@Nick-Nuon I solved the performance problem with ARM/NEON. Lesson learned: do not use Vector128.Shuffle for anything.

@Nick-Nuon Note that our routines are likely now buggy, but I am hoping it is easily fixable. I recommend keeping the benchmarks alive so that we no longer regress. The current speed is good and we should aim to keep it.

@Nick-Nuon My last commit f037b96 simplifies the logic of the Avx2 validator and should fix the tests.

Can you review?

@Nick-Nuon The failures on Arm64 are expected.

I'll work on the Arm64 bit this morning. :-)

@Nick-Nuon My last commit f037b96 simplifies the logic of the Avx2 validator and should fix the tests.

Can you review?

Im on it!

@Nick-Nuon I expect to have the arm component in a few minutes.

@Nick-Nuon I have pushed a commit. Arm should be good now.

@Nick-Nuon Great stuff. I recommend moving iteratively from this point forward. We have a good basis. I trust these tests. We just need to add Sse and AVX-512... but it is somewhat easier than the work done so far. :-)

Merging!

Yeah, I think so too. I really don't regret the time we spent on the the tests. SSE in particular should be straightforward, given how close it is too avx-2.