operator.cc 25.2 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
#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

Y
Yi Wang 已提交
19 20 21 22 23
#include "paddle/fluid/framework/data_transform.h"
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/shape_inference.h"
#include "paddle/fluid/framework/var_type.h"
24
#include "paddle/fluid/platform/profiler.h"
Q
Qiao Longfei 已提交
25

D
dzhwinter 已提交
26
DECLARE_bool(benchmark);
C
chengduoZH 已提交
27 28 29
DEFINE_bool(check_nan_inf, false,
            "Checking whether operator produce NAN/INF or not. It will be "
            "extremely slow so please use this flag wisely.");
D
dzhwinter 已提交
30

Q
Qiao Longfei 已提交
31 32 33
namespace paddle {
namespace framework {

34 35 36 37 38 39
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
40

Q
qiaolongfei 已提交
41 42 43 44 45 46 47 48 49 50 51
proto::VarType::Type GetDataTypeOfVar(const Variable* var) {
  if (var->IsType<framework::LoDTensor>()) {
    return framework::ToDataType(var->Get<framework::LoDTensor>().type());
  } else if (var->IsType<framework::SelectedRows>()) {
    return framework::ToDataType(
        var->Get<framework::SelectedRows>().value().type());
  } else {
    PADDLE_THROW("Var should be LoDTensor or SelectedRows");
  }
}

52 53
static DDim GetDims(const Scope& scope, const std::string& name,
                    bool get_actual_dim = false) {
54
  Variable* var = scope.FindVar(name);
Q
qiaolongfei 已提交
55 56
  if (var == nullptr) {
    return DDim({-1});
Q
Qiao Longfei 已提交
57 58 59
  }

  if (var->IsType<LoDTensor>()) {
60 61
    return var->Get<LoDTensor>().dims();
  } else if (var->IsType<SelectedRows>()) {
62 63 64 65 66
    if (get_actual_dim) {
      return var->Get<SelectedRows>().value().dims();
    } else {
      return var->Get<SelectedRows>().GetCompleteDims();
    }
67 68 69 70 71
  } else {
    return DDim({-1});
  }
}

72 73 74 75 76 77 78 79 80 81 82 83 84
static int GetRowSize(const Scope& scope, const std::string& name) {
  Variable* var = scope.FindVar(name);
  if (var == nullptr) {
    return -1;
  }

  if (var->IsType<SelectedRows>()) {
    return var->Get<SelectedRows>().rows().size();
  }

  return -1;
}

Q
Qiao Longfei 已提交
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99
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;
  }
}

100
void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
101
  VLOG(10) << "- " << DebugStringEx(&scope);
102 103 104 105 106 107 108 109 110
  if (platform::is_gpu_place(place)) {
#ifndef PADDLE_WITH_CUDA
    PADDLE_THROW("Cannot run operator on place %s", place);
#else
    auto dev_id = boost::get<platform::CUDAPlace>(place).device;
    platform::SetDeviceId(dev_id);
#endif
  }
  RunImpl(scope, place);
111
  VLOG(10) << "+ " << DebugStringEx(&scope);
112 113
}

114 115 116 117 118 119 120 121
bool OperatorBase::HasInputs(const std::string& name) const {
  if (inputs_.find(name) != inputs_.end()) {
    return true;
  } else {
    return false;
  }
}

122
std::string OperatorBase::Input(const std::string& name) const {
Y
Yu Yang 已提交
123
  auto& ins = Inputs(name);
Y
Yu Yang 已提交
124
  PADDLE_ENFORCE_LE(ins.size(), 1UL,
125 126
                    "Operator %s's input %s should contain only one variable.",
                    type_, name);
Y
Yu Yang 已提交
127
  return ins.empty() ? kEmptyVarName : ins[0];
Y
Yan Chunwei 已提交
128 129
}

