programming

[acoustic mango]

At least in New Zealand

[prime ridge]

nevertheless, pretty kool. Good luck!

also holy so many lurkers in chat rn

just... watching

[acoustic mango]

Thank you!

[wild wadi]

[amber fractal]

[knotty current]

[prime ridge]

非常based bio btw 😎

[knotty current]

that might change soon in the future

i have fallen down the ARG rabbit hole

[olive sable]

this code is now as fast as i can get it

def getLightSpaceMatrix(self, view):

  position = self.player.camera.position + self.player.camera.forwards * self.frustumdepth
  self.lightview[:, 3, 0:3] = np.stack([-position @ self.lightright, -position @ self.lightup, position @ self.lightforwards], axis=1)
  
  transView = view[:3,:3].T
  self.invView[:3,:3] = transView
  self.invView[3,:3] = -view[3,:3] @ transView
  
  viewcorners = self.preFrustum @ self.invView @ self.lightview
  
  mins = viewcorners.min(axis=1)[:, :3]
  maxs = viewcorners.max(axis=1)[:, :3]
  
  minZ = -10 * mins[:, 2]
  maxZ = -10 * maxs[:, 2]
  
  return self.lightview @ np.array([create_orthogonal_projection(mins[i, 0], maxs[i, 0], mins[i, 1], maxs[i, 1], maxZ[i], minZ[i]) for i in range(4)])

fuck

[amber fractal]

[olive sable]

pressed enter too fast

[prime ridge]

noooo

[olive sable]

the ctrl+v didnt work before and now i look like a fool

[knotty current]

how did u know i understand mandarin?

[prime ridge]

I can sense a mandarin brother anywhere

[knotty current]

im chinese

[olive sable]

hi chinese, im sam

[hoary lion]

drop the paper 🔥 🔥

[hoary lion]

not this again

[prime ridge]

Can I try to guess city?

[knotty current]

sure

[olive sable]

its like 6am so i should be getting to bed sometime soon

[hoary lion]

im sleepy -> hi sleepy, im sam
im happy -> hi happy, im sam
im [ethinicity] -> hi [ethinicity], im sam

[olive sable]

i am sam tho

never lied

[amber fractal]

Hi hyper

[prime ridge]

Probably not north. I can't see Fuzhou. Probably either Chongqing or Guangzhou. Surely not Shanghai or anything . Final answer: Guangzhou

[knotty current]

wrong

[prime ridge]

it's hard over text

[knotty current]

it is

[prime ridge]

also I just learned Mandarin cuz I was bored ☠️

im not ethnically Chinese at all

[knotty current]

want me to reveal where im originally from?

[prime ridge]

If u want. Probably southern

[knotty current]

actually i have been out of china for 5 years now so i somehow forgor

[knotty current]

mid-eastern i think

[prime ridge]

I might know

Def Henan province

[amber fractal]

Waiting for create_orthogonal_projection to not need to be looped over

[prime ridge]

so maybe like Hebei?

CHina ia massive

[knotty current]

no

[prime ridge]

damn

[knotty current]

that is where my grandparents are

[prime ridge]

aw cmon bruh

[hoary lion]

welcome to doxxing stage of #programming

[prime ridge]

nah it's calmmmm

[knotty current]

i dont actually live in china now

[amber fractal]

[olive sable]

ye idk about that one.
that function is just this

def create_orthogonal_projection(left, right, bottom, top, near, far, dtype= np.float32):

    rml = right - left
    tmb = top - bottom
    fmn = far - near

    A = 2. / rml
    B = 2. / tmb
    C = -2. / fmn
    Tx = -(right + left) / rml
    Ty = -(top + bottom) / tmb
    Tz = -(far + near) / fmn

    return np.array((( A, 0., 0., 0.),
                     (0.,  B, 0., 0.),
                     (0., 0.,  C, 0.),
                     (Tx, Ty, Tz, 1.),),
                     dtype=dtype)

so the only way to not loop is to hardcode the 4

[maiden geyser]

the acceleration is MASSIVE

[hoary lion]

yk what else is massive?

[maiden geyser]

low taper fade

[olive sable]

what?

[maiden geyser]

mini soda

[hoary lion]

mods twist these fake neuro fans' balls

[olive sable]

was that modded or deleted by apollo?

[hoary lion]

self-destruction

[prime ridge]

deleted by me lol

mods only lurk

[hoary lion]

https://tenor.com/view/mods-testicular-torsion-lowtiergod-fool-gif-17829099525522895736

[olive sable]

why react with neuro british tho?
do americans not say "what?"

[hoary lion]

yes

[prime ridge]

seems like something a non-american would say

[hoary lion]

im in canada but like yk

[olive sable]

i am indeed not american

[hoary lion]

same continent

[amber fractal]

yes (I'm spreading misinformation)

[olive sable]

but also not british so idk

[prime ridge]

Any EDT time zone fellows?

