[maleadt]

oneAPI can't precompile AcceleratedKernels' operations for this extension: the extension isn't loaded while oneAPI precompiles, so kernels cached there would use the default tunings rather than the ones set here (predicate_tuning even selects different kernels for any/all).

When oneAPI can use the GPU while precompiling, which it reports with oneAPI.can_precompile (JuliaGPU/oneAPI.jl#753), this runs sum of a vector and along each dimension of a matrix, cumsum, any, all and an Int sum in a @compile_workload of the extension, at sizes that select the same kernels as a vector of 10^6 elements and a 1000×1000 matrix (checked by recording the kernels compiled at both sizes). With oneAPI versions without can_precompile the workload is skipped. Like oneAPI's own workload it needs Julia 1.12.

On an Arc A770 with Julia 1.13, together with that oneAPI PR, the first sum of 10^6 elements takes 0.17 instead of 1.2 s, sum along both dimensions 0.25 instead of 2.0 s, cumsum 0.41 instead of 2.2 s and all 0.25 instead of 1.7 s. The extension's image grows by 11 MB, its precompilation from 1.4 to 6.9 s, and using oneAPI takes about 80 ms longer. Most of what remains is oneAPI compiling the kernels' SPIR-V in every session, as it doesn't persist compiled kernels.