Y
Yu Yang 已提交
130 131
const std::vector<std::string>& OperatorBase::Inputs(
    const std::string& name) const {
Y
Yu Yang 已提交
132
  auto it = inputs_.find(name);
133 134
  PADDLE_ENFORCE(it != inputs_.end(), "Operator %s does not have the input %s.",
                 type_, name);
Y
Yu Yang 已提交
135
  return it->second;
Y
Yan Chunwei 已提交
136 137
}

138 139 140 141 142 143 144 145
bool OperatorBase::HasOutputs(const std::string& name) const {
  if (outputs_.find(name) != outputs_.end()) {
    return true;
  } else {
    return false;
  }
}

146
std::string OperatorBase::Output(const std::string& name) const {
Y
Yu Yang 已提交
147
  auto& outs = Outputs(name);
Y
Yu Yang 已提交
148
  PADDLE_ENFORCE_LE(outs.size(), 1UL,
149 150
                    "Operator %s's output %s should contain only one variable.",
                    type_, name);
Y
Yu Yang 已提交
151
  return outs.empty() ? kEmptyVarName : outs[0];
Y
Yan Chunwei 已提交
152 153
}

Y
Yu Yang 已提交
154 155
const std::vector<std::string>& OperatorBase::Outputs(
    const std::string& name) const {
Y
Yu Yang 已提交
156
  auto it = outputs_.find(name);
157 158
  PADDLE_ENFORCE(it != outputs_.end(),
                 "Operator %s does not have an output called %s.", type_, name);
Y
Yu Yang 已提交
159
  return it->second;
Y
Yan Chunwei 已提交
160 161
}

162
std::string OperatorBase::DebugStringEx(const Scope* scope) const {
Q
Qiao Longfei 已提交
163
  std::stringstream ss;
Y
Yu Yang 已提交
164
  ss << "Op(" << type_ << "), inputs:{";
Y
Yu Yang 已提交
165 166
  for (auto it = inputs_.begin(); it != inputs_.end();) {
    auto& input = *it;
Y
Yu Yang 已提交
167 168 169
    ss << input.first << "[";
    for (size_t i = 0; i < input.second.size(); ++i) {
      ss << input.second[i];
170
      if (scope) {
171 172 173 174
        int row_size = GetRowSize(*scope, input.second[i]);
        if (row_size >= 0) {
          ss << "[row_size=" << row_size << "]";
        }
175
        ss << "[" << GetDims(*scope, input.second[i], true) << "]";
Q
Qiao Longfei 已提交
176
        ss << "(" << GetLoD(*scope, input.second[i]) << ")";
177
      }
Y
Yu Yang 已提交
178 179 180
      if (i != input.second.size() - 1) {
        ss << ", ";
      }
181
    }
Y
Yu Yang 已提交
182
    ss << "]";
Y
Yu Yang 已提交
183 184
    ++it;
    if (it != inputs_.end()) {
185 186
      ss << ", ";
    }
Q
Qiao Longfei 已提交
187
  }
Y
Yu Yang 已提交
188
  ss << "}, outputs:{";
Y
Yu Yang 已提交
189 190
  for (auto it = outputs_.begin(); it != outputs_.end();) {
    auto& output = *it;
Y
Yu Yang 已提交
191 192 193
    ss << output.first << "[";
    for (size_t i = 0; i < output.second.size(); ++i) {
      ss << output.second[i];
194
      if (scope) {
195 196 197 198
        int row_size = GetRowSize(*scope, output.second[i]);
        if (row_size >= 0) {
          ss << "[row_size=" << row_size << "]";
        }
199
        ss << "[" << GetDims(*scope, output.second[i], true) << "]";
Q
Qiao Longfei 已提交
200
        ss << "(" << GetLoD(*scope, output.second[i]) << ")";
201
      }
Y
Yu Yang 已提交
202 203 204
      if (i != output.second.size() - 1) {
        ss << ", ";
      }
205
    }
Y
Yu Yang 已提交
206
    ss << "]";
Y
Yu Yang 已提交
207 208
    ++it;
    if (it != outputs_.end()) {
209 210
      ss << ", ";
    }
Q
Qiao Longfei 已提交
211
  }
Y
Yu Yang 已提交
212
  ss << "}.";
Q
Qiao Longfei 已提交
213 214 215
  return ss.str();
}

