Complete the GPU extensions: atomic get/set, swap, float min/max

[vchuravy]

The CUDA, oneAPI, OpenCL and Metal extensions implemented Atomix.get and Atomix.set! as error("not implemented"), and modify! errored for any operation without a listed backend intrinsic. So inside kernels @atomic A[i], @atomic A[i] = v, @atomicswap A[i] = v (which is modify! with right) and float @atomic max(A[i], x) / min all threw — silently on backends that do not surface device exceptions (pocl leaves the array untouched), and with a MethodError on CUDA, whose atomic_max!/atomic_min! only exist for integers. This surfaced in JuliaGPU/KernelAbstractions.jl#726 and #737.

Changes

• get / set!: the overrides are removed, so Atomix's generic implementation applies: an LLVM atomic load/store on the device pointer through UnsafeAtomicsLLVM, which honors the requested ordering. Metal uses its atomic_load_explicit / atomic_store_explicit intrinsics instead, since Metal.jl does not lower LLVM atomics.
• Swap: modify! with right dispatches to the backend's exchange intrinsic. On CUDA, floats are exchanged through their integer representation (atomic_xchg! is integer-only there).
• No native instruction (float min/max where the backend lacks them, arbitrary functions such as *): a compare-and-swap loop on the backend's CAS. It compares bitwise (===), so a stored NaN or a signed zero cannot make it spin. On OpenCL/oneAPI the float min/max intrinsics are used when the installed SPIRVIntrinsics provides them (checked with hasmethod at extension load), otherwise the loop.
• Dispatch by element type: native intrinsics are only selected for the types the backend actually supports (Int32/Int64/UInt32/UInt64, Float32/Float64; Int32/UInt32/Float32 on Metal), everything else takes the CAS path instead of a MethodError. Float64 add on CUDA checks the compute capability, as CUDA.jl's own atomic_arrayset does.
• Orderings are still ignored on the intrinsic paths (they have no ordering parameter); this is now stated in each extension instead of a TODO.

Tests

Each backend test file gains test_get_set (including the @atomic A[i] / @atomic :monotonic A[i] = v sugar), test_swap (Atomix.swap! and @atomicswap), test_ordering, and test_float (add, sub, max, min, and * via the CAS loop). The stale "Not implemented" comment blocks are gone.

Run locally on Julia 1.10:

• --CUDA (CUDA v6.3.1, Quadro RTX 4000): 8/8 testsets pass.
• --OpenCL (OpenCL v0.10.11, pocl 7.1): 8/8 testsets pass.
• KernelAbstractions' new atomics testsuite on OpenCLBackend/pocl: 38/38 pass, where 5 failed before.
• oneAPI and Metal are untested locally; relying on Buildkite.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DnYMbgSmZQr7KtRUwei63a

[christiangnrd]

@vchuravy I compiled a list of PRs/issues I think this may have closed:

• Missing @atomicswap implementation for CUDA #49
• Implement atomicswap for CUDA #55
• CUDA: Fallback to UnsafeAtomics #56
• Implement @atomicswap on all extensions #62
• add atomic swap support for OpenCL #66