Wednesday, 2023-04-05

*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc00:15
gnucodesadoon[m]: good morning.  :)01:42
sadoon[m]It's almost dawn here so :)01:47
gnucodesadoon[m]: it's 2 hours from my bedtime.  :)02:05
*** gnucode <gnucode!~gnucode@user/jab> has quit IRC03:00
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC06:26
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc06:26
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC08:18
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc08:19
lkclprogrammerjake, fiiinally - in extremely dark colours, an extremely small near-invisible "tick" icon can be found next to your name on hackaday :)08:44
lkcltook me over 5 minutes to find it!08:45
programmerjakeyup, it worked! thx!08:52
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC09:44
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc09:45
lkcltoshywoshy, about learning SVP64: if you happened to write a tutorial (or improve an existing one) it's easy to justify that as an NLnet "documentation" task10:27
lkcl(i.e. you don't have to be out-of-pocket)10:27
lkcland if you spot anything in the next 2 weeks, that easily justifies putting you on the "reviewed" list of recipients for an RFP as well10:31
lkclmarkos, do you know this guy - - just spotted he's in charge of hyperscan11:33
markosyes, he's one of the original authors, he left Intel and then returned11:34
markoswe did have an email exchange at one point11:34
markoshe said that they were planning to do similar things to hyperscan and port it to other arches and create a microbenchmark to help tuning, etc etc11:35
markosbut there were 'objections' within Intel to supporting other architectures, sigh11:35
markos"I think your overall approach is sound – I would apply the same approach myself in many cases if I had the same constraints or goals. Not having PMOVMSKB is a big headache."11:42
lkclwe've half a dozen ways of achieving that.  grevluti is one - i found it produces 100010001000 and other patterns11:45
markosapparently VSX also has something similar11:45
markosnot quite but makes emulation much faster11:45
lkcl1000s of them, including a5a5a5 aa55aa55 77337733 70000007000000 70007000 70707070 77777777 etc. etc.11:46
lkclyes i remember you mentioned, it was toshywoshy who found it?11:46
markosgather bits11:47
markosvery useful actually11:48
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC12:10
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc12:12
*** octavius <octavius!> has joined #libre-soc12:54
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC13:02
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc13:04
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC14:23
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc14:24
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC14:47
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc14:48
lkcldoes anyone know the twin 64-bit multiply instructions in VSX?19:15
lkcltoshywoshy, markos ?19:15
markos64-bit doubles or ints?19:16
markosscratch that, there is none for 64-bit ints, but for doubles it's  xvmuldp19:18
lkcl2x 64-in producing 2x 64-out (low-half)19:18
lkclyyeah, i found vmuleud VRT,VRA,VRB19:19
lkclreally odd design19:19
markosin the intrinsics reference vec_mul says that 64-bit multiplication is [scalarized]19:20
markosand it's not in 3.0 ISA manual19:21
lkcli know what they're doing - vmuleud is 64-in x 64-in -> 128-out19:21
markosbut it's in 3.1B19:21
lkclwhere vmuloud takes the higher-halves19:21
markosso Power1019:21
*** octavius <octavius!> has quit IRC19:22
markosthat was news to me, I always assumed that just because vec_mul has a prototype for 64-bit ints, that Power9/VSX *actually* supported 64-bit multiplication and I *have* been using it in the past19:26
markosso that's why performance sucked19:26
markosmoral of the story: don't trust the intrinsics, read the assembly!19:26
lkcljust found vmsumudm VRT,VRA,VRB,VRC19:30
lkclp. 35619:30
lkclahh they're for performing 256-bit horizontal multiply-and-accumulate19:32
lkcl(of 64-bit dual-packed-SIMD source operands)19:33
programmerjakewe could probably ask jean-paul to endorse us when we have a paper to submit21:58
*** octavius <octavius!> has joined #libre-soc22:02
programmerjakehmm, maybe not, can't find him on arxiv22:04
toshywoshymeeting now?22:05
toshywoshyare you coming lkcl and djac?22:05
programmerjakeit was canceled, see the email luke sent22:06
markos? I didn't receive anything22:07
programmerjakeLibreSOC / Microwatt / OpenPOWER Sync MkIII22:08
programmerjakeafaict you're cc-ed on the email22:08
markosit was on top of the body22:08
markosbut it seems everyone in the meeting room read the agenda but *not* the cancelation :D22:09
octaviusHeheh, needs to be in **bold** next time22:10
programmerjakelkcl: maybe change the subject to include "Canceled" next time?22:10
programmerjakesince not everyone may read the email before the meeting, assuming someone will just paste the contents into chat like usual22:11
programmerjakeI checked all the people from lip6 in utils.git, afaict none of them can endorse on arxiv22:22
*** octavius <octavius!> has quit IRC22:22
toshywoshyapparently we were so focussed on the agende nobody read the first line22:22

Generated by 2.17.1 by Marius Gedminas - find it at!