[maleadt]

Each KernelAbstractions launch calls device_limits several times, each requiring task-local storage and Dict lookups. Store the limits in a global Base.ImmutableDict, replaced under a lock when adding a device. Cached lookups need no lock. __init__ clears the cache because its keys are driver handles. In KI.max_work_group_size(kernel), query device limits only when the kernel's limit is unknown; coalesce evaluated both arguments.

Host time per launch in batches of 64 launches without synchronization. Arc A770, Julia 1.12.7, KernelAbstractions 0.10.3; median of 3 fresh processes:

	main	PR
@oneapi	2294 ns	2298 ns
KA, 1D, ndrange=1000	3103 ns	2938 ns
KA, 1D, static group size 256	2928 ns	2727 ns
KA, 3D, 30×30×30	3833 ns	3548 ns

The full test suite passes on an Arc A770 (Julia 1.12.7).

[codecov]

Codecov Report

❌ Patch coverage is 93.33333% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 86.65%. Comparing base (446f8ac) to head (03afc1a).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/oneAPIKernels.jl	92.85%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #747      +/-   ##
==========================================
- Coverage   86.89%   86.65%   -0.25%     
==========================================
  Files          55       55              
  Lines        5090     5005      -85     
==========================================
- Hits         4423     4337      -86     
- Misses        667      668       +1     

☔ 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.