operator.h 10.2 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 22
#include <boost/variant.hpp>
#include <string>
#include <unordered_map>
#include <vector>

Q
qijun 已提交
23 24
#include "paddle/framework/attr_checker.h"
#include "paddle/framework/op_desc.pb.h"
Y
Yan Chunwei 已提交
25
#include "paddle/framework/op_proto.pb.h"
Q
qijun 已提交
26 27 28 29 30
#include "paddle/framework/scope.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/device_context.h"
#include "paddle/platform/place.h"
#include "paddle/utils/Error.h"
Q
Qiao Longfei 已提交
31 32 33 34 35

namespace paddle {
namespace framework {

class OperatorBase;
36 37
class InferShapeContext;
class ExecutionContext;
Q
Qiao Longfei 已提交
38 39 40 41 42 43 44 45
/**
 * 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:
46 47 48 49 50 51 52
  /// If a variable is a empty variable, that name will be used.
  static std::string EMPTY_VAR_NAME() { return "@EMPTY@"; }

  /// 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.
  static std::string TMP_VAR_NAME() { return "@TEMP@"; }

F
fengjiayi 已提交
53 54 55 56 57
  /// 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".
  static std::string GRAD_VAR_SUFFIX() { return "@GRAD"; }

58 59 60
  /// Variables with this suffix are supposed to be filled up with zeros.
  static std::string ZERO_VAR_SUFFIX() { return "@ZERO"; }

Q
Qiao Longfei 已提交
61 62 63 64 65 66 67 68 69
  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));
  }

70
  virtual std::string DebugString() const;
Q
Qiao Longfei 已提交
71

Q
Qiao Longfei 已提交
72 73 74 75
  /// Init will be called after CreateOperator, you can put some initialization
  /// logic here.
  virtual void Init() {}

Q
Qiao Longfei 已提交
76 77
  /// InferShape infer the size of Variables used by this Operator with
  /// information inside scope
Y
Yu Yang 已提交
78
  virtual void InferShape(const Scope& scope) const = 0;
Q
Qiao Longfei 已提交
79 80

  /// Net will call this function to Run an op.
Y
Yu Yang 已提交
81
  virtual void Run(const Scope& scope,
Y
Yu Yang 已提交
82 83
                   const platform::DeviceContext& dev_ctx) const = 0;

Y
Yu Yang 已提交
84 85
  virtual bool IsNetOp() const { return false; }

D
dongzhihong 已提交
86 87 88
  /// rename inputs outputs name
  void Rename(const std::string& old_name, const std::string& new_name);

Y
Yu Yang 已提交
89
  //! Get a input with argument's name described in `op_proto`
Y
Yan Chunwei 已提交
90
  const std::string& Input(const std::string& name) const;
Y
Yu Yang 已提交
91

Y
Yu Yang 已提交
92 93
  //! Get a input which has multiple variables.
  //! TODO add a vector_view to prevent memory copy.
Y
Yan Chunwei 已提交
94
  std::vector<std::string> Inputs(const std::string& name) const;
Y
Yu Yang 已提交
95
  //! Get a output with argument's name described in `op_proto`
Y
Yan Chunwei 已提交
96
  const std::string& Output(const std::string& name) const;
Y
Yu Yang 已提交
97 98
  //! Get an output which has multiple variables.
  //! TODO add a vector_view to prevent memory copy.
Y
Yan Chunwei 已提交
99 100
  std::vector<std::string> Outputs(const std::string& name) const;

