Questions about topk_per_row_kernel in topk_plain_kernels.cu

[zjin-lcf]

In the source file https://github.com/ROCm/aiter/blob/main/csrc/kernels/topk_plain_kernels.cu, the kernel is launched when

       // Use topk_per_row kernel when:
        // n + K log²K ≥ 3 × Factor(n) × n
        // where Factor(n) = 1/3 + 1.6/(log₂(n) - 9.5)

Can you please explain these magic numbers ?

When this HIP kernel is converted to CUDA kernel, can you suggest where users need to do performance tuning targeting an NVIDIA GPU ?

Thanks