Set algorithms, revert to set_a_write algorithm for small sizes

[danhoeflinger]

This PR makes some infrastructure changes to allow a runtime decision based upon number of elements to select the set algorithm selected, in addition to limitations based upon platform and environment.

1. Relocate algorithm choice from hetero level to sycl backend and make __parallel_set_op blocking. (temporarily required for type alignment, see [Draft] RFC future type and keepalives #2300)
2. Create sycl backend entry points to call "merge" (__parallel_merge_impl) and "copy" (__parallel_copy_impl).
3. Add a threshold to switch algorithms from set_a_scan to balanced_path based upon set tag in __set_op_impl using __consider_write_a_alg

This results in an overall performance improvement, as small numbers of input elements do not perform as well with the balanced path algorithm.

edit: this incorporates a fix from #2386 for the reduce_then_scan temporary buffer for set_a_write

[Copilot]

Pull Request Overview

This PR reworks the set algorithms in the SYCL backend to dynamically select between different implementation strategies based on input sizes. Key changes include removing the boolean template parameter from the set tag types, updating call sites and helper functions to use non-templated tags, and introducing a threshold switch in __set_op_impl to choose between the write_a_scan and balanced_path algorithms.

Reviewed Changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
test/general/implementation_details/balanced_path_unit_tests.pass.cpp	Removed instantiations with std::true_type from set tag calls in tests
include/oneapi/dpl/pstl/hetero/dpcpp/unseq_backend_sycl.h	Dropped the _IsOneShot template parameter from set tag types and updated __brick_set_op accordingly
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h	Updated comments and template parameters (renaming to _SetTag) to align with new tag type definitions
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h	Adjusted __parallel_merge overloads to remove extraction of _ExecutionPolicy and use a custom name parameter
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_for.h	Updated parallel_for overloads to use an explicit queue parameter and _CustomName for kernel naming
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h	Refactored the __set_op_impl and related functions to support algorithm switching based on input size
include/oneapi/dpl/pstl/hetero/algorithm_impl_hetero.h	Modified set algorithm patterns to remove the boolean tag instantiation and use the new tag types

Comments suppressed due to low confidence (3)

include/oneapi/dpl/pstl/hetero/dpcpp/unseq_backend_sycl.h:1230

• Since the _IsOneShot template parameter and the associated __can_write_from_rng2_v flag have been removed from the set tag types, please add a brief comment explaining that the write capability is now determined elsewhere for clarity and maintainability.

struct _IntersectionTag

include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h:532

• With the introduction of the _CustomName parameter and the new overloads for __parallel_merge, please add inline documentation to clarify when each overload should be used and how _CustomName is derived from the execution policy.

template <typename _ExecutionPolicy, typename _Range1, typename _Range2, typename _Range3, typename _Compare,

include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl.h:1332

• The magic number used as a threshold for switching algorithms (1024 * 1024) could benefit from an inline comment that explains its rationale and expected impact on performance, to improve code maintainability.

    static constexpr std::size_t __threshold = 1024 * 1024;

[mmichel11]

Only minor comments from me.

[danhoeflinger]

I fixed the rebase with main, Now the diff reasonable is reasonable. It was previously clashing with a partial implementation of the partitioning PR which made the diff messy and incorrect.

I have found a same-mangled name issue with set_union and set_symmetric_difference I am investigating.
Edit: this is now resolved.

[SergeyKopienko]

I have found a same-mangled name issue with set_union and set_symmetric_difference I am investigating.

@danhoeflinger I can help you with this topic.

[danhoeflinger]

I've resolved the same mangled name issue.

[mmichel11]

I have done another pass over this PR, and it LGTM assuming green CI.

[mmichel11]

Reapproving after bugfix.

[SergeyKopienko]

LGTM