Bug in "paddle/fluid/framework/tensor_impl.h" and inefficient code in "paddle/fluid/framework/reader.cc"

[sneaxiy]

Bug in paddle/fluid/framework/tensor_impl.h

In functions const T* Tensor::data() const and T* Tensor::data(), the inner type holder_ holds must be the same as T. It is implemented by comparing hash codes in the current version, i.e., holder_->type().hash_code() == typeid(T).hash_code().

However, C++ does not guarantee the hash codes of different types must be different. It is recommended to change it to be holder_->type() == std::type_index(typeid(T)).

Inefficient code in paddle/fluid/framework/reader.cc

Inside the first for loop of function void FileReader::ReadNext(std::vector<LoDTensor> *out), std::vector<LodTensor>::at is called to access the data in the read-out result out. Therefore, std::vector<LodTensor>::at would check the size bound of out during each loop and thus it is inefficient.

A better choice is to add PADDLE_ENFORCE_EQ to compare out->size() to dims_.size() before the loop, rather than check the bounds in each loop.