Y
Yu Yang 已提交
216
OperatorBase::OperatorBase(const std::string& type,
Y
Yu Yang 已提交
217 218
                           const VariableNameMap& inputs,
                           const VariableNameMap& outputs,
Y
Yu Yang 已提交
219 220
                           const AttributeMap& attrs)
    : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
221 222
  GenerateTemporaryNames();
  CheckAllInputOutputSet();
Y
Yu Yang 已提交
223
}
224

Q
qijun 已提交
225 226
std::vector<std::string> OperatorBase::InputVars() const {
  std::vector<std::string> ret_val;
Y
Yu Yang 已提交
227
  for (auto& o : inputs_) {
Q
qijun 已提交
228 229 230 231 232 233
    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
234 235 236 237 238 239 240 241 242 243
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
244
  auto& info = OpInfoMap::Instance().Get(Type());
Y
Yu Yang 已提交
245 246

  // get all OpProto::Var for outputs
Y
Yu Yang 已提交
247
  for (auto& o : info.Proto().outputs()) {
Y
Yu Yang 已提交
248 249 250 251 252 253 254 255 256
    // 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
257 258
}

259 260 261
void OperatorBase::CheckAllInputOutputSet() const {
  auto& info_map = OpInfoMap::Instance();
  auto* op_info = info_map.GetNullable(Type());
Y
Yu Yang 已提交
262
  if (op_info == nullptr || op_info->proto_ == nullptr) return;
263 264

  for (auto& in : op_info->Proto().inputs()) {
265 266 267 268
    if (!in.dispensable()) {
      PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(),
                     "Operator %s's input, %s, is not set", Type(), in.name());
    }
269 270 271
  }

  for (auto& out : op_info->Proto().outputs()) {
272 273 274 275 276
    if (!out.dispensable()) {
      PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(),
                     "Operator %s's output, %s, is not set", Type(),
                     out.name());
    }
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
  }
}

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));
      }
    }
  }
}

293 294 295 296
static bool VarIsTensor(const Variable* var) {
  return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
}

297
static const Tensor* GetTensorFromVar(Variable* var) {
Q
QI JUN 已提交
298
  if (var->IsType<LoDTensor>()) {
299
    return var->GetMutable<LoDTensor>();
Q
QI JUN 已提交
300
  } else if (var->IsType<SelectedRows>()) {
301
    return var->GetMutable<SelectedRows>()->mutable_value();
Q
QI JUN 已提交
302
  } else {
Y
Yang Yang 已提交
303 304
    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
                 var->Type().name());
Q
QI JUN 已提交
305 306 307 308 309
  }
}

static Tensor* GetMutableTensorFromVar(Variable* var) {
  if (var->IsType<LoDTensor>()) {
310
    return var->GetMutable<LoDTensor>();
Q
QI JUN 已提交
311
  } else if (var->IsType<SelectedRows>()) {
312
    return var->GetMutable<SelectedRows>()->mutable_value();
Q
QI JUN 已提交
313
  } else {
Y
Yang Yang 已提交
314 315
    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
                 var->Type().name());
Q
QI JUN 已提交
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
bool ExecutionContext::HasInput(const std::string& name) const {
  if (!op_.HasInputs(name)) {
    return false;
  }
  auto& ins = Inputs(name);
  size_t length = ins.size();
  if (length == 0) {
    return false;
  }
  PADDLE_ENFORCE_EQ(length, 1UL,
                    "Input %s should not have more than one inputs", name);
  auto arg = ins[0];
  auto* var = arg == kEmptyVarName ? nullptr : scope_.FindVar(arg);
  return var != nullptr;
}