Q
Qiao Longfei 已提交
101
 public:
Q
Qiao Longfei 已提交
102
  std::string type_;
D
dongzhihong 已提交
103 104 105 106
  // NOTE: in case of OpGrad, inputs_ contains:
  // I (Inputs)
  // O (Outputs)
  // OG (Output Gradients)
Q
Qiao Longfei 已提交
107
  std::vector<std::string> inputs_;
D
dongzhihong 已提交
108 109
  // NOTE: in case of OpGrad, outputs_ contains
  // IG (Inputs Gradients)
Q
Qiao Longfei 已提交
110 111
  std::vector<std::string> outputs_;
  AttributeMap attrs_;
Y
Yan Chunwei 已提交
112
  // store the arguments' offset described in op_desc.
Y
Yu Yang 已提交
113
  std::shared_ptr<std::unordered_map<std::string, int>> in_out_idxs_;
Y
Yan Chunwei 已提交
114 115
};

116
class OperatorContext {
Y
Yan Chunwei 已提交
117
 public:
118
  OperatorContext(const OperatorBase* op, const Scope& scope)
119 120 121
      : op_(*op), scope_(scope) {}

  size_t InputSize() const { return op_.inputs_.size(); }
Y
Yan Chunwei 已提交
122

123 124
  size_t OutputSize() const { return op_.outputs_.size(); }

125
  const Variable* InputVar(const size_t index) const {
126
    return scope_.FindVar(op_.inputs_.at(index));
Y
Yan Chunwei 已提交
127 128
  }

129
  Variable* OutputVar(const size_t index) const {
130
    return scope_.FindVar(op_.outputs_.at(index));
Y
Yan Chunwei 已提交
131 132
  }

133
  const Variable* InputVar(const std::string& name) const {
Y
Yu Yang 已提交
134
    return scope_.FindVar(op_.Input(name));
Y
Yan Chunwei 已提交
135 136
  }

137
  Variable* OutputVar(const std::string& name) const {
Y
Yu Yang 已提交
138
    return scope_.FindVar(op_.Output(name));
Y
Yan Chunwei 已提交
139 140
  }

141 142
  const std::vector<const Variable*> MultiInputVar(
      const std::string& name) const {
Y
Yan Chunwei 已提交
143 144
    auto names = op_.Inputs(name);
    std::vector<const Variable*> res;
145
    res.reserve(names.size());
Y
Yan Chunwei 已提交
146
    std::transform(
147
        names.begin(), names.end(), std::back_inserter(res),
Y
Yu Yang 已提交
148
        [this](const std::string& name) { return scope_.FindVar(name); });
Y
Yan Chunwei 已提交
149 150 151
    return res;
  }

152
  std::vector<const Variable*> MultiOutputVar(const std::string& name) const {
Y
Yan Chunwei 已提交
153 154
    auto names = op_.Outputs(name);
    std::vector<const Variable*> res;
155
    res.reserve(names.size());
Y
Yan Chunwei 已提交
156
    std::transform(
157
        names.begin(), names.end(), std::back_inserter(res),
Y
Yu Yang 已提交
158
        [this](const std::string& name) { return scope_.FindVar(name); });
Y
Yan Chunwei 已提交
159 160 161
    return res;
  }

162
  template <typename T>
163 164 165 166
  const T* Input(const size_t index) const {
    auto var = InputVar(index);
    PADDLE_ENFORCE(var != nullptr, "Input(%d) should not be nullptr", index);
    return &var->Get<T>();
167 168 169
  }

  template <typename T>
170 171 172 173
  T* Output(const size_t index) const {
    auto var = OutputVar(index);
    PADDLE_ENFORCE(var != nullptr, "Output(%d) should not be nullptr", index);
    return var->GetMutable<T>();
174 175 176 177
  }

  template <typename T>
  const T* Input(const std::string& name) const {
178 179 180
    auto var = InputVar(name);
    PADDLE_ENFORCE(var != nullptr, "Input(%s) should not be nullptr", name);
    return &var->Get<T>();
181 182 183 184
  }

  template <typename T>
  T* Output(const std::string& name) const {
185 186 187
    auto var = OutputVar(name);
    PADDLE_ENFORCE(var != nullptr, "Output(%s) should not be nullptr", name);
    return var->GetMutable<T>();
188 189 190 191 192 193 194 195
  }

