brandon-b-miller
commented
3 years ago I think there can be three scenarios when astype is called on a cudf scalar s:
shost value is present, and its device value is not. This returns a new python object with a new host value, but still no device value.- both the host and device values are present for
s. In this case, we cast the host value, and construct a new cudf scalar from the result. The question is should the device value be set. My initial thought is that there's no reason to set it until it's actually needed, and.device_valueis accessed. sdevice value is present, and the host value is not. Here I think we need to move the device value to host, cast it there and construct a new scalar again. The question is again should the device value be set - and I lean towards no for the same reasons as above - it shouldn't harm anything to leave it unsynchronized until it's used.
Most of the time we'll obtain a numpy scalar and we can just use numpy casting to get the new scalar. But we'll need special handling for strings and decimals.
github-actions[bot]
What is the expected behavior of
Scalar.astype?During implementing https://github.com/rapidsai/cudf/pull/7182, I realized
cudf.Scalar.astype's behavior is not well-defined. Since cudf Scalar has a host-side and a device-side, which oneastypeshould operate on and what is the expected behavior? In addition, device side scalar is not really operable since libcudf does not cast scalars types. How to resolve for the discrepancy?cc @kkraus14 @brandon-b-miller