raid/riscv64: Optimize xor_gen_rvv

[leiwen2025]

sg2044:
      new: xor_gen_warm: runtime = 3062177 usecs, bandwidth 50580 MB in 3.0622 sec = 17579.80 MB/s
      old: xor_gen_warm: runtime = 3062535 usecs, bandwidth 31581 MB in 3.0625 sec = 10564.11 MB/s

[sunyuechi]

This adds manual unrolling on top of the standard RVV path, but the extra
gain isn't consistent across machines (likely because sg2044's vector unit
is out-of-order while k3's is in-order):

sg2044  #412 = 16517 MB/s   #418 = 17579 MB/s   (418 faster)
k3      #412 = 16295 MB/s   #418 = 15406 MB/s   (418 slower)

For a portable implementation I'd lean toward the plain standard-RVV style
(#412) unless there's a stronger case across more machines that the manual
unrolling is a net win.

[leiwen2025]

This adds manual unrolling on top of the standard RVV path, but the extra gain isn't consistent across machines (likely because sg2044's vector unit is out-of-order while k3's is in-order):

sg2044  #412 = 16517 MB/s   #418 = 17579 MB/s   (418 faster)
k3      #412 = 16295 MB/s   #418 = 15406 MB/s   (418 slower)

For a portable implementation I'd lean toward the plain standard-RVV style (#412) unless there's a stronger case across more machines that the manual unrolling is a net win.

Thanks for the review and for sharing the benchmark numbers.
How about creating a VLEN=128 branch for the unrolled loop? The unrolled version is modeled after ARM NEON 128-bit implementation, so the gains maybe are expected primarily on VLEN=128 machines like sg2044. Platforms with wider vectors (like k3) would continue using the plain standard-RVV style(#412).

[sunyuechi]

This might be more of a microarchitecture / memory-throughput / OoO-vs-in-order
thing rather than VLEN, maybe? As for the NEON style.. I'm just guessing, but
maybe it's because NEON is fixed-length and can't be written the RVV way.

One example is k230 (VLEN=128, in-order, RVV 1.0):

pr 412
  xor_gen_warm: runtime =    3062030 usecs, bandwidth 2280 MB in 3.0620 sec = 744.90 MB/s
pr 418
  xor_gen_warm: runtime =    3061614 usecs, bandwidth 1936 MB in 3.0616 sec = 632.42 MB/s