Created by: qingqing01

The following comments are addressing three problems:

• Necessary change to LoDTensor
• Replace uses of Tensor into LoDTensor in operators who don't need LoD
• How to pass LoD from operators' inputs to outputs

LoDTensor

The definition of LoDTensor is as follows:

class LodTensor {
 public:
  LoDTensor() {}
  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
 private
  Tensor* tensor_;
  Lod lod_;
};

The code snippet shows how we use it in operators:

void InferShape(const framework::InferShapeContext &ctx) const override {
    auto tensor_x = ctx.Input<LoDTensor>("X")->tensor();
    auto tensor_y = ctx.Input<LoDTensor>("Y")->tensor();
    PADDLE_ENFORCE_EQ(tensor_x.dims(),
                      tensor_y.dims(),
                      "Two inputs of Add Op's dimension must be same.");
    auto* lod_tensor_out =   ctx.Output<LodTensor>("Out");
    Tensor* out = new Tensor();
    out->Resize(tensor_x->dims());
    lod_tensor_out->set_tensor(out);
  }
};

We see that LoDTensor::set_tensor is kind of ugly because the caller needs to destruct the tensor at right time.

A solution is to make LoDTensor owns the tensor and delete it at destruction time:

class LodTensor {
 public:
  LoDTensor() : tensor_(new Tensor()) {}
  ~LoDTensor() { delete tensor_; }
  //  删掉set_tensor接口。
  // void set_tensor(Tensor* tensor) { tensor_ = tensor; } 
 private
  Tensor* tensor_;
  Lod lod_;
};