egg|nomz|egg changed the topic of #kspacademia to: https://gist.github.com/pdn4kd/164b9b85435d87afbec0c3a7e69d3e6d | Dogs are cats. Spiders are cat interferometers. | Космизм сегодня! | Document well, for tomorrow you may get mauled by a ネコバス. | <UmbralRaptor> egg|nomz|egg: generally if your eyes are dewing over, that's not the weather. | <ferram4> I shall beat my problems to death with an engineer.
TonyC1 has joined #kspacademia
TonyC has quit [Ping timeout: 198 seconds]
EntropicRaptor is now known as DeadRaptor
<DeadRaptor>
Currently brewing a pot of coffee at 20:27 because I'm excellent at managing priorities and time! >_>
<DeadRaptor>
(Okay, it's more because of caffeine withdrawl plus planning to be at the observatory)
e_14159- has joined #kspacademia
e_14159 has quit [Ping timeout: 383 seconds]
e_14159- is now known as e_14159
<bofh>
DeadRaptor: I mean I do that often
DeadRaptor is now known as ObservatoryRaptor
<StCipher>
DeadRaptor, for yuri's night?
<StCipher>
i went to earthrise in LA last weekend. was.. kinda rad? for the most part.
<ObservatoryRaptor>
StCipher: not exactly. Going to be helping look for transiting exoasteroid around a white dwarf.
<egg>
bofh: note that the benchmarks are with MSVC and microsoft's libc, and that therefore the performance of things like unionizing double and uint64_t, or frexp/ldexp, etc. in some of the implementations will be affected by that
<egg>
still, it should give a general idea of what the speed is
<kmath>
<acapellascience> "Not retweeting this because my advisor follows me" says one reply. Hemingway hardly penned such a concise tragedy. https://t.co/gquD4KvXVK
<egg>
!wpn UmbralRaptor
* Qboid
gives UmbralRaptor an extragalactic coffee/hydrofluoroolefin hybrid
<whitequark>
that
<whitequark>
doesn't seem like it would work, no
egg|phone|egg has joined #kspacademia
egg|cell|egg has joined #kspacademia
egg|phone|egg has quit [Read error: Connection reset by peer]
egg|mobile|egg has quit [Ping timeout: 198 seconds]
<egg>
whitequark: what, coffee and hydrofluoroolefin?
<kmath>
<modocache> @barrelshifter I refer back to this and https://t.co/yBRklNcZhX a lot. Would you consider doing one for lvalue/rvalue/glvalue/prvalue??
<egg>
!wpn whitequark
* Qboid
gives whitequark a B-stable dynamic 📠
<egg>
!wpn котя
* Qboid
gives котя a neon overflow
<egg>
bofh: interestingly the max error and percentage of incorrect roundings seem to be somewhat independent: sun and householder 10th have the same max error, but very different incorrectness rates
<bofh>
Wait, the *same*? I'm curious if it's on the same imput too.
<egg>
bofh: so like it seems *all* cbrts out there are meh :-p
<bofh>
egg: yeah I'm genuinely surprised
<bofh>
particularly as it doesn't seem too hard to do it well, I mean either high-order Householder or 2x Newton + lower-order Householder seems to do a good job if you do it right.
TonyC2 has quit [Ping timeout: 186 seconds]
<egg>
bofh: or even Halley + 6th Householder, tbh the double-inverse-newton is tricky and doesn't gain much perf (and technically loses a bit of precision)
<awang_>
egg: What's different about your implementation that makes it so good?
<egg>
bofh: 49.6 misroundings, doesn't quite make it to "more often incorrect than not", but close :-p
<egg>
awang_: it's based on Kahan and takes advantage of reciprocal throughput >> latency
<egg>
um. reciprocal throughput << latency rather
<egg>
or throughput >> reciprocal latency, whatever :-p
<awang_>
It looks all the same to me :P
<egg|cell|egg>
!Wpn whitequark
* Qboid
gives whitequark a sufficiently smart Zener orb with a bimap attachment
<egg|cell|egg>
Bofh: any other cbrt implementations that you can think of?
<bofh>
nope, I think atlas, sun, msft, glibc & cephes covers all the libm impls I can think of.
<whitequark>
!wpn egg
* Qboid
gives egg a private bit-addressable paraboloid
<egg|cell|egg>
Bofh: what is cephes though, I had never heard of it
<bofh>
a special function collection sonewhat common in scientific computing. I've stolen part of their J_{nu}(x) code in the past.
<bofh>
J_ν(x)*
TonyC has quit [Ping timeout: 186 seconds]
<egg|cell|egg>
Bofh: does fnlib count as a libm
<egg|cell|egg>
Should I add it
<bofh>
meh? like impl doesn't seem too special or good but might as well.
