[AzamatB]

Since GPUArrays 12, Base's any, all and findall of GPU arrays go through AK, and on main a warm call still runs inference every time:

julia> f(x) = any(>(0), x);   # x::CuVector{Int32}, 10_000 elements

julia> @time f(x)
  0.000113 seconds (62 allocations: 2.828 KiB, 0.52% compilation time: 100% of which was recompilation)

any, all, count, findall and most of the internal methods they call only pass the predicate on, so Julia compiles them for pred::Function. In that code, check_bool_result's Base.promote_op(pred, eltype(v)) can't fold and runs inference on each call. This PR annotates the predicate as pred::F ... where {F} in those methods, as the reductions already do with f::F, op::OP. A warm call then compiles nothing, and allocates about half as much on the host: on an RTX 5080, any(>(0), x) goes from 63 allocations (2.9 KB) and a median of 12.3–12.7 µs to 48 allocations (1.5 KB) and 11.6–11.9 µs, and AK.any(isnan, x) on 10,000 Float32s from 12.1–12.5 µs to 11.4–11.6 µs.

Separately, since #189 the tunings are looked up behind Base.inferencebarrier. Auto() picks a GPU's predicate algorithm, ConcurrentWrite or ViaReduce, from the tuning, so the plan doesn't infer, and neither does the result: Base.infer_return_type(all, Tuple{typeof(isfinite), CuVector{Float64, CUDA.DeviceMemory}}) is Any on main, and so is every caller that returns or uses the result, e.g. checkbounds(Bool, x, i::CuVector{Int32}). Hiding the tuning is intended (it avoids invalidations when loading a backend), so any and all now assert that their result is a Bool, which every algorithm returns. Run-time dispatch of the algorithm stays. The other operations that use tunings are not affected: reductions, scans, sorts, count and findall still infer with Auto() on CUDA, because their tuning only picks values or the algorithm's type doesn't change the result type.

The tests check that a warm call of any, all and findall by dynamic dispatch compiles nothing, and that any and all infer with each algorithm, including Auto(). Both fail on main (the second only on GPUs). The CPU tests (--cpu --cpu-ka) and the CUDA tests for the predicates, findall and the reductions pass on Julia 1.12, and the new tests also pass on 1.10.

sort!'s by has the same problem (promote_op(by, ...) in the sort setup still runs inference on each call). That's left for a separate PR.

[maleadt]

Thanks. Rebased and slightly reworked.