  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),
196
                   [&](const std::string& sub_name) {
197
                     auto var = scope_.FindVar(sub_name);
198 199 200 201
                     PADDLE_ENFORCE(var != nullptr,
                                    "MultiInput(%s:%s) should not be nullptr",
                                    name, sub_name);
                     return &var->Get<T>();
202 203 204 205 206 207 208 209 210 211
                   });
    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),
212
                   [&](const std::string& sub_name) {
213
                     auto var = scope_.FindVar(sub_name);
214 215 216 217
                     PADDLE_ENFORCE(var != nullptr,
                                    "MultiOutput(%s:%s) should not be nullptr",
                                    name, sub_name);
                     return var->GetMutable<T>();
218 219 220 221 222
                   });
    return res;
  }

  const OperatorBase& op_;
223
  const Scope& scope_;
224 225 226 227
};

class InferShapeContext : public OperatorContext {
 public:
228
  InferShapeContext(const OperatorBase* op, const Scope& scope)
229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
      : OperatorContext(op, scope) {}
};

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

class ExecutionContext : public OperatorContext {
 public:
249
  ExecutionContext(const OperatorBase* op, const Scope& scope,
250 251 252
                   const platform::DeviceContext& device_context)
      : OperatorContext(op, scope), device_context_(device_context) {}

Q
qijun 已提交
253 254 255 256 257 258 259
  template <typename PlaceType,
            typename DeviceType =
                typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
  DeviceType* GetEigenDevice() const;

  platform::Place GetPlace() const { return device_context_.GetPlace(); }

Y
Yan Chunwei 已提交
260
  const platform::DeviceContext& device_context_;
Q
Qiao Longfei 已提交
261 262
};

Q
qijun 已提交
263 264
class OpKernel {
 public:
Q
qijun 已提交
265
  /**
266
   * ExecutionContext is the only parameter of Kernel Run function.
Q
qijun 已提交
267 268
   * Run will get input/output variables, state such as momentum and
   * device resource such as CUDA stream, cublas handle, etc. from
269
   * ExecutionContext. User should construct it before run the Operator.
Q
qijun 已提交
270 271
   */

272
  virtual void Compute(const ExecutionContext& context) const = 0;
Y
Yu Yang 已提交
273 274 275 276

  virtual ~OpKernel() {}
};

Q
Qiao Longfei 已提交
277 278
class OperatorWithKernel : public OperatorBase {
 public:
Y
Yu Yang 已提交
279 280
  struct OpKernelKey {
    platform::Place place_;
Q
Qiao Longfei 已提交
281

Y
Yu Yang 已提交
282 283 284 285 286
    OpKernelKey() = default;
    OpKernelKey(const platform::DeviceContext& dev_ctx) {
      place_ = dev_ctx.GetPlace();
    }

Q
qijun 已提交
287 288 289
    bool operator==(const OpKernelKey& o) const {
      return platform::places_are_same_class(place_, o.place_);
    }
Y
Yu Yang 已提交
290 291 292 293 294 295 296 297 298 299 300
  };

  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
301

302
  void InferShape(const Scope& scope) const {
303 304 305
    InferShape(InferShapeContext(this, scope));
  }

Y
Yu Yang 已提交
306
  void Run(const Scope& scope,
Y
Yu Yang 已提交
307
           const platform::DeviceContext& dev_ctx) const final {
Q
Qiao Longfei 已提交
308
    auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
309
    opKernel->Compute(ExecutionContext(this, scope, dev_ctx));
Q
Qiao Longfei 已提交
310 311
  }

Y
Yu Yang 已提交
312 313 314 315
  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
316
  }
Y
Yan Chunwei 已提交
317

Y
Yu Yang 已提交
318
 protected:
319
  virtual void InferShape(const InferShapeContext& ctx) const = 0;
Q
Qiao Longfei 已提交
320 321 322 323
};

}  // namespace framework
}  // namespace paddle