operator.h 13.3 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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. */

#pragma once

D
dongzhihong 已提交
17
#include <algorithm>
Q
Qiao Longfei 已提交
18 19 20 21
#include <string>
#include <unordered_map>
#include <vector>

Y
Yu Yang 已提交
22
#include "op_info.h"
Y
Yi Wang 已提交
23
#include "paddle/framework/attribute.h"
Y
Yu Yang 已提交
24
#include "paddle/framework/framework.pb.h"
Q
qijun 已提交
25 26 27 28
#include "paddle/framework/scope.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/device_context.h"
#include "paddle/platform/place.h"
Y
Yu Yang 已提交
29
#include "paddle/platform/variant.h"
Q
qijun 已提交
30
#include "paddle/utils/Error.h"
Q
Qiao Longfei 已提交
31 32 33 34

namespace paddle {
namespace framework {

35
/// If a variable is a empty variable, that name will be used.
36
constexpr char kEmptyVarName[] = "@EMPTY@";
37 38 39

/// If a variable is a temporary variable, that name will be set in Python,
/// but it will be convert to a unique name in scope after OpCreator.
40
constexpr char kTempVarName[] = "@TEMP@";
41 42 43 44

/// If a variable's name has a certain suffix, it means that the
/// variable is the gradient of another varibale.
/// e.g. Variable "x@GRAD" is the gradient of varibale "x".
45
constexpr char kGradVarSuffix[] = "@GRAD";
46 47

/// Variables with this suffix are supposed to be filled up with zeros.
48
constexpr char kZeroVarSuffix[] = "@ZERO";
49 50 51 52 53

inline std::string GradVarName(const std::string& var_name) {
  return var_name + kGradVarSuffix;
}

Q
Qiao Longfei 已提交
54
class OperatorBase;
55 56
class InferShapeContext;
class ExecutionContext;
57

Q
Qiao Longfei 已提交
58 59 60 61 62 63 64 65
/**
 * OperatorBase has the basic element that Net will call to do computation.
 * Only CreateOperator from OpRegistry will new Operator directly. User
 * should always construct a proto message OpDesc and call
 * OpRegistry::CreateOp(op_desc) to get an Operator instance.
 */
class OperatorBase {
 public:
Y
Yu Yang 已提交
66 67
  OperatorBase(const std::string& type, const VariableNameMap& inputs,
               const VariableNameMap& outputs, const AttributeMap& attrs);
68

Q
Qiao Longfei 已提交
69 70 71 72 73 74 75 76 77
  virtual ~OperatorBase() {}

  template <typename T>
  inline const T& GetAttr(const std::string& name) const {
    PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
                   name);
    return boost::get<T>(attrs_.at(name));
  }

78
  virtual std::string DebugString() const;
Q
Qiao Longfei 已提交
79 80 81

  /// InferShape infer the size of Variables used by this Operator with
  /// information inside scope
Y
Yu Yang 已提交
82
  virtual void InferShape(const Scope& scope) const = 0;
Q
Qiao Longfei 已提交
83 84

  /// Net will call this function to Run an op.
Y
Yu Yang 已提交
85
  virtual void Run(const Scope& scope,
Y
Yu Yang 已提交
86 87
                   const platform::DeviceContext& dev_ctx) const = 0;

Y
Yu Yang 已提交
88 89
  virtual bool IsNetOp() const { return false; }

90 91
  virtual bool SupportGPU() const { return false; }

D
dongzhihong 已提交
92 93 94
  /// rename inputs outputs name
  void Rename(const std::string& old_name, const std::string& new_name);

Y
Yu Yang 已提交
95 96
  const VariableNameMap& Inputs() const { return inputs_; }
  const VariableNameMap& Outputs() const { return outputs_; }
Y
Yu Yang 已提交
97
  //! Get a input with argument's name described in `op_proto`
98
  std::string Input(const std::string& name) const;
Y
Yu Yang 已提交
99
  //! Get a input which has multiple variables.
Y
Yu Yang 已提交
100
  const std::vector<std::string>& Inputs(const std::string& name) const;
Y
Yi Wang 已提交
101

Y
Yu Yang 已提交
102
  //! Get a output with argument's name described in `op_proto`
103
  std::string Output(const std::string& name) const;
Y
Yu Yang 已提交
104 105
  //! Get an output which has multiple variables.
  //! TODO add a vector_view to prevent memory copy.
Y
Yu Yang 已提交
106
  const std::vector<std::string>& Outputs(const std::string& name) const;
Y
Yan Chunwei 已提交
107

