在OneFlow实现数据类型自动提升

x_tensor = torch.ones((3, ), dtype=torch.int8)
y1_tensor = torch.tensor(1, dtype=torch.float64)
out1 = torch.mul(x_tensor, y1_tensor)

y2_tensor = torch.tensor(1, dtype=torch.int64)
out2 = torch.mul(x_tensor, y2_tensor)

out3 = torch.mul(x_tensor, 1.0)

out4 = torch.mul(x_tensor, 2^63-1(the max value of int64))

out1.dtype: torch.float64
out2.dtype: torch.int8
out3.dtype: torch.float32
out4.dtype: torch.int8

x_tensor = torch.ones((3, ), dtype=torch.int16)
out1 = x_tensor + 2 # out.dtype = torch.int16
out2 = x_tensor + 2.0 # out.dtype = torch.float32

x_tensor = torch.ones((3, ), dtype=torch.int16)
y1_tensor = torch.tensor(2)
y2_tensor = torch.tensor(2.0)

out1 = x_tensor + y1_tensor # out.dtype = torch.int16
out2 = x_tensor + y2_tensor # out.dtype = torch.float32

import numpy as np

x = np.ones((3, 3), dtype=np.int32)
out = x + (2**31-1) # dtype: np.int32
out = x + (2**31) # dtype: np.int64

RuntimeError: dot : expected both vectors to have same dtype, but found Short and Float

TORCH_META_FUNC(threshold)(const Tensor& self, const Scalar& threshold, const Scalar& value) {
  const Tensor& result = maybe_get_output();
  build(TensorIteratorConfig()
    ...
    .promote_inputs_to_common_dtype(true)
}

...
compute_types(config);
...

      // Promotes inputs by creating temporaries of the correct dtype
      if (config.promote_inputs_to_common_dtype_ && !op.is_output && op.current_dtype != common_dtype_) {
        op.original_tensor = op.tensor;
        op.tensor = c10::MaybeOwned<Tensor>::owned(op.tensor->to(common_dtype_));
        op.current_dtype = common_dtype_;
        op.target_dtype = common_dtype_;
      }

class TensorProcessor final {
 public:
  TensorProcessor()
      : common_dtype_(DType::InvalidDataType()), promote_inputs_to_common_dtype_(false){};
  TensorProcessor& AddInputs(const TensorTuple& init_list);
  TensorProcessor& AddInputs(const TensorTuple& init_list, Symbol<DType> tensor_lowest_dtype);

  Maybe<void> Apply();
  TensorProcessor& PromoteInputsToCommonDtype(bool is_promote);
  Maybe<TensorTuple&> GetInputs() { return tensor_tuple_; };

 private:
  TensorTuple tensor_tuple_;
  Symbol<DType> common_dtype_;
  std::vector<Symbol<DType>> inputs_lowest_dtype_vec_;

  bool promote_inputs_to_common_dtype_;
};

class BinaryFunctor{
 public:
  Maybe<Tensor> operator()(const std::shared_ptr<one::Tensor>& x,
                           const std::shared_ptr<one::Tensor>& y) const {
    TensorProcessor tensor_processor;
    JUST(tensor_processor.PromoteInputsToCommonDtype(true).AddInputs({x, y}).Apply());
    TensorTuple input_tuple = JUST(tensor_processor.GetInputs());
    return OpInterpUtil::Dispatch<Tensor>(*op_, input_tuple);
  ...
  }
  ...
}; 

Maybe<void> CastToSameType(TensorTuple& tensor_tuple, const Symbol<DType>& common_dtype) {
  for (auto& tensor_ptr : tensor_tuple) {
    if (tensor_ptr->dtype() != common_dtype) {
      tensor_ptr = JUST(functional::Cast(tensor_ptr, common_dtype));
    }
  }
  return Maybe<void>::Ok();
}

Maybe<void> TensorProcessor::Apply() {
  if (promote_inputs_to_common_dtype_) {
    bool has_different_input_dtype = CheckHasDifferentInputDType(tensor_tuple_);
    if (has_different_input_dtype) {
      common_dtype_ = ComputeCommonDType(tensor_tuple_);
      JUST(CastToSameType(tensor_tuple_, common_dtype_));
    }
  } else {
    for (int i = 0; i < tensor_tuple_.size(); ++i) {
      // Cast all the inputs to it's attribute `lowest_dtype` if the input tensor dtype is lower
      // than attribute `lowest_dtype`.
      Symbol<DType> base_dtype = inputs_lowest_dtype_vec_.at(i);
      if (base_dtype->data_type()
          && DType::priority_order[base_dtype->data_type()]
                 > DType::priority_order[tensor_tuple_.at(i)->dtype()->data_type()]) {
        tensor_tuple_.at(i) = JUST(one::functional::Cast(tensor_tuple_.at(i), base_dtype));
      }
    }
  }
  return Maybe<void>::Ok();
}

class ReduceSumFunctor{
public: 
  Maybe<Tensor> operator()(const std::shared_ptr<one::Tensor>& x, const std::vector<int32_t>& axis,
                           const bool& keepdims) const {
    ...
    TensorProcessor tensor_processor;
    JUST(tensor_processor.AddInputs({x}, /*lowest_dtype=*/DType::Int64()).Apply());
    TensorTuple input_tuple = JUST(tensor_processor.GetInputs());
  }
  ...
};