I guess it comes down to application. If you don't attempt to find the most general solution, you can dodge those pitfalls. Case in point, abstracting over AVX-512, SVE, and RVV may be tough, but picking one is fine (On nightly only for now), can with the right abstractions can be almost as straightforward as using normal scalar values. I don't have a solution on the CUDA variants either; have been hard-coding that as well... (Cudarc lib with CUDA-version feature gates and GPU-series-specific code). Haven't hit a brick wall yet, but might... or might not.