Use merge-path in merge/inplace_merge with host policies

[dmitriy-sobolev]

Most important changes:

1. Redirect dpl::merge to the merge-path algorithm, already used by dpl::ranges::merge.
2. Adapt the merge-path for inplace_merge for both dpl::inplace_merge and dpl::ranges::inplace_merge.
3. Remove now unused backend-level parallel_merge.
4. Handle more edge cases in both algorithms: serial cut offs, sequences are ordered between each other, skipping already ordered head/tail in inplace_merge.

Other:

1. Lift the requirement to have a sized index type in the diagonal search. There was an assert for a signed index type, which depended on a user type - not reliable.
2. Improve stability testing in the merge test. It used to cover only serial cut offs.

---

Performance results

call	CPU	n < 4M	n >= 4M
dpl::merge	GNR (128 thr)	1.38x	3.53x
dpl::merge	PTL (16 thr)	1.12x	1.20x
dpl::inplace_merge	GNR	1.37x	1.34x
dpl::inplace_merge	PTL	1.47x	1.12x

>1x is a speedup.

The results is a geomean of the combinations of these runs:

• Types: int16_t, float, uint64_t, std::string (5-60 characters, lexicographical comparator), std::tuple<int, uint64_t> merged by the first element.
• Sizes: 4K, 16K, 64K, 256K, 1M, 4M, 16M, 67M, 268M total elements n = n1 + n2, 4x per step; std::string up to 16M.
• Distributions:
  • interleaved_equal (|A| = |B| = n/2),
  • interleaved_unequal (|A| = 3n/4, |B| = n/4),
  • non_interleaved (all of A precede all of B),
  • few_unique (|A| = |B| = n/2, only 1% of the keys distinct).

non_interleaved measurements are excluded from the summary for inplace_merge - this is a short-circuit now.

Roofline for dpl::merge:

type	system	main / roofline	PR / roofline
int16_t	GNR	20.54x	1.52x
int16_t	PTL	2.98x	1.44x
float	GNR	11.38x	1.90x
float	PTL	2.21x	1.79x
uint64_t	GNR	5.11x	1.34x
uint64_t	PTL	1.20x	1.17x
std::string	GNR	1.01x	1.01x
std::string	PTL	1.18x	1.19x
std::tuple<int, uint64_t>	GNR	2.05x	1.13x
std::tuple<int, uint64_t>	PTL	1.00x	1.01x

This is a "soft" roofline: tbb::parallel_for with memcpy (trivial types) or std::copy (std::string) and the same chunk size as the merge-path implementation. Geomean of all distributions and >=4M sizes.

---

Investigation notes

I've tried these optimization strategies which have proved to be futile:

• Cache-optimized segmented merge from the merge-path paper.
• Affinity partitioner in inplace_merge.
• Transposing the segments in inplace_merge at 2nd stage to make the sequence to move adjacent at the 3rd stage.
• Blocked read and writes (aka GPU approach).
• Making inplace_merge truly inplace by rotating merged segments to make them adjacent and using a serial inplace_merge later.

[Copilot]

Pull request overview

Redirects host-policy merge operations to the shared merge-path algorithm.

Changes:

• Extracts merge-path intersection logic.
• Uses parallel merge-path partitioning for iterator and ranges APIs.
• Handles empty and bounded output ranges.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
algorithm_ranges_impl.h	Implements bounded ranges merge-path execution.
algorithm_impl.h	Adds shared partitioning and redirects host merge.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

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

[Copilot]

🟡 Changes recommended

The small parallel inplace_merge path no longer preserves required execution-policy exception handling.

Get a fresh assessment by requesting another Copilot review.

Review details

Suppressed comments (1)

include/oneapi/dpl/pstl/algorithm_impl.h:3309

• This serial-cutoff path runs outside __except_handler, so a comparator exception escapes from std::inplace_merge for par/par_unseq. The previous backend path executed the merge inside __except_handler, and execution-policy overloads must terminate on non-bad_alloc exceptions. Wrap this call so the small and large paths preserve the same exception semantics.

        std::inplace_merge(__first, __middle, __last, __comp);

• Files reviewed: 10/10 changed files
• Comments generated: 1
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

New fast paths bypass required host execution-policy exception handling.

Get a fresh assessment by requesting another Copilot review.

Review details

Suppressed comments (1)

include/oneapi/dpl/pstl/algorithm_impl.h:3329

• The new preflight comparator calls, range-narrowing searches, and serial-cutoff merge all execute before the only __except_handler below. Consequently, a comparator exception can propagate from the par/par_unseq overload instead of terminating as required; previously comparator work was performed under the handler. Enclose the nontrivial algorithm body in __except_handler, retaining the empty-range early return if desired.

    if (!__comp(*__middle, *(__middle - 1)))

• Files reviewed: 10/10 changed files
• Comments generated: 2
• Review effort level: Balanced