Merge Tensor and LoDTensor class

[tonyyang-svail]

In protocol buffer, tensor and lodtensor are of the same type.

Paddle/paddle/framework/framework.proto

Lines 118 to 126 in e445b3f

	enum VarType {
	LOD_TENSOR = 1;
	SELECTED_ROWS = 2;
	FEED_MINIBATCH = 3;
	FETCH_LIST = 4;
	STEP_SCOPES = 5;
	LOD_RANK_TABLE = 6;
	LOD_TENSOR_ARRAY = 7;
	}

This consistency should stay true at the class level.

[jacquesqiao]

why do we need to merge these two classes or Why do we need to keep these two classes?

[tonyyang-svail]

@jacquesqiao In the current implementation:

Paddle/paddle/operators/mul_op.h

Line 30 in 61ec0b9

const Tensor* x = context.Input<Tensor>("X");

Calls

Paddle/paddle/framework/operator.h

Lines 314 to 315 in e445b3f

	template <>
	const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const;

Calls

Paddle/paddle/framework/operator.cc

Lines 202 to 206 in e445b3f

	template <>
	const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
	auto* var = InputVar(name);
	return var == nullptr ? nullptr : GetTensorFromVar(var);
	}

Calls

Paddle/paddle/framework/operator.cc

Lines 178 to 188 in e445b3f

	static const Tensor* GetTensorFromVar(const Variable* var) {
	const Tensor* t = nullptr;
	if (var->IsType<LoDTensor>()) {
	t = &(var->Get<LoDTensor>());
	} else if (var->IsType<SelectedRows>()) {
	t = &(var->Get<SelectedRows>().value());
	} else {
	PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
	}
	return t;
	}

Meaning, even though context.Input<Tensor>("X"); is called on the surface, deep down the code the type returned is LoDTensor. This is fundamentally wrong.

We have to remove Tensor class so that the code is less error-prone. For example, let's say someone calls var->Mutable<Tensor>, but the underlying type is LoDTensor, the call would return a brand new tensor. This is VERY likely since we set LoDTensor as a default varType.