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

351
template <>
352
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
353
  auto* var = InputVar(name);
354 355
  return var == nullptr ? nullptr
                        : GetTensorFromVar(const_cast<Variable*>(var));
356 357 358
}

template <>
359
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
360 361 362 363
    const std::string& name) const {
  auto names = op().Inputs(name);
  std::vector<const Tensor*> res;
  res.reserve(names.size());
364 365 366 367 368
  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);
                 });
369 370 371 372
  return res;
}

template <>
373
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
374
  auto var = OutputVar(name);
Q
QI JUN 已提交
375
  return var == nullptr ? nullptr : GetMutableTensorFromVar(var);
376 377 378
}

template <>
379
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
380 381 382 383
    const std::string& name) const {
  auto names = op().Outputs(name);
  std::vector<Tensor*> res;
  res.reserve(names.size());
384 385
  std::transform(names.begin(), names.end(), std::back_inserter(res),
                 [&](const std::string& sub_name) {
386 387
                   auto var = scope_.FindVar(sub_name);
                   return var == nullptr ? nullptr
Q
QI JUN 已提交
388
                                         : GetMutableTensorFromVar(var);
389
                 });
390 391 392
  return res;
}

Y
Yu Yang 已提交
393 394 395 396 397 398 399 400 401 402 403 404 405 406 407
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;
}

408 409 410 411 412 413
class RuntimeInferShapeContext : public InferShapeContext {
 public:
  RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope)
      : op_(op), scope_(scope) {}

  bool HasInput(const std::string& name) const override {
414 415 416
    if (!op_.HasInputs(name)) {
      return false;
    }
417 418 419 420 421
    auto& ins = Inputs(name);
    size_t length = ins.size();
    if (length == 0) {
      return false;
    }
F
fengjiayi 已提交
422 423
    PADDLE_ENFORCE_EQ(length, 1UL,
                      "Input %s should not have more than one inputs", name);
424 425 426 427 428 429
    auto ipt = ins[0];
    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
    return var != nullptr;
  }

  bool HasOutput(const std::string& name) const override {
430 431 432
    if (!op_.HasOutputs(name)) {
      return false;
    }
433 434 435 436 437
    auto& outs = Outputs(name);
    size_t length = outs.size();
    if (length == 0) {
      return false;
    }
F
fengjiayi 已提交
438 439
    PADDLE_ENFORCE_EQ(length, 1UL,
                      "Output %s should not have more than one inputs", name);
440 441 442 443 444 445
    auto ipt = outs[0];
    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
    return var != nullptr;
  }

  bool HasInputs(const std::string& name) const override {
446 447 448
    if (!op_.HasInputs(name)) {
      return false;
    }
449 450 451 452 453 454 455 456 457 458 459 460 461
    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 {
462 463 464
    if (!op_.HasOutputs(name)) {
      return false;
    }
465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488
    auto outputs = op_.Outputs(name);
    if (outputs.empty()) {
      return false;
    }
    for (auto& output : outputs) {
      if (scope_.FindVar(output) == nullptr) {
        return false;
      }
    }
    return true;
  }

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

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

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

Q
Qiao Longfei 已提交
489 490 491 492 493 494 495 496 497 498 499 500
  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
501

M
mozga-intel 已提交
502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520
// 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
521 522 523 524 525 526 527 528 529 530 531 532 533 534
  }

  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
535 536
  }

537 538 539
  bool IsRuntime() const override { return true; }

 protected:
540 541
  DDim GetDim(const std::string& name) const override {
    Variable* var = scope_.FindVar(name);
F
fengjiayi 已提交
542
    PADDLE_ENFORCE_NOT_NULL(var);
543 544 545 546 547
    if (var->IsType<LoDTensor>()) {
      return var->Get<LoDTensor>().dims();
    } else if (var->IsType<SelectedRows>()) {
      return var->Get<SelectedRows>().GetCompleteDims();
    } else {
F
fengjiayi 已提交
548 549 550 551 552 553 554
      PADDLE_THROW(
          "Only LoDTensor/SelectedRows support 'GetDim', but Variable %s's "
          "type_id is %s.",
          name, var->Type().name());
    }
  }

F
fengjiayi 已提交
555
  std::vector<DDim> GetRepeatedDims(const std::string& name) const override {
Y
Yu Yang 已提交
556
    PADDLE_THROW("Only compile time support this method");
557 558 559 560 561 562 563 564 565
  }

