Optimize scan KT and support user defined binary operators

[mmichel11]

This PR reworks our scan kernel template by switching from an SLM driven scan where all data is loaded and operated in SLM to a sub-group driven scan where data is primarily stored and operated on in hardware registers. Additionally, this PR adds the building blocks necessary to implement future scan-like kernel templates.

The following changes are made:

• internal/cooperative_lookback.h is introduced to separate the lookback implementation from inclusive scan. We additionally provide an optimized path based on bit packing status and prefix flags for certain types.
• internal/work_group/work_group_scan.h and internal/sub_group/sub_group_scan.h implement optimized work-group and sub-group level scans, enabling support for user defined operators beyond SYCL function objects with known identities. This implementation and API may be used to propose improvements to SYCL itself as it has been shown to perform significantly better.
• The reworked implementation also serves as a single work-group scan without any extra kernels compiled. At runtime we determine if a single tile is present and skip temporary storage allocation and the init kernel if so. The existing single work-group path is removed as it does not perform better.

The initial implementation only supported SYCL function objects with known identities. User defined operators are now supported so long as the type satisfies the following conditions:

• Scan input type must be convertible to output type
• Type must be copy constructible / assignable.

[Copilot]

Pull Request Overview

This PR reworks the scan kernel template implementation by transitioning from an SLM-driven approach to a sub-group-driven approach that primarily operates on hardware registers. It also adds support for user-defined binary operators and optimizes the implementation.

Key changes:

• Introduces modular cooperative lookback implementation with bit packing optimizations
• Implements optimized work-group and sub-group level scans with register-based operations
• Adds support for custom binary operators beyond SYCL function objects with known identities

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
test/kt/single_pass_scan.cpp	Adds custom binary operator test (my_bit_xor) to verify user-defined function object support
test/kt/CMakeLists.txt	Expands test coverage by adding uint8_t and uint16_t data types and additional test configurations
include/oneapi/dpl/experimental/kt/single_pass_scan.h	Core implementation rewrite - switches to sub-group driven scan with register-based operations and modular lookback
include/oneapi/dpl/experimental/kt/internal/work_group/work_group_scan.h	New optimized work-group scan implementation using sub-group level operations
include/oneapi/dpl/experimental/kt/internal/sub_group/sub_group_scan.h	New sub-group scan implementation operating on register arrays in strided layout
include/oneapi/dpl/experimental/kt/internal/cooperative_lookback.h	Extracted cooperative lookback implementation with specialized bit-packed storage for small types

Comments suppressed due to low confidence (1)

include/oneapi/dpl/experimental/kt/single_pass_scan.h:314

• There is an extra closing brace that doesn't match any opening brace. This will cause a compilation error.

    }

[adamfidel]

Made it through cooperative_lookback.h. I still need to review the rest.

[adamfidel]

I've gone through the changes and I don't see any glaring issues. I just have a few nitpicks but I think it's in a good shape to merge.

[adamfidel]

Can you move this comment to line 128 (right before the direct write)? I think it would be clearer what it is referring to. Same for the set_init function

[mmichel11]

Yes, done.

[adamfidel]

Suggested change

	std::size_t __num_tiles;
	_TileIdxT __num_tiles;

The tile index is 32-bits, so we can probably restrict the size of the number of tiles here.

[mmichel11]

Done, I had to change a few other types as well.

[danhoeflinger]

Overall this looks good. If we fix the outdated comment mentioning default initialized value type #2357 (comment), and either fix or make an issue for the zeroth tile performance issue #2357 (comment), then I'm ready to approve.

The rest are nitpicks which you can apply or ignore.

[danhoeflinger]

LGTM with green CI. Thanks for the quick responses.