volk VS sliceslice-rs

I wonder how much of the performance will improve when compilers get better at RISC-V.

It's been a long time since I could beat the compiler at optimizing assembly on x86, yet in the end merely unrolling a loop and keeping an eye on write-read stalls I managed to get a simple "multiply array by const" about 56% faster:

https://github.com/gnuradio/volk/pull/619

And that's with hardware that doesn't even have vector instructions! I'd understand GCC not supporting that yet.

Some other quickstart docs and hot takes from me on this hardware: https://blog.habets.se/2023/01/VisionFive-2-quickstart.html

Aside from that, their SIMD implementation is better optimized than the one I wrote. Aside from the codegen problem I talked about on that PR, sliceslice does better with its confirmation step by specializing calls to memcmp for all needles up to length 16. This repeats the entire implementation 16 times or so (for each of SSE2 and AVX2, so 32 in total I believe), but lets the memcmp call be a bit better than a generic one. We could do the same in memchr, but I wanted to see how much mileage we could get with fewer copies of the code and a lower latency implementation of memcmp.