support GpuSize

[sperlingxx]

This PR is to support GpuSize.
The intention is to fix test failures on GpuExplode under spark 3.1. The root cause of the failure is that in spark 3.1 newly added rule InferFiltersFromGenerate will insert a group by expressions include GPU unsupported expression Size.

Signed-off-by: sperlingxx [email protected]

[sperlingxx]

build

[revans2]

The code itself looks fine, but I just would prefer for us to do the right thing instead of putting in a hack like this.

[revans2]

We need to removed UDT from this for the non-spark use case. This is just to make the docs correct, because we don't support UDT anywhere.

[revans2]

Lets just add in JNI for this instead.

https://github.com/rapidsai/cudf/blob/6fe8b1b7be71c0f3587d3ef43c535aec376d1e45/cpp/include/cudf/lists/count_elements.hpp#L49

It does everything we want, except for the legacy null behavior, and we can fix that up without too much work.

var ret = intput.getBase.countElements()
if (legacySizeOfNull) {
    withResource(input.getBase.isNull()) { isNull =>
        withResource(GpuScalar.from(-1)) { negOne =>
            withResource(ret) { count =>
                ret = isNull.ifElse(negOne, count)
            }
        }
    }
}

[revans2]

I missed the part about this is blocking issues with the nightly tests. Lets merge this in as is and have a follow on to make ti work correctly. I'll get a PR up for the JNI code ASAP so we can get a real fix in soon too.

[revans2]

I filed #1974 as a follow on issue for this.