[Libre-soc-dev] harmful SIMD again

[Luke Kenneth Casson Leighton]

https://www.reddit.com/r/programming/comments/p0yn45/three_fundamental_flaws_of_simd/h9ncoft/?utm_source=reddit&utm_medium=web2x&context=3

FUZxxl is advocating the case for SIMD by implementing this algorithm
in AVX512, SSE and NEON:
https://github.com/clausecker/pospop/blob/master/safe.go

i initially made the mistake of thinking it was a straight vectorised popcount,
it's not.  for count8safe there are 8 accumulators:
* accumulator 0 receives the count of the number of bit zeros of ALL
  the input vector.
* accumulator 1 receives the count of the number of bit 1s of the entire
  input vector

    for i := range buf {
        for j := 0; j < 8; j++ {
            counts[j] += int(buf[i] >> j & 1)
        }
    }

NOT

    for i := range buf {
        for j := 0; j < 8; j++ {
            # counts is NOT the same length as buf.
            counts[>>>>I<<<<] += int(buf[i] >> j & 1)
        }
    }

turns out that SVP64 can do this algorithm in about 12 or so instructions,
which is 35x less than the instruction count needed for AVX512 or NEON.

l.