[FEA] Construct column_view from rmm::device_uvector and device_span

[ttnghia]

Sometimes, we need to call some API that only accepts column_view input but we only have data stored in rmm::device_uvector (which can implicitly casted into device_span). As a result, we need to construct a column_view from rmm::device_uvector manually like this, in many places:

 auto gather_map = column_view(
      cudf::data_type{cudf::type_to_id<cudf::size_type>()}, output_size, indices.data());

This is lengthy and error-prone.

A better solution is to add a utility API to do that automatically. Internally, that utility may call type_dispatcher to construct the output column_view depending on the input type.

[ttnghia]

Tag @jrhemstad and @davidwendt as this idea is from you guys and it's great if you can work on it 😄

[jrhemstad]

A better solution is to add a utility API to do that automatically. Internally, that utility may call type_dispatcher to construct the output column_view depending on the input type.

There's no need for a type dispatch here because you already know the type from the device_span.

[davidwendt]

Make sure you really need this. If you doing this for calling cudf::detail::gather then you can just pass the device_uvector directly to

cudf/cpp/include/cudf/detail/gather.hpp

Lines 78 to 84 in ba2b51d

	std::unique_ptr<table> gather(
	table_view const& source_table,
	device_span<size_type const> const gather_map,
	out_of_bounds_policy bounds_policy,
	negative_index_policy neg_indices,
	rmm::cuda_stream_view stream = rmm::cuda_stream_default,
	rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());

which will automatically convert the device_uvector to a device_span.

Example:

cudf/cpp/src/copying/copy.cu

Lines 162 to 175 in ba2b51d

	auto gather_map = rmm::device_uvector<size_type>{static_cast<std::size_t>(size), stream};
	auto const gather_map_end = thrust::copy_if(rmm::exec_policy(stream),
	thrust::make_counting_iterator(size_type{0}),
	thrust::make_counting_iterator(size_type{size}),
	gather_map.begin(),
	is_left);
	gather_map.resize(thrust::distance(gather_map.begin(), gather_map_end), stream);
	auto const scatter_src_lhs = cudf::detail::gather(table_view{std::vector<column_view>{lhs}},
	gather_map,
	out_of_bounds_policy::DONT_CHECK,
	negative_index_policy::NOT_ALLOWED,
	stream);

[jrhemstad]

I think there are plenty of other places where people are constructing a column or column_view from a device_uvector that make this an important and useful thing.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.