PiPNN currently uses f32x16 for both ranking kernels across architectures. Validate whether that width is appropriate on AArch64/NEON before changing the SIMD policy.
- Check TopK, leaf selection, and partition assignment correctness on NEON, including partial vectors and fixed/runtime capacities.
- Compare the current width with suitable NEON widths on representative workloads; record the hardware and results, then retain or adjust the width based on that evidence.
Parent: #878. Follow-up to #1287; this validation is not completed by the numerical-kernel PR.
Review request: #1287 (comment)
PiPNN currently uses f32x16 for both ranking kernels across architectures. Validate whether that width is appropriate on AArch64/NEON before changing the SIMD policy.
Parent: #878. Follow-up to #1287; this validation is not completed by the numerical-kernel PR.
Review request: #1287 (comment)