-
Aedil
I loaded a LiquidAI 4-bit, a tiny one, into my GPU on Linux. Now I get plausible answers and it works good.
-
jbk
hrm.. did i miss it, or does the mlxcx driver not expose any kstats beyond the standard mac ones?
-
jbk
i wanted to make sure it's not dropping things because of lack of descriptors, dma buffers, etc.
-
jbk
(which don't have standard kstats, but many drivers expose stuff like that as driver specific kstats, but i didnt' see any w/ mlxcx looking briefly)
-
josephholsten
anyone heading to the openzfs summit later this week?
-
danmcd
It's been a bit since I've dug into mlxcx, but nope, no kstats from it per se.
-
danmcd
kebe(io/mlxcx)[0]% git grep kstat
-
danmcd
kebe(io/mlxcx)[1]%
-
» danmcd wonders if its in the wad of not-yet-upstreamed arekinath changes?
-
danmcd
With cxgbe getting a lot of love, fewer immediate reasons to dive into mlxcx. I was lucky to get a spec sheet from a customer who buys enough nvidia gear to get them to cough one up.
-
richlowe
have any of you used SF1_SUNW_ADDR32?
-
jbk
so disappointing -- it appears mellanox used to make it pretty easily available...
-
jbk
thankfully i was scratching my head because it turned out to be a firewall problem...
-
jbk
but it kinda resembled a bit like a ring might have been stuck somehow
-
richlowe
further from asking if any of you have used SF1_SUNW_ADDR32, do any of you have institutional knowledge about what the hell it was for?
-
richlowe
(separate from what it does, which I know, and what it is probably meant to do, which I think I know)
-
richlowe
also if you just want to play along at home, be aware it looks like Rod screwed his commit messages when he did this, so if you read the _other_ commit, you'll just be very puzzled
-
sommerfeld
richlowe: based on a comment in machdep_sparc.h it would appear to be a variant 64-bit ABI which was limited to a 32-bit address space. Big registers, small pointers
-
richlowe
yeah, that's what it just about does. I was wondering if we knew who for
-
richlowe
6755162 is the good one
-
richlowe
where raf implemented the rest of it
-
richlowe
not clear where it just refers to a compiler RFE if they had something in mind, or just if they thought it was a good idea maybe
-
sommerfeld
richlowe: benchmarketing. For what it's worth linux copied this idea with the "x32" ABI; looks good on benchmarks but nobody uses it in practice and they were talking about ripping it out.