aten_floor_divide_int may cause precision loss

[wodesuck]

Current implement will first cast inputs to float, do a true divide then floor. If inputs are big int64, it may case precision loss.

@torch_op("aten::floor_divide", trace_only=True)
def aten_floor_divide_int(self: TInt, other: TInt) -> TInt:
    """floor_divide(Tensor self, Tensor other) -> Tensor"""

    # TODO(justinchuby): This can be simplified if we can constrain the
    # inputs to be positive integers. Consider how we can embed constraints in the model.
    dtype = self.dtype
    self = op.Cast(self, to=FLOAT.dtype)
    other = op.Cast(other, to=FLOAT.dtype)
    result = op.Floor(op.Div(self, other))
    return op.Cast(result, to=dtype)

[wodesuck]

Maybe we can refer to the implementation in here.

[justinchuby]

Looks like we can use Sign which is simpler: https://github.com/pytorch/pytorch/blob/ffc645c870f0abd368606ba1e2b3b58cacb03046/torch/_refs/__init__.py#L1401C1-L1409C70