  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
566 567
      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
                   name, var->Type().name());
568 569 570
    }
  }

F
fengjiayi 已提交
571 572
  void SetRepeatedDims(const std::string& name,
                       const std::vector<DDim>& dims) override {
Y
Yu Yang 已提交
573
    PADDLE_THROW("Only compile time support this method");
F
fengjiayi 已提交
574 575
  }

576
  proto::VarType::Type GetVarType(const std::string& name) const override {
577 578 579 580
    auto* var = scope_.FindVar(name);
    return ToVarType(var->Type());
  }

F
fengjiayi 已提交
581 582 583 584
  InferShapeVarPtr GetVarPtr(const std::string& name) override {
    return scope_.FindVar(name);
  }

585
 private:
586 587 588 589
  const OperatorBase& op_;
  const Scope& scope_;
};

C
chengduoZH 已提交
590 591 592 593 594
static void CheckTensorNANOrInf(const std::string& name,
                                const framework::Tensor& tensor) {
  if (tensor.memory_size() == 0) {
    return;
  }
S
sneaxiy 已提交
595
  if (!IsType<float>(tensor.type()) && !IsType<double>(tensor.type())) {
C
chengduoZH 已提交
596 597 598 599 600 601 602 603
    return;
  }
  PADDLE_ENFORCE(!framework::TensorContainsInf(tensor),
                 "Tensor %s contains Inf", name);
  PADDLE_ENFORCE(!framework::TensorContainsNAN(tensor),
                 "Tensor %s contains NAN", name);
}

604 605
void OperatorWithKernel::RunImpl(const Scope& scope,
                                 const platform::Place& place) const {
606 607
  RuntimeInferShapeContext infer_shape_ctx(*this, scope);
  this->InferShape(&infer_shape_ctx);
Y
Yu Yang 已提交
608
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
609
  auto* dev_ctx = pool.Get(place);
610 611 612 613

  // For profiling, don't move out of this function because that will result
  // in the failure of multi-GPU profiling.
  platform::RecordEvent record_event(Type(), dev_ctx);
614 615 616 617
  // 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
618 619
    PADDLE_THROW(
        "There are no kernels which are registered in the %s operator.", type_);
620 621
  }

Q
qiaolongfei 已提交
622 623
  OpKernelMap& kernels = kernels_iter->second;

624 625
  // TODO(dzhwinter) : kernel fallback mechanism will be added when all the
  // transform functions are ready.
Q
qiaolongfei 已提交
626

627 628 629 630
  // for (auto& candidate : kKernelPriority) {
  //   Do selection
  // }

Y
yuyang18 已提交
631 632
  auto expected_kernel_key =
      this->GetExpectedKernelType(ExecutionContext(*this, scope, *dev_ctx));
Q
qiaolongfei 已提交
633 634
  VLOG(3) << "expected_kernel_key:" << expected_kernel_key;

635
  auto kernel_iter = kernels.find(expected_kernel_key);
636
#ifdef PADDLE_WITH_MKLDNN
P
Paweł Żelazko 已提交
637
  // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
638 639
  if (kernel_iter == kernels.end() &&
      expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
P
Paweł Żelazko 已提交
640
    VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
641 642 643 644 645
    expected_kernel_key.library_type_ = LibraryType::kPlain;
    expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
    kernel_iter = kernels.find(expected_kernel_key);
  }
#endif
646 647 648 649 650
  if (kernel_iter == kernels.end()) {
    PADDLE_THROW("op %s does not have kernel for %s", type_,
                 KernelTypeToString(expected_kernel_key));
  }

Y
yuyang18 已提交
651 652 653 654
  // do data transformScope &transfer_scope;
  std::vector<std::string> transfered_inplace_vars;
  auto* transfer_scope =
      TryTransferData(scope, expected_kernel_key, &transfered_inplace_vars);
655