<bofh>
the impl*
<egg>
bofh: yeah the implementation seems boring, but if I'm going to test ALL the cbrts,
<egg>
bofh: otoh not sure how to make MSVC compile FORTRAN 77 :-p
<egg>
(no i am not writing a FORTRAN compiler for principia)
<bofh>
grab ifort or use f2c or I can do the latter myself in a half-hour since it's easy
* egg
pets whitequark
<egg>
bofh: linking random things is probably going to get messy, unless you can compile it to masm with ifort f2c is probably the way out
<bofh>
like just let me get back to my desk, sec.
<UmbralRaptor>
Linking random things is for silliness in IRC channels, not coding.
<UmbralRaptor>
Reason computers are bad number 768: the idea of truly generic fasteners that are used for *all* hardware in a mass/stress/material range is incomprehensible to a skilled GPU driver dev.
<egg>
!wpn bofh
* Qboid
gives bofh a transuranic semigroup which vaguely resembles a pidgin
<rqou>
somehow the "scientific" world seems to _suck_ at "basic IT issues"
<UmbralRaptor>
o_O
<APlayer>
Uh, is an excess of data really a problem?
<bofh>
rqou: so like by photographs they mean, like, Kodachrome K-14
<UmbralRaptor>
rqou: That the macbook air I was issued by my advisor cost ~1/9 my stipend at Missouri State University may have something to do with that.
<bofh>
APlayer: so like NIMBUS is like 1964 - ~1978
<APlayer>
I may or may not have lost a 1 kilogram mass from our school's physics materials.
<bofh>
the Vidicon and IR radiometer, if downlinked as raw analog data, could very easily overwhelm disk capacity even after lossy digitization
<APlayer>
Looks like I'll be killed by mad scientists.
<APlayer>
bofh: "Too much data" can be easily turned into "just enough data", can't it?
<UmbralRaptor>
Quick, check sat data for 21 megaton explosions in 🇩🇪
<rqou>
btw random offtopic questionrant: why does it seem like there's some section of the "scientific python ecosystem" that's "infected" by conda?
<rqou>
e.g. i wanted to install cvxpy the other day
<APlayer>
There was a 21 MT explosion here and I don't know?
<rqou>
and the instructions say "use conda"
<rqou>
but i just said "f*ck that, i'll use pip" and it works fine?
<bofh>
rqou: b/c some people in the academic world insist upon claiming conda is a net benefit somehow.
<bofh>
I do not understand how one can be that insane or that far into Stockholm syndrome to believe that, my only guess is it's an attempt by them to slow down the pace of research so they can publish their results faster.
<UmbralRaptor>
!wa 1 kg*c^2 in megatons of tnt
<Qboid>
UmbralRaptor: convert 1 kg c^2 (kilogram speed of light squared) to megatons of TNT: 21.48 megatons of TNT
<rqou>
also why do i get the impression that there's this whole "scientific python ecosystem" that i don't really know about and doesn't really seem connected to "mainstream" software engineering
<APlayer>
UmbralRaptor: Antimatter?
<UmbralRaptor>
bofh: I blame it on how we're all using macs instead of something with package management built in.
<UmbralRaptor>
APlayer: direct matter to energy conversion. You said you lost a kilogram…
<APlayer>
Woops
<rqou>
btw anybody here a huge masochist and want to debug why some matlab code doesn't work in octave?
* APlayer
is sorry for not getting that
<rqou>
(yes this is the thing i was complaining about a few days ago)
<UmbralRaptor>
rqou: a SF software engineer would cost the same as an entire department. ¯\_(ツ)_/¯
<APlayer>
So much for "Death by mad scientist assault"
<UmbralRaptor>
Astropy exists in part because a bunch of grad students are fed up with IDL.
<bofh>
UmbralRaptor: so like, literally exchanging tarballs over 3.5" floppy is a better package management system than conda, to paraphrase whitequark
<bofh>
I'm not kidding when I hear someone complaining about it breaking within earshot IRL at least once or twice a week (irritatingly, I often wind up getting asked to help).
<UmbralRaptor>
bofh: modern macs don't support floppies. Or CDs. =|
<rqou>
although to be fair "mainstream" package management isn't great either
<bofh>
UmbralRaptor: USB floppy drives.
<rqou>
"it works on my machine, so let's use containers to give you my machine" :P
* UmbralRaptor
🔪 homebrew.
<rqou>
bofh: won't always work if it's a 720k floppy :P
<UmbralRaptor>
Yes, you're very clever. Now how about terminology that doesn't assume I'm big on craft beers?
<rqou>
oh yeah, f*cking homebrew
<rqou>
presumably some sf software engineers smoked too many joints before coding it
<bofh>
UmbralRaptor: yeah that pisses me off to no end too
<rqou>
although the python ecosystem _also_ has stupid special words
<rqou>
like "wheel" and "egg"
<egg>
meow
<UmbralRaptor>
egg.py vs egg.egg
<rqou>
vs <egg|egg> :P
<UmbralRaptor>
≡ 1 (I hope)
<egg>
bofh: hmmm I had run my analysis with G = 0x553EF0FF289DD796 (optimizing for the error after the integer arithmetic approximation) rather than G = 0x553EEE713FD44BE6 (optimizing the error after two rounds of Newton and inversion)
APlayer has quit [Read error: Connection reset by peer]
BPlayer has joined #kspacademia
<bofh>
egg: how much of a difference does it make?