Y
Yu Yang 已提交
108
  virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
109

Q
qiaolongfei 已提交
110
  const std::string& Type() const { return type_; }
Q
qiaolongfei 已提交
111
  void SetType(const std::string& type) { type_ = type; }
Y
Yi Wang 已提交
112 113
  const AttributeMap& Attrs() const { return attrs_; }

Y
Yu Yang 已提交
114
  // Return a new operator instance, which is as same as this.
Y
Yu Yang 已提交
115 116
  // Use unique_ptr to prevent caller forget to delete this pointer.
  virtual std::unique_ptr<OperatorBase> Clone() const = 0;
Y
Yu Yang 已提交
117

Q
qiaolongfei 已提交
118
 protected:
Q
Qiao Longfei 已提交
119
  std::string type_;
D
dongzhihong 已提交
120
  // NOTE: in case of OpGrad, inputs_ contains:
Y
Yu Yang 已提交
121
  // I (Inputs)opear
D
dongzhihong 已提交
122 123
  // O (Outputs)
  // OG (Output Gradients)
Y
Yu Yang 已提交
124
  VariableNameMap inputs_;
Y
Yu Yang 已提交
125

D
dongzhihong 已提交
126 127
  // NOTE: in case of OpGrad, outputs_ contains
  // IG (Inputs Gradients)
Y
Yu Yang 已提交
128
  VariableNameMap outputs_;
Q
Qiao Longfei 已提交
129
  AttributeMap attrs_;
Y
Yan Chunwei 已提交
130 131
};

Y
Yu Yang 已提交
132 133
// Macro for define a clone method.
// If you are writing an kernel operator, `Clone` will be defined when you
134
// register it. i.e. `Clone` method is not needed to define by yourself.
Y
Yu Yang 已提交
135
#define DEFINE_OP_CLONE_METHOD(cls)                       \
Y
Yu Yang 已提交
136
  std::unique_ptr<OperatorBase> Clone() const final {     \
Y
Yu Yang 已提交
137
    return std::unique_ptr<OperatorBase>(new cls(*this)); \
Y
Yu Yang 已提交
138
  }
Y
Yu Yang 已提交
139

Y
Yu Yang 已提交
140 141 142 143
// Macro for define a default constructor for Operator.
// You can also use
//   using PARENT_CLASS::PARENT_CLASS;
// to use parent's constructor.
Y
Yu Yang 已提交
144 145
#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls)             \
  cls(const std::string& type,                             \
Y
Yu Yang 已提交
146 147 148
      const ::paddle::framework::VariableNameMap& inputs,  \
      const ::paddle::framework::VariableNameMap& outputs, \
      const paddle::framework::AttributeMap& attrs)        \
Y
Yu Yang 已提交
149
      : parent_cls(type, inputs, outputs, attrs) {}
Y
Yu Yang 已提交
150

151 152
class NOP : public OperatorBase {
 public:
153
  using OperatorBase::OperatorBase;
154 155 156
  void InferShape(const Scope& scope) const override {}
  void Run(const Scope& scope,
           const platform::DeviceContext& dev_ctx) const override {}
157 158 159
  std::unique_ptr<OperatorBase> Clone() const override {
    return std::unique_ptr<OperatorBase>(new NOP(*this));
  }
160 161
};

162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
// this class not only make proto but also init attribute checkers.
class OpProtoAndCheckerMaker {
 public:
  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
      : proto_(proto), op_checker_(op_checker) {}

  ~OpProtoAndCheckerMaker() {
    PADDLE_ENFORCE(validated_, "should call Validate after build");
  }

  void Validate();

 protected:
  struct VariableBuilder {
    OpProto::Var* var_;

    VariableBuilder& AsDuplicable() {
      var_->set_duplicable(true);
      return *this;
    }

    VariableBuilder& AsIntermediate() {
      var_->set_intermediate(true);
      return *this;
    }

188 189
    VariableBuilder& NotInGradient() {
      var_->set_not_in_gradient(true);
190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226
      return *this;
    }
  };

  VariableBuilder AddInput(const std::string& name, const std::string& comment);

  VariableBuilder AddOutput(const std::string& name,
                            const std::string& comment);

  template <typename T>
  TypedAttrChecker<T>& AddAttr(const std::string& name,
                               const std::string& comment,
                               bool generated = false) {
    auto* attr = proto_->add_attrs();
    attr->set_name(name);
    attr->set_comment(comment);
    attr->set_generated(generated);
    attr->set_type(AttrTypeID<T>());
    return op_checker_->AddAttrChecker<T>(name);
  }

  void AddComment(const std::string& comment) { proto_->set_comment(comment); }

 private:
  void CheckNoDuplicatedInOutAttrs();