Y
yuyang18 已提交
656 657 658 659 660 661
  // exec scope is the scope that kernel actually executed on.
  const Scope& exec_scope =
      (transfer_scope == nullptr ? scope : *transfer_scope);

  if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) {
    dev_ctx = pool.Get(expected_kernel_key.place_);
662
  }
Q
QI JUN 已提交
663

Y
yuyang18 已提交
664
  kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx));
D
dzhwinter 已提交
665

Y
yuyang18 已提交
666 667 668
  if (!transfered_inplace_vars.empty()) {
    // there is inplace variable has been transfered.
    TransferInplaceVarsBack(scope, transfered_inplace_vars, *transfer_scope);
669 670
  }

D
dzhwinter 已提交
671
  /*For profiling/benchmark only*/
D
dzhwinter 已提交
672
  if (FLAGS_benchmark) {
Y
yuyang18 已提交
673
    dev_ctx->Wait();
D
dzhwinter 已提交
674
  }
C
chengduoZH 已提交
675 676 677

  if (FLAGS_check_nan_inf) {
    for (auto& vname : OutputVars(true)) {
Y
yuyang18 已提交
678
      auto* var = exec_scope.FindVar(vname);
C
chengduoZH 已提交
679 680 681
      if (var == nullptr) continue;
      if (var->IsType<framework::LoDTensor>()) {
        CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
682 683
      } else if (var->IsType<framework::SelectedRows>()) {
        CheckTensorNANOrInf(vname, var->Get<framework::SelectedRows>().value());
C
chengduoZH 已提交
684 685 686
      }
    }
  }
Q
Qiao Longfei 已提交
687
}
Y
yuyang18 已提交
688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738
void OperatorWithKernel::TransferInplaceVarsBack(
    const Scope& scope, const std::vector<std::string>& inplace_vars,
    const Scope& transfer_scope) const {
  for (auto& var_name : inplace_vars) {
    VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
    auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
    auto* transformed_tensor =
        GetTensorFromVar(transfer_scope.FindVar(var_name));
    original_tensor->ShareDataWith(*transformed_tensor);
  }
}

Scope* OperatorWithKernel::TryTransferData(
    const Scope& scope, const OpKernelType& expected_kernel_key,
    std::vector<std::string>* transfered_inplace_vars) const {
  Scope* new_scope = nullptr;
  for (auto& var_name_item : Inputs()) {
    for (auto& var_name : var_name_item.second) {
      auto* var = scope.FindVar(var_name);
      // Only tensor can be tranfer to another device.
      if (var == nullptr || !VarIsTensor(var)) {
        continue;
      }

      auto* tensor_in = GetTensorFromVar(var);
      if (!tensor_in->IsInitialized()) {
        continue;
      }

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

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

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

      VLOG(3) << "Transform Variable " << var_name << " from "
              << kernel_type_for_var << " to " << expected_kernel_key;

      if (new_scope == nullptr) {
        new_scope = &scope.NewScope();
      }

      auto* trans_var = new_scope->Var(var_name);
      Tensor out;
Y
yuyang18 已提交
739
      TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in, &out);
Y
yuyang18 已提交
740 741 742 743 744 745
      SetTensorToVariable(*var, out, trans_var);
    }
  }

  return new_scope;
}
Q
Qiao Longfei 已提交
746

747
proto::VarType::Type OperatorWithKernel::IndicateDataType(
Y
Yu Yang 已提交
748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764
    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()));
765 766 767 768
          PADDLE_ENFORCE(
              tmp == data_type || data_type == -1,
              "DataType of Paddle Op %s must be the same. Get %d != %d", Type(),
              data_type, tmp);
Y
Yu Yang 已提交
769 770 771 772 773 774
          data_type = tmp;
        }
      }
    }
  }
  PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
775
  return static_cast<proto::VarType::Type>(data_type);
Y
Yu Yang 已提交
776
}
777

778 779 780 781 782 783 784 785
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
786 787
  return OpKernelType(expected_kernel_type.data_type_, tensor.place(),
                      tensor.layout());
788 789
}

Q
Qiao Longfei 已提交
790
}  // namespace framework
L
liaogang 已提交
791
}  // namespace paddle