[olive sable]

european daylightsavings time

[prime ridge]

Eastern US

[olive sable]

shhhhhh

[prime ridge]

or UTC -5 ig I should say

[amber fractal]

Nope, in PDT

[prime ridge]

Cali moment

[olive sable]

post-daylightsavings time

[hoary lion]

friend found

[prime ridge]

Please god don't tell me you live in Vancouver 😭

[hoary lion]

NAH

[prime ridge]

Prob worst city

[hoary lion]

he doxxing me

[prime ridge]

literally only city in PDT

[hoary lion]

chat do anything 😭

[olive sable]

whats wrong about vancouver?

[prime ridge]

everything

shit, i'd rather live in the bronx

[hoary lion]

too big ong

[prime ridge]

a lotta Chinese people live in Vancouver tho. So pretty nice

[olive sable]

everything but brussels

[amber fractal]

[prime ridge]

brussle sprouts

[olive sable]

nah, scratch that

[hoary lion]

bushels?

[olive sable]

everything but charleroi

all my homies hate charleroi

[prime ridge]

European city "crime": 😊
American city crime:

[amber fractal]

[prime ridge]

at least it aint vancouver tho

[olive sable]

idk bro, charleroi is pretty crazy

[hoary lion]

except london 🔪

[amber fractal]

Not me visiting LA and taking all measures

[prime ridge]

ngl I live nearish an area with a crime rating of 99/100 on the FBI list

literally one of the highest in the country

for context the bronx is 40

LA is like 50 I think

[amber fractal]

A good stabbing time

[maiden geyser]

verbrechen, Deutschland: 🤮💩🚨⚠️
kuraimu, 日本: 🌸🌸♥️😍🥰✨✨🌈

o algo

[prime ridge]

Most dangerous Belgian City

[olive sable]

ye thats what they want you to think, but noticve how the picture is empty

its fake

[hoary lion]

ai generated images are going too far

😥

[prime ridge]

that was off wikipedia

[olive sable]

irl charleroi has at least 20 homeless people doing hard drugs

[prime ridge]

sign me up

only 20

[olive sable]

i meant the location in the picture

[amber fractal]

(Can we not get modchecked please)

[olive sable]

ive been there

[olive sable]

oh ye

hmmm

[prime ridge]

mods never go in #programming

[olive sable]

bottom line is, the town center looks okay, but thew moment you go off the main roads its a shithole

[amber fractal]

[olive sable]

just like france

[prime ridge]

At least from when I was in Europe, people are a lot nicer

[olive sable]

no, we just dont say everything

in reality we're just as mean

[prime ridge]

The dutch were the most frank people I met

[amber fractal]

-# why is modcheck not on this version of bingo?

[prime ridge]

mods prolly just enjoying the convo prolly

[olive sable]

we just dont "rotate around the jar"

[prime ridge]

what a phrase

[amber fractal]

It is dead hours for modding rn

[olive sable]

we dont like spinning around the pot

[prime ridge]

probably morning in the UK

[olive sable]

beat around the bush in english i think?

[prime ridge]

yep

[amber fractal]

yep

[olive sable]

yep, 6am there

here its 7am

[prime ridge]

1 am

[olive sable]

10% of exe time

pretty decent

[prime ridge]

Writing performance code in python

[wild wadi]

what do you use for profiling?

[olive sable]

it is actually good in terms of performance

[prime ridge]

You can probably do some caching

[olive sable]

py-spy

py-spy record -o profile.svg -- python main.py

[wild wadi]

thanks

[olive sable]

i think almost everything that can be cached is cached rn, in terms of CSM code

[prime ridge]

Anybody can cache but seldom are these caches efficient

or "optimized"

[olive sable]

main improvement now is batching the drawcalls

[unkempt citrus]

https://fixupx.com/tirayed/status/1927861621755420877

Tea 😼🎗🍉 (@Tirayed)

LMFAOOOO

**💬 601 🔁 6.6K ❤️ 148.0K 👁️ 1.75M **

FixupX · May 28, 2025, 10:56 PM

[prime ridge]

are u writing to a buffer rn

[amber fractal]

@olive sable Could we get a visual of the code
https://nodezator.com/

[olive sable]

what

sure?

how?

[amber fractal]

Visual python is crazy

I heard you can import code, but not sure if you just need to include the path or what.

[olive sable]

its loading

i think

its doing something at least

[prime ridge]

visual python is better for complex algos. might be weird for rendering code

either really good or really baf

never used it myself

[amber fractal]

I do want to see how stupid it looks

It is probably dying due to the hot reloading if I had to guess

[olive sable]

i cant figure out how to import an existing file into it

[amber fractal]

Yeah it doesn't seem easy to do

looking through docs I still got no clue

[olive sable]

cant find it

anyways, the code is all on github so https://github.com/samvanmaele/game-engine

imma sleep

the demo on github is not up-to-date btw

goodnight

[prime ridge]

[fast pagoda]

"i know y'all have cheated before"
classic "well i do it so clearly everyone does" moment

[amber fractal]

Classic indeed

[tender river]

@grok is this true

exponentials are like the only part of linear logic i struggle to understand atm

[hoary lion]

wtf is that symbol

[dry charm]

null coalesce is a upside down and b popsicle to nullcoalesced A upsidedown and nullcoalesced B

