NEON implementation for TopK/TopKV

[allnes]

Hi,

We rely on it in Arm CPU pipelines. I see TopK only via CPPTopKV (arm_compute/runtime/CPP/functions/CPPTopKV.*, arm_compute/core/CPP/kernels/CPPTopKVKernel.*) and no NEON path in the current main branch (checked 2025-12-08).

Question: is a NEON (or other accelerated) TopK/TopKV already available in a private/dev branch or planned on your roadmap? If not, could you consider adding it?

Use case: detection/NMS post-processing; TopK is a visible CPU bottleneck.

[allnes]

@milpuz01 could you help with that problem?

[alvoron]

This issue has been observed as a result of Phoronix report analisys.
This is the x86 TopK implementation: https://github.com/openvinotoolkit/openvino/blob/3c4e6ae070d626986af5c094d06242b23d675908/src/plugins/intel_cpu/src/nodes/topk.cpp

[morgolock]

Hi @alvoron

It would be helpful if you could share with us a reproducer in C++ with the exact configuration that you experience the performance problems.