Preliminary results:
Ryzen 1700X @ stock settings with 8 threads set in the batch file.
Lyra2v2 = 1001 Kh/s on avx2
= 1016 Kh/s on avx
Timetravel = 799 Kh/s on avx2
= 843 Kh/s on avx
X11 = 686 Kh/s on avx2
= 711 Kh/s on avx
X17 = 294 Kh/s on avx2
= 294 Kh/s on avx
X17 was the same for both options, but I was surprised to see avx beating avx2 for the other algorithms tested. My understanding was that it would likely be the other way around....
Disappointing results. AVX2 provides 256 bit vectors vs 128 bit for AVX. Ignoring any memory accesses the AVX2
code uses half the instructions to do the same processing so should be twice as fast.
Most of the AVX2 code is in the x11 chain so it should see a bigger difference than x17.
I had heard that Ryzen had a "faked" AVX2 and your test seems to prove it and that it performs worse than AVX.
I was concerned it wouldn't be any faster than AVX but never considered that it would be slower.
The only reason I can think of to include it is for compatibility with any hard coded AVX2. IMO they should have
left it out.
It appears that AVX2 should continue to be avoided on AMD CPUs.
Thanks for testing.