Partition Set Algorithm Balanced Path

[danhoeflinger]

Add partitioning kernel to set APIs balanced path algorithm.

Adds a partitioning phase which does a sparse pass over the input data to establish binary search boundaries for the main run. This allows memory access pattern to fit within L1 cache for the main kernels when performing the binary searches to establish balanced path intersections.

This improves performance for large sizes of the set algorithms. (When combined with #2317, it provides a nice combination of performance improvements for both large and small sizes of the set algorithms)

[Copilot]

Pull Request Overview

This PR introduces a partitioning phase to the set algorithms, enhancing performance for large input sizes by establishing binary search boundaries that optimize cache usage. Key changes include updating the __gen_set_balanced_path template to accept an additional bounds provider parameter, adding new helper functions (__decode_balanced_path_temp_data, __encode_balanced_path_temp_data) for balanced path processing, and integrating a new partition kernel for the balanced path phase.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
test/general/implementation_details/device_copyable.pass.cpp	Updated static_asserts to include new bounds provider parameter
include/oneapi/dpl/pstl/hetero/dpcpp/sycl_traits.h	Modified __gen_set_balanced_path specialization to include _BoundsProvider
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h	Added helper functions & modified balanced path computation to include partitioning support and safeguard against out-of-bound element access
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h	Updated __parallel_set_reduce_then_scan integration with new bounds provider and partitioning kernel

Comments suppressed due to low confidence (4)

include/oneapi/dpl/pstl/hetero/dpcpp/sycl_traits.h:458

• The specialization of __gen_set_balanced_path now includes the _BoundsProvider parameter; please verify that all downstream usages are updated accordingly to maintain consistent API behavior.

template <typename _SetOpCount, typename _BoundsProvider, typename _Compare>

include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h:1086

• [nitpick] When constructing _GenReduceInput with the new _BoundsProvider, it would be helpful to document the role of __diagonal_spacing and __partition_size in determining partition sizes, ensuring that readers understand how these values impact performance.

                                       _BoundsProvider{__diagonal_spacing, __partition_size}, __comp};

include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h:698

• The change from returning 0 to clamping __i_elem to __rng1.size() + __rng2.size() - 1 may affect the algorithm's edge-case handling; please confirm that this adjusted behavior correctly reflects the intended semantics.

        if (__i_elem >= __rng1.size() + __rng2.size())

include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h:821

• [nitpick] Retrieving __tile_size from __gen_input.__get_bounds is critical for partitioning; please ensure that __tile_size is always correctly initialized and consistent for various input sizes to avoid unexpected partition boundaries.

        std::size_t __tile_size = __gen_input.__get_bounds.__tile_size;

[mmichel11]

I have made a first pass over the implementation. I like how the patch unifies handling between the partitioned and non-partitioned bounds.

[mmichel11]

This PR looks near ready to me. The conflicts with main need to be addressed.

[mmichel11]

No necessary action on your side here, but I see we use something along the lines of decltype(__rng.size())) frequently in this patch. It looks like we have a oneapi::dpl::__internal::__range_size_t but only for C++20 or later. Perhaps it would be worth extending this for C++17 with internal ranges.

[mmichel11]

LGTM

[mmichel11]

Reapproving

[SergeyKopienko]

LGTM

[SergeyKopienko]

LGTM