[maleadt]

Implements KernelInterface's sub-group API (JuliaGPU/KernelAbstractions.jl#831) for OpenCL.jl. Stacked on #519.

• Shuffles (shfl, shfl_up, shfl_down, shfl_xor) for the types sub_group_shuffle supports, with cl_khr_subgroup_shuffle. They're overridden only for those types, so other types reach KernelInterface's fallbacks, which shuffle them as UInt32 words. shfl_down/shfl_up return the work-item's own value past the width, also for very large offsets.
• Float16/Float64 aren't reported as shuffleable on devices without cl_khr_fp16/cl_khr_fp64. A kernel can't even move such values there: GPUCompiler rejects the IR after LLVM turns an integer load that is reinterpreted as a float into a float load.
• Votes. The ballot uses cl_khr_subgroup_ballot where available, and otherwise two 32-bit sub_group_reduce_adds.
• Native sub_group_reduce/sub_group_scan/sub_group_exclusive_scan for + (integers and floats), and min/max on integers only, because OpenCL's float min/max don't propagate NaN or order signed zeros the way Julia does.
• Sub-groups are still only reported with cl_intel_required_subgroup_size, which kernels need to get a fixed width. They're now also detected when the device reports them as an OpenCL 3.0 core feature, and only widths of at most 64 are reported, so the ballot has a bit for every lane. Requiring the extension is an initial limitation; other ways to fix the width can be added later.
• supports_linear_subgroups: PoCL and Intel's CPU and GPU runtimes, all of which pass KernelInterface's layout tests.
• supports_independent_subgroups: false on PoCL, whose sub-group operations are work-group barriers. It's also false on Intel's CPU runtime, which loses the writes after a sub_group_barrier that other sub-groups returned before ([OpenCL CPU Runtime] Writes after a sub-group OpControlBarrier are lost when other sub-groups returned early intel/llvm#23449). Intel's GPU runtime passes the independence test.
• supports_partitionable_subgroups: only on Intel's GPU runtime with cl_khr_subgroup_ballot, where a sub-group is the SIMD lanes of one EU thread and the shuffles and ballot of the width functions only involve the active lanes. OpenCL has no barrier for part of a sub-group, so a new :partitionable_sub_groups device feature makes KI.sub_group_barrier(width) a sub-group-scope memory barrier there for segments narrower than the sub-group. Elsewhere (PoCL, Intel's CPU runtime, unknown runtimes) it's false, and KernelAbstractions' tiles aren't available on PoCL and Intel's CPU runtime, whose sub-groups aren't independent.
• cl.sub_group_size returns 0 for devices without sub-groups (fixes cl.sub_group_size doesn't return 0 for devices without sub-groups #553).

The TEMP commit gets KernelAbstractions and KernelInterface from KA's vc/ki-subgroup-ops branch.

Tested locally, KI and KA testsuites, Julia 1.12 (and 1.10 on PoCL and Intel's CPU runtime):

• PoCL (pocl_jll 7.2.2): all pass.
• PoCL via OpenCL C (poclc): KA passes. KI's sub-group tests fail with UNIMPLEMENTED instruction 262/264 and UNIMPLEMENTED capability 64, because spirv2clc doesn't translate the votes, collectives and ballot yet. These failures, and the same ones on 'PoCL next' (arm64), which also goes through spirv2clc, go away with Untranslated sub-group instructions: OpGroupAll/Any/Broadcast, OpGroupIAdd and other group arithmetic, GroupNonUniformBallot spirv2clc#25.
• Intel's GPU runtime (Iris Xe, NEO_jll): KI 5730 pass. KA 2627 pass, 4 broken.
• Intel's CPU runtime: KI passes. The KA failures (3-D launches, Intel's CPU runtime skips the partial last work-group along x in N-d launches #521) also happen without this PR.
• NVIDIA (no sub-groups): KI passes. The KA @print failure also happens without this PR.

[codecov]

Codecov Report

❌ Patch coverage is 94.64286% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.06%. Comparing base (e867580) to head (9c807cf).

Files with missing lines	Patch %	Lines
src/OpenCLKernels.jl	94.44%	3 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           ka-0.10     #554      +/-   ##
===========================================
- Coverage    87.13%   87.06%   -0.08%     
===========================================
  Files           18       18              
  Lines         1664     1716      +52     
===========================================
+ Hits          1450     1494      +44     
- Misses         214      222       +8     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.