInputDim(const std::string& name) const override {
    return GetDim(op_.Input(name));
  }

  void SetOutputDim(const std::string& name, const DDim& dim) override {
    SetDim(op_.Output(name), dim);
  }

  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);
  }

Q
Qiao Longfei 已提交
391 392 393 394 395 396 397 398 399 400 401 402
  void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
                size_t j = 0) const override {
    PADDLE_ENFORCE_LT(i, Inputs(in).size());
    PADDLE_ENFORCE_LT(j, Outputs(out).size());
    Variable* in_var = scope_.FindVar(Inputs(in)[i]);
    Variable* out_var = scope_.FindVar(Outputs(out)[j]);
    if (!in_var->IsType<LoDTensor>()) return;
    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
403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421

    // 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?
    out_tensor->set_layout(in_tensor.layout());
  }

  void ShareLayout(const std::string& in, const std::string& out, size_t i = 0,
                   size_t j = 0) const {
    PADDLE_ENFORCE_LT(i, Inputs(in).size());
    PADDLE_ENFORCE_LT(j, Outputs(out).size());
    Variable* in_var = scope_.FindVar(Inputs(in)[i]);
    Variable* out_var = scope_.FindVar(Outputs(out)[j]);
    if (!in_var->IsType<LoDTensor>()) return;
    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_layout(in_tensor.layout());
Q
Qiao Longfei 已提交
422 423
  }

424 425 426
  bool IsRuntime() const override { return true; }

 protected:
427 428 429 430 431 432 433
  DDim GetDim(const std::string& name) const override {
    Variable* var = scope_.FindVar(name);
    if (var->IsType<LoDTensor>()) {
      return var->Get<LoDTensor>().dims();
    } else if (var->IsType<SelectedRows>()) {
      return var->Get<SelectedRows>().GetCompleteDims();
    } else {
Y
Yang Yang 已提交
434 435
      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
                   name, var->Type().name());
436 437 438 439 440 441 442 443 444 445
    }
  }

  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
446 447
      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
                   name, var->Type().name());
448 449 450
    }
  }

451
  proto::VarDesc::VarType GetVarType(const std::string& name) const override {
452 453 454 455 456
    auto* var = scope_.FindVar(name);
    return ToVarType(var->Type());
  }

 private:
457 458 459 460 461
  const OperatorBase& op_;
  const Scope& scope_;
};

void OperatorWithKernel::Run(const Scope& scope,
D
dzhwinter 已提交
462
                             const platform::Place& place) const {
463 464
  RuntimeInferShapeContext infer_shape_ctx(*this, scope);
  this->InferShape(&infer_shape_ctx);
Y
Yu Yang 已提交
465 466
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
  auto dev_ctx = pool.Get(place);
467 468 469 470 471

  // 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
472 473
    PADDLE_THROW(
        "There are no kernels which are registered in the %s operator.", type_);
474 475
  }

D
dzhwinter 已提交
476
  ExecutionContext ctx(*this, scope, *dev_ctx);
477

Q
qiaolongfei 已提交
478 479
  OpKernelMap& kernels = kernels_iter->second;

480 481
  // TODO(dzhwinter) : kernel fallback mechanism will be added when all the
  // transform functions are ready.
Q
qiaolongfei 已提交
482

483 484 485 486 487
  // for (auto& candidate : kKernelPriority) {
  //   Do selection
  // }

  auto expected_kernel_key = this->GetExpectedKernelType(ctx);
Q
qiaolongfei 已提交
488 489
  VLOG(3) << "expected_kernel_key:" << expected_kernel_key;

490 491 492 493 494 495 496
  auto kernel_iter = kernels.find(expected_kernel_key);
  if (kernel_iter == kernels.end()) {
    PADDLE_THROW("op %s does not have kernel for %s", type_,
                 KernelTypeToString(expected_kernel_key));
  }

  // do data transform
497 498 499 500 501 502 503 504 505 506
  Scope& new_scope = scope.NewScope();

  for (auto& var_name_item : this->Inputs()) {
    for (auto& var_name : var_name_item.second) {
      auto* var = scope.FindVar(var_name);
      if (var && VarIsTensor(var)) {
        auto* tensor_in = GetTensorFromVar(var);
        if (tensor_in->IsInitialized()) {
          auto kernel_type_for_var = this->GetKernelTypeForVar(
              var_name_item.first, *tensor_in, expected_kernel_key);
507
          if (TransFromNeeded(kernel_type_for_var, expected_kernel_key)) {
508 509 510 511 512 513 514 515
            auto out_var_names = OutputVars(true);
            if (std::find(out_var_names.begin(), out_var_names.end(),
                          var_name) != out_var_names.end()) {
              PADDLE_THROW(
                  "var %s is both input and output, "
                  "does not support transform",
                  var_name);
            }
516 517
            VLOG(3) << "Transform Variable " << var_name << " from "
                    << kernel_type_for_var << " to " << expected_kernel_key;
518
            auto* trans_var = new_scope.Var(var_name);
519 520 521 522
            std::shared_ptr<Tensor> out(new Tensor);
            DataTransform(expected_kernel_key, kernel_type_for_var, *tensor_in,
                          out.get());
            CopyVariableWithTensor(*var, *(out.get()), *trans_var);
523
          }
Q
QI JUN 已提交
524 525
        }
      }
526 527
    }
  }
Q
QI JUN 已提交
528

D
dzhwinter 已提交
529 530 531 532 533 534 535 536
  auto* new_dev_ctx = pool.Get(expected_kernel_key.place_);
  kernel_iter->second->Compute(
      ExecutionContext(*this, new_scope, *new_dev_ctx));

  /*For profiling/benchmark only*/
  if (FLAGS_op_sync) {
    new_dev_ctx->Wait();
  }
Q
Qiao Longfei 已提交
537 538
}

539
proto::DataType OperatorWithKernel::IndicateDataType(
Y
Yu Yang 已提交
540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564
    const ExecutionContext& ctx) const {
  auto& scope = ctx.scope();
  int data_type = -1;
  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) {
          int tmp = static_cast<int>(ToDataType(t->type()));
          PADDLE_ENFORCE(tmp == data_type || data_type == -1,
                         "DataType of Paddle Op %s must be the same.", Type());
          data_type = tmp;
        }
      }
    }
  }
  PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
565
  return static_cast<proto::DataType>(data_type);
Y
Yu Yang 已提交
566
}
567

568 569 570 571 572 573 574 575 576 577 578
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 {
  return OpKernelType(expected_kernel_type.data_type_, tensor.place());
}

Q
Qiao Longfei 已提交
579
}  // namespace framework
L
liaogang 已提交
580
}  // namespace paddle