  OpProto* proto_;
  OpAttrChecker* op_checker_;
  bool validated_{false};
};

class NOPMaker : public OpProtoAndCheckerMaker {
 public:
  NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {}
};

227
class InferShapeContext {
Y
Yan Chunwei 已提交
228
 public:
229 230
  InferShapeContext(const OperatorBase& op, const Scope& scope)
      : op_(op), scope_(scope) {}
231

Y
Yu Yang 已提交
232
  size_t InputSize(const std::string& name) const {
Y
Yu Yang 已提交
233
    return op_.Inputs(name).size();
Y
Yan Chunwei 已提交
234 235
  }

Y
Yu Yang 已提交
236
  size_t OutputSize(const std::string& name) const {
Y
Yu Yang 已提交
237
    return op_.Outputs(name).size();
Y
Yan Chunwei 已提交
238 239
  }

240
  const Variable* InputVar(const std::string& name) const {
241
    auto ipt = op_.Input(name);
Y
Yu Yang 已提交
242
    return ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
Y
Yan Chunwei 已提交
243 244
  }

245
  Variable* OutputVar(const std::string& name) const {
246
    auto opt = op_.Output(name);
Y
Yu Yang 已提交
247
    return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt);
Y
Yan Chunwei 已提交
248 249
  }

250 251
  const std::vector<const Variable*> MultiInputVar(
      const std::string& name) const {
Y
Yan Chunwei 已提交
252 253
    auto names = op_.Inputs(name);
    std::vector<const Variable*> res;
254
    res.reserve(names.size());
255 256
    std::transform(names.begin(), names.end(), std::back_inserter(res),
                   [this](const std::string& name) {
Y
Yu Yang 已提交
257 258
                     return name == kEmptyVarName ? nullptr
                                                  : scope_.FindVar(name);
259
                   });
Y
Yan Chunwei 已提交
260 261 262
    return res;
  }

263
  std::vector<const Variable*> MultiOutputVar(const std::string& name) const {
Y
Yan Chunwei 已提交
264 265
    auto names = op_.Outputs(name);
    std::vector<const Variable*> res;
266
    res.reserve(names.size());
267 268
    std::transform(names.begin(), names.end(), std::back_inserter(res),
                   [this](const std::string& name) {
Y
Yu Yang 已提交
269 270
                     return name == kEmptyVarName ? nullptr
                                                  : scope_.FindVar(name);
271
                   });
Y
Yan Chunwei 已提交
272 273 274
    return res;
  }

275 276
  template <typename T>
  const T* Input(const std::string& name) const {
Y
Yu Yang 已提交
277
    auto* var = InputVar(name);
278
    return var == nullptr ? nullptr : &var->Get<T>();
279 280 281 282
  }

  template <typename T>
  T* Output(const std::string& name) const {
283
    auto var = OutputVar(name);
284
    return var == nullptr ? nullptr : var->GetMutable<T>();
285 286 287 288 289 290 291 292
  }

  template <typename T>
  const std::vector<const T*> MultiInput(const std::string& name) const {
    auto names = op_.Inputs(name);
    std::vector<const T*> res;
    res.reserve(names.size());
    std::transform(names.begin(), names.end(), std::back_inserter(res),
293
                   [&](const std::string& sub_name) {
294
                     auto var = scope_.FindVar(sub_name);
295
                     return var == nullptr ? nullptr : &var->Get<T>();
296 297 298 299 300 301 302 303 304 305
                   });
    return res;
  }

  template <typename T>
  std::vector<const T*> MultiOutput(const std::string& name) const {
    auto names = op_.Outputs(name);
    std::vector<const T*> res;
    res.reserve(names.size());
    std::transform(names.begin(), names.end(), std::back_inserter(res),
306
                   [&](const std::string& sub_name) {
307
                     auto var = scope_.FindVar(sub_name);
308
                     return var == nullptr ? nullptr : var->GetMutable<T>();
309 310 311 312 313
                   });
    return res;
  }

  const OperatorBase& op_;
314
  const Scope& scope_;
315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331
};

template <typename T>
struct EigenDeviceConverter;

template <>
struct EigenDeviceConverter<platform::CPUPlace> {
  using EigenDeviceType = Eigen::DefaultDevice;
};

