Target support is good, though AFAIK non-constexpr-size vectors (SVE/RVV) are not supported.
The biggest concern I'd have is predictability. A thin wrapper over intrinsics makes it easy to experiment with codegen and see what's going on. We've seen some compiler bugs but usually narrow scope and relatively easy to reproduce. By contrast, Halide involves extensive compiler transforms. When that breaks, it's much harder to move forward. There is also a risk of performance cliffs because the language is very flexible and AFAIK lacks 'this is going to be slow' guardrails.
It's great for prototyping and image processing and this is probably what you're using it for? But I also would not expect peak performance, at one point I reimplemented parts of RAISR with explicit SIMD and saw a surprisingly high speedup.