<egg>
bofh: currently rerunning the tests, probably not much
<egg>
(also did you know computing a billion cube roots with my really slow correctly-rounded method is slow)
<SilverFox>
nice
<SilverFox>
what does this difference mean?
<egg>
bofh: 854 misroundings instead of 873 at 100 million test values
<bofh>
egg: so effectively a negligible difference in the end
<egg>
bofh: well, it's an improvement and it's *really* free, so
<egg>
(replace *magic number* by *other magic number* :-p)
<egg>
bofh: Halley instead of double-Newton should be more significant, I'll try that
<egg>
(but that's slightly slower)
<egg>
not much tho
<egg>
bofh: I'm fascinated at how bad all those cbrts are, since aside from atlas's they're worse than Kahan's on both precision and performance
<egg>
and that qbrt paper is not new
<bofh>
egg: yeah, I'm confused especially at the one from Sun which uses basically exactly the algo in the qbrt paper.
<egg>
bofh: it hasn't picked up on the "last iterate must be high order" bit
<bofh>
ah.
<bofh>
that would make sense then.
<egg>
Kahan doesn't really go all in on that one either (only 4th order) but that's because he's in Hornerland where things get slow quickly
<egg>
(and accumulate rounding error)
<egg>
bofh: with positive terms the rounding error bounds on estrin are better than those on horner in pretty much the same way as it's faster afaict?
<egg>
or maybe I'm confused
BPlayer has quit [Ping timeout: 182 seconds]
awang_ has quit [Ping timeout: 383 seconds]
UmbralRaptor has quit [Ping timeout: 186 seconds]
UmbralRaptop has joined #kspacademia
<egg>
bofh: yeah I think that's right (even without FMAs), I'm getting (1+u)**5 for Estrin vs (1+u)**6 for Horner on a 3rd degree polynomial
<egg>
bofh: ... nevermind, Halley is faster than two newtons on the inverse :D
<egg>
and it rounds noticeably better, so
<egg>
bofh: also I should zzz
* egg
flying to LHR tomorrow
<egg>
!wpn bofh
* Qboid
gives bofh an intensional dust
<bofh>
why to LHR / my condolences?
<egg>
bofh: vacation
<bofh>
And yeah, Estrin *should* give better error bounds in most (every?) case (at least with +ve coeffs, I think every).
<egg>
bofh: going to spend the next week eating in fancy restaurants and spending absurd amounts of time staring at things in the british museum and the national gallery :-p
<egg>
bofh: and afaict the other complaint ones aren't actually a thing?
<egg>
(aside from the fanfics)
<bofh>
so that's the thing. there *ARE* many other cuneiform tablets found in that same excavation location, just most don't seem to be transcribed or even photographed.
<bofh>
that reminds me, I've been asked to look into how2get Google's IME system into implementing an Akkadian IME (and ofc have not had time to, but a lot of that is unfamiliarity with the system).
<egg>
bofh: I mean if I don't even know whether there are other tablets it's going to be tricky to find them (and they might not be on display, so short of hunting down a curator...)
<egg>
ah the complaint itself is not on display :-/
<bofh>
this reminds me of the time I wandered around randomly in the NYPL into the archival processing room and wound up reading unprocessed microfiche for an hour before getting shoed out
<egg>
:D
<bofh>
the librarian that kicked me out complimented me on how careful I was and told me who to email if I wanted access formally, tho admitted they usually don't grant it for reasons of "I wanted to read some old Russian spy documents from the 1880s out of boredom"
<UmbralRaptop>
Hah!
<UmbralRaptop>
Also, microfische without a reader?
<UmbralRaptop>
(I think that makes you a witch?)
<bofh>
UmbralRaptop: no, as in I was unpacking it and running it thru the reader for the first time, then carefully refolding it.
<bofh>
I *can* read microfische with just a magnifying glass tho.
<egg>
microfische?
<egg>
when you have several smol german fish
<bofh>
ROFL
<UmbralRaptop>
Yes. Eggsactly that.
<Iskierka>
µ𓆟
<egg>
!u 𓆟
<Qboid>
U+1319F EGYPTIAN HIEROGLYPH K005 (𓆟)
<egg>
!u µ
<Qboid>
U+00B5 MICRO SIGN (µ)
<egg>
!wpn Iskierka
* Qboid
gives Iskierka a nonpolar FITS knife
<egg>
bofh: anyway I should zzz
<egg>
bofh: ping me if you manage to translate the FNLIB CBRT to either masm or C++
<bofh>
oh right I forgot to do that. sec
egg|cell|egg has quit [Read error: -0x1: UNKNOWN ERROR CODE (0001)]