#ifndef PADDLE_ONLY_CPU
template <>
struct EigenDeviceConverter<platform::GPUPlace> {
  using EigenDeviceType = Eigen::GpuDevice;
};
#endif

332
class ExecutionContext : public InferShapeContext {
333
 public:
334
  ExecutionContext(const OperatorBase& op, const Scope& scope,
D
dongzhihong 已提交
335
                   const platform::DeviceContext* device_context)
336
      : InferShapeContext(op, scope), device_context_(device_context) {}
337

Q
qijun 已提交
338 339 340
  template <typename PlaceType,
            typename DeviceType =
                typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
341
  DeviceType& GetEigenDevice() const;
Q
qijun 已提交
342

D
dongzhihong 已提交
343
  platform::Place GetPlace() const { return device_context_->GetPlace(); }
Q
qijun 已提交
344

Q
qijun 已提交
345
  const platform::DeviceContext* device_context() const {
Q
qijun 已提交
346
    return device_context_;
Q
qijun 已提交
347
  }
Q
qijun 已提交
348

D
dongzhihong 已提交
349
  const platform::DeviceContext* device_context_;
Q
Qiao Longfei 已提交
350 351
};

Q
qijun 已提交
352 353
class OpKernel {
 public:
Q
qijun 已提交
354
  /**
355
   * ExecutionContext is the only parameter of Kernel Run function.
Q
qijun 已提交
356 357
   * Run will get input/output variables, state such as momentum and
   * device resource such as CUDA stream, cublas handle, etc. from
358
   * ExecutionContext. User should construct it before run the Operator.
Q
qijun 已提交
359 360
   */

361
  virtual void Compute(const ExecutionContext& context) const = 0;
Y
Yu Yang 已提交
362 363 364 365

  virtual ~OpKernel() {}
};

Q
Qiao Longfei 已提交
366 367
class OperatorWithKernel : public OperatorBase {
 public:
Y
Yu Yang 已提交
368 369
  struct OpKernelKey {
    platform::Place place_;
Q
Qiao Longfei 已提交
370

Y
Yu Yang 已提交
371
    OpKernelKey() = default;
L
liaogang 已提交
372
    explicit OpKernelKey(const platform::DeviceContext& dev_ctx) {
Y
Yu Yang 已提交
373 374 375
      place_ = dev_ctx.GetPlace();
    }

Q
qijun 已提交
376 377 378
    bool operator==(const OpKernelKey& o) const {
      return platform::places_are_same_class(place_, o.place_);
    }
Y
Yu Yang 已提交
379 380 381 382 383 384 385 386 387 388 389
  };

  struct OpKernelHash {
    std::hash<bool> hash_;
    size_t operator()(const OpKernelKey& key) const {
      return hash_(platform::is_gpu_place(key.place_));
    }
  };

  using OpKernelMap =
      std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
Q
Qiao Longfei 已提交
390

Y
Yu Yang 已提交
391 392
  OperatorWithKernel(const std::string& type, const VariableNameMap& inputs,
                     const VariableNameMap& outputs, const AttributeMap& attrs)
Y
Yu Yang 已提交
393 394
      : OperatorBase(type, inputs, outputs, attrs) {}

395
  void InferShape(const Scope& scope) const override {
396
    InferShape(InferShapeContext(*this, scope));
397 398
  }

Y
Yu Yang 已提交
399
  void Run(const Scope& scope,
Y
Yu Yang 已提交
400
           const platform::DeviceContext& dev_ctx) const final {
Q
Qiao Longfei 已提交
401
    auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
402
    opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
Q
Qiao Longfei 已提交
403 404
  }

Y
Yu Yang 已提交
405 406 407 408
  static std::unordered_map<std::string /* op_type */, OpKernelMap>&
  AllOpKernels() {
    static std::unordered_map<std::string, OpKernelMap> g_all_op_kernels;
    return g_all_op_kernels;
Y
Yu Yang 已提交
409
  }
Y
Yan Chunwei 已提交
410

411 412 413 414 415 416
  bool SupportGPU() const override {
    OperatorWithKernel::OpKernelKey key;
    key.place_ = platform::GPUPlace();
    return OperatorWithKernel::AllOpKernels().at(type_).count(key) != 0;
  }

Y
Yu Yang 已提交
417
 protected:
418
  virtual void InferShape(const InferShapeContext& ctx) const = 0;
Q
Qiao Longfei 已提交
419 420 421 422
};

}  // namespace framework
}  // namespace paddle