Caffe2 has a concept operator, which corresponds to TensorFlow's Op.

Different from Op, an operator usualy accompany with a gradient operator (GradientOp).

Let us take ReluOp and ReluGradientOp as an example.

All operators are classes derived from Operaotr<Context>.

template <typename T, class Context>
class ReluOp final : public Operator<Context> {
 public:
  USE_SIMPLE_CTOR_DTOR(ReluOp);
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  bool RunOnDevice() override;
};

template <typename T, class Context>
class ReluGradientOp final : public Operator<Context> {
 public:
  USE_SIMPLE_CTOR_DTOR(ReluGradientOp);
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  bool RunOnDevice() override;
};

Operator<Context> has a data member Context context_, which records the current device (or GPU). The constructor of Operator initializes context_ by passing in its constructor a proto message OperatorDef. Then Operator::Operator calls context_.SwitchToDevice(0).

Operator<Context has three virtual functions:

1. RunOnDevice() = 0 is what you want to override,
2. Run(stream_id) calls context_.SwitchToDevice(stream_id), RunOnDevice, and context_.FinishDeviceComputation, and
3. RunAsync calls context_.SwitchToDevice(stream_id) and RunOnDevice.

[TODO: Check what Context::FinishDeviceComputation does.]

Operator<Context> also allows user overriden RunOnDevice to access inputs and outputs through:

• const Tensor<Context>& Input(idx), and
• Tensor<Context>* Output(idx)

Operator<Context> derives from class OperatorBase.