perf: use uint MAX_INDEX on arm / aarch64 for uint SIMD

[jvdd]

Enhances the performance for arm / aarch64 uint operations by using uint as MAX_INDEX instead of int.
This results in 2x fewer exits from the inner SIMD loop (and thus 2x less - much slower - horizontal SIMD ops) 🎉

Illustration of the performance gain

main branch ⬇️

scalar_u8_argminmax     time:   [274.45 µs 274.47 µs 274.48 µs]
scalar_u8_argmin        time:   [274.52 µs 274.53 µs 274.54 µs]
scalar_u8_argmax        time:   [274.51 µs 274.52 µs 274.54 µs]
neon_u8_argminmax       time:   [58.060 µs 58.072 µs 58.089 µs]
neon_u8_argmin          time:   [38.840 µs 38.843 µs 38.846 µs]
neon_u8_argmax          time:   [38.836 µs 38.837 µs 38.840 µs]
impl_u8_argminmax       time:   [58.059 µs 58.062 µs 58.065 µs]
impl_u8_argmin          time:   [38.869 µs 38.870 µs 38.871 µs]
impl_u8_argmax          time:   [38.873 µs 38.886 µs 38.912 µs]

this PR ⬇️

scalar_u8_argminmax     time:   [274.43 µs 274.44 µs 274.44 µs]
scalar_u8_argmin        time:   [274.53 µs 274.54 µs 274.54 µs]
scalar_u8_argmax        time:   [274.49 µs 274.50 µs 274.52 µs]
neon_u8_argminmax       time:   [39.193 µs 39.194 µs 39.195 µs]
neon_u8_argmin          time:   [29.295 µs 29.295 µs 29.296 µs]
neon_u8_argmax          time:   [29.298 µs 29.301 µs 29.307 µs]
impl_u8_argminmax       time:   [39.194 µs 39.195 µs 39.196 µs]
impl_u8_argmin          time:   [29.299 µs 29.300 µs 29.300 µs]
impl_u8_argmax          time:   [29.299 µs 29.300 µs 29.301 µs]

We see here 30-40% better performance. This improvement becomes smaller when increasing in bit-size of the underlying uint datatype.

[codspeed-hq]

CodSpeed Performance Report

Merging #54 max_index_arm (4a59183) will not alter performances.

Summary

🔥 0 improvements
❌ 0 regressions
✅ 168 untouched benchmarks

🆕 0 new benchmarks
⁉️ 0 dropped benchmarks