operator.h 5.0 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

Y
Yu Yang 已提交
17 18 19 20 21 22
#include <paddle/framework/attr_checker.h>
#include <paddle/framework/op_desc.pb.h>
#include <paddle/framework/scope.h>
#include <paddle/platform/device_context.h>
#include <paddle/platform/place.h>
#include <paddle/utils/Error.h>
Q
Qiao Longfei 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
#include <boost/variant.hpp>
#include <string>
#include <unordered_map>
#include <vector>

namespace paddle {
namespace framework {

class OperatorBase;

/**
 * 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:
  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));
  }

  std::string DebugString() const;

Q
Qiao Longfei 已提交
52 53 54 55
  /// Init will be called after CreateOperator, you can put some initialization
  /// logic here.
  virtual void Init() {}

Q
Qiao Longfei 已提交
56 57 58 59 60 61
  /// InferShape infer the size of Variables used by this Operator with
  /// information inside scope
  virtual void InferShape(const std::shared_ptr<Scope>& scope) const = 0;

  /// Net will call this function to Run an op.
  virtual void Run(const std::shared_ptr<Scope>& scope,
Y
Yu Yang 已提交
62 63 64 65
                   const platform::DeviceContext& dev_ctx) const = 0;

 protected:
  std::string Type() const { return desc_.type(); }
Q
Qiao Longfei 已提交
66 67 68 69 70 71 72 73

 public:
  OpDesc desc_;
  std::vector<std::string> inputs_;
  std::vector<std::string> outputs_;
  AttributeMap attrs_;
};

Y
Yu Yang 已提交
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105
class OpKernel {
 public:
  /**
   * KernelContext is the only parameter of Kernel Run function.
   * Run will get input/output variables, state such as momentum and
   * device resource such as CUDA stream, cublas handle, etc. from
   * KernelContext. User should construct it before run the Operator.
   */
  class KernelContext {
   public:
    KernelContext(const OperatorBase* op, const std::shared_ptr<Scope>& scope,
                  const platform::DeviceContext& device_context)
        : op_(*op), scope_(scope), device_context_(device_context) {}

    const Variable* Input(int index) const {
      return scope_->GetVariable(op_.inputs_[index]);
    }

    Variable* Output(int index) const {
      return scope_->GetVariable(op_.outputs_[index]);
    }

    const OperatorBase& op_;
    const std::shared_ptr<Scope>& scope_;
    const platform::DeviceContext& device_context_;
  };

  virtual void Compute(const KernelContext& context) const = 0;

  virtual ~OpKernel() {}
};

Q
Qiao Longfei 已提交
106 107
class OperatorWithKernel : public OperatorBase {
 public:
Y
Yu Yang 已提交
108 109
  struct OpKernelKey {
    platform::Place place_;
Q
Qiao Longfei 已提交
110

Y
Yu Yang 已提交
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127
    OpKernelKey() = default;
    OpKernelKey(const platform::DeviceContext& dev_ctx) {
      place_ = dev_ctx.GetPlace();
    }

    bool operator==(const OpKernelKey& o) const { return place_ == o.place_; }
  };

  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
128 129

  void Run(const std::shared_ptr<Scope>& scope,
Y
Yu Yang 已提交
130 131 132
           const platform::DeviceContext& dev_ctx) const final {
    auto& opKernel = AllOpKernels().at(Type()).at(OpKernelKey(dev_ctx));
    opKernel->Compute(OpKernel::KernelContext(this, scope, dev_ctx));
Q
Qiao Longfei 已提交
133 134
  }

Y
Yu Yang 已提交
135 136 137 138 139
  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;
  };
Q
Qiao Longfei 已提交
140 141 142 143
};

}  // namespace framework
}  // namespace paddle
Y
Yu Yang 已提交
144 145 146 147 148 149 150 151 152 153 154

#define REGISTER_OP_KERNEL(type, PlaceType, KernelType)                   \
  struct __op_kernel_register__##type##__ {                               \
    __op_kernel_register__##type##__() {                                  \
      ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
      key.place_ = PlaceType();                                           \
      ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
          .reset(new KernelType());                                       \
    }                                                                     \
  };                                                                      \
  static __op_kernel_register__##type##__ __reg_kernel_##type##__