tevador you might be interested: xmrig/xmrig #1986

sech1: is it basically due to faster mov? dataset init only happens twice per week anyways

It's faster because it initializes 5 items at a time (1 item using integer registers, 4 items using AVX 256-bit registers)

so it also has more time to prefetch data from cache

5 parallel fetches per thread

but it's not faster on all CPUs...

It would've been much much faster if there were 64-bit multiplication instructions for AVX...

Basically main loop is ~4 times slower than the regular loop, but it initializes 5 elements

older CPUs with only 128bit pathways for AVX

I guess I could try to test AVX2 with HashX, it could be a bit faster there since it doesn't use mulh