< bitcoin-git>
[bitcoin] lucash-dev opened pull request #13404: [tests] speed up of tx_validationcache_tests by reusing of CTransaction. (master...speedup-tx_validationcache_tests) https://github.com/bitcoin/bitcoin/pull/13404
< wumpus>
MarcoFalke: I hope that happens automatically? otherwise, it sounds like a nightmare to keep it up to date
< wumpus>
MarcoFalke: also, it's already possible to use bitcoinacks.com to keep track of that
< wumpus>
how are things with 0.16.0rc1? do we have anything that needs to be backported for rc2? I haven't heard any reports of bugs at least.
< wumpus>
if not, we should do a very fast rc2 for the translations issue and then tag final
< wumpus>
eh, 0.16.1 obvs
< bitcoin-git>
[bitcoin] ken2812221 opened pull request #13406: travis: Add make step so that travis can build all executables for Mac. (master...travis_make_mac) https://github.com/bitcoin/bitcoin/pull/13406
< MarcoFalke>
wumpus: I know it is on bitcoinacks.com, but I don't like the idea to switch back and forth between websites when you could have it all in one place
< MarcoFalke>
And yes, will be automated in the future
< sipa>
(Ryzen actually only has 4-way parallel arithmetic, so avx2 doesn't have that much of a gain)
< cfields>
interesting that it's almost the same as avx2
< cfields>
ah
< jarthur>
Yep, it's practically just "API compatible" with AVX2.
< sipa>
well it also gives you 256-bit registers
< sipa>
but i guess those exist at AVX already
< jarthur>
sipa: gmaxwell mentioned that zen might be able to do parallel sha-ni runs if each step is loaded up side by side. Do you know if anyone has played with that yet? I volunteered at some point but didn't get around to it.
< sipa>
jarthur: yup, 2-way SHA-NI is faster than 1-way on my system
< jarthur>
nice! Is that code in your branch atm?
< sipa>
yup
< jarthur>
rockin
< sipa>
(not quite 2x - the implementation needs 10 registers-ish, so 2-way needs 20, while there are only 16 addressable ones, resulting in spills)
< sipa>
IIRC it took a benchmark from 0.83ms to 0.61ms by doing the 2-way
< sipa>
it just calls every round function twice in a row
< jarthur>
Thanks. Looks nice and clean with those inlines.
< sipa>
who is DrahtBot?
< wumpus>
sipa: MarcoFalke's bot
< sipa>
ah, nice
< MarcoFalke>
[ ] I'm not a robot
< sipa>
oh, i meant "who is running Drahtbot"
< sipa>
i did realize it was a bot :)
< sipa>
how does it figure out conflicts?
< sipa>
does it try every combination of 2 PRs?
< MarcoFalke>
sipa: Yes, rn. I might implement a smart solution when I have time. Though, the compute overhead is trivial compared to the latency by the github api for now...
< cfields>
that has the libs split out, rebuilt for each isn set, and adds avx
< cfields>
and yea, i see your point now, we need to pass in a flag for the namespace
< jarthur>
sipa cfields, are we going down the direction that optimal instruction set would be picked at runtime, and default b86_64 build would have the lot of them compiled?
< cfields>
jarthur: yes. that's already the case, we're just diving deeper.