Popsicle is my favourite math operation

[maiden geyser]

evil &

[tender river]

i understood it

i just had to correct my understanding of ?

using unary ~ for negation and -> for lollipop and par for ⅋

?(A par B) -> ?A par ?B
~(!(~(A par B))) -> ~(~(?A par ?B))
~(?A par ?B) -> !(~(A par B))
~(?~A par ?~B) -> !(~(~A par ~B))
~(?~A par ?~B) -> !(A * B)
~(?~A) * ~(?~B) -> !(A * B)
!A * !B -> !(A * B)

you could definitely make a shorter proof but i'm just glad i figured it out

[stark needle]

Morning

[maiden geyser]

morbing

funny how 3/5 morbings in this chat are mine

i hold a morb monopoly

[opaque sigil]

https://github.com/HigherOrderCO/Bend

[autumn ore]

yo programmers wanna help me test smth

[knotty current]

what thing?

[autumn ore]

a little server i put on my raspberry piu

i dont know if i did the fowarding right

so i need someone not on my wifi to help check

[knotty current]

ok

[autumn ore]

ok

first

go to https://ytsodacan.github.io/120.html and wait for it to load

shut up desscription

then make your profile and go to multiplayer

go network settings in the top left

click add relay

[knotty current]

ok

[autumn ore]

then put in this link existing-kentucky.gl.at.ply.gg:35863

it should say Soda's Relay

go back to multiplayer

and it should say YtSodaCans world

join iy

[knotty current]

wait

[autumn ore]

i am at -31 63 213

[knotty current]

it didnt

[autumn ore]

wdym

it didnt do what

[knotty current]

the relay is not online

[autumn ore]

did it say no connection

[knotty current]

i dont think i can help any further because im not an expert in networking

sorry

[autumn ore]

my fault use this link ws://existing-kentucky.gl.at.ply.gg:35863/

@knotty current

[knotty current]

ok

[autumn ore]

tell me if it works

@knotty current ?

[knotty current]

still trying

[autumn ore]

ok

dosent work, try this

ws://existing-kentucky.gl.at.ply.gg:35863/

[knotty current]

that is the same timg

[autumn ore]

dang

[knotty current]

check ur port fowarding and other stuff

also

[autumn ore]

ok

[knotty current]

i kinda have other stuff

sorry

[autumn ore]

ok

[dense cosmos]

i <3 linux

[knotty current]

nice rice

[dense cosmos]

thank you, though I didn't really do much beyond installing stuff

[knotty current]

that is already an acheivement, certain people fail to install stuff

[maiden geyser]

the most atrocious way to display cpu frequency

[dry charm]

it works ¯_(ツ)_/¯

I assume you can put processes under than

[dry charm]

Welcome to the other side, we hope you enjoy your time and don't forget to not break your install

[rigid snow]

fuckin apple watch rings

[dry charm]

and even if you break your system, don't panic, just google and fix it

it is 98% of the time fixable

the other 2% you just deleted your whole root partition

[knotty current]

at least u dont have a half broken hyprland install that is still using KDE GTK and Qt themes like me

[dry charm]

wait you are supposed to fix that?

[knotty current]

i cant

and i dont have time

i would do a full reinstall at a later date though

[maiden geyser]

i think i burnt my 67w phone charger's fast charging circuitry

xd

[fast pagoda]

✍️ 🔥

[trim valve]

neat

[nocturne olive]

Now you have blazing fast charging

[dense cosmos]

    # If variable is `ON`, `YES`, `TRUE`, `Y`, or non zero number
else()
    # If variable is `0`, `OFF`, `NO`, `FALSE`, `N`, `IGNORE`, `NOTFOUND`, `""`, or ends in `-NOTFOUND`
endif()
# If variable does not expand to one of the above, CMake will expand it then try again``` bruh, what is this

[opaque sigil]

cmake :)

[desert wave]

I made my charger rusty and now it doesn't charge at all

[dry charm]

I've seen this before

[midnight sigil]

https://www.youtube.com/watch?v=SyjzArtKqEw

[maiden geyser]

14 competing standards and their consequences

[rigid snow]

psa: having rust on your phone charger will NOT make it faster

[quaint iron]

Really, it's not even recommended to use ferrous contacts on your phone charger to begin with.

[dense cosmos]

the hypocrisy

also, was curious how long the <random.h> header file was

beeg

[ruby timber]

Random number generation is crazy, turns out

[dense cosmos]

this link doesn't work kek

[stark needle]

https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-ceo-says-ai-could-cause-up-to-20-percent-unemployment-within-five-years-wipe-out-half-of-all-entry-level-white-collar-jobs

✅

[ruby timber]

Yay

This is a w, because it means we won't have to work anymore!

[maiden geyser]

i sneezed at my phone with a mouth full of water, soon after i cleaned it i connected it to the charger
not enough time for rust to commence

it looks kind of good

i gave it an example task from an exam i will have in a month and it kind of solved it

10t/s is very slow though

[stark needle]

only 10t/s??

doesnt deepseek have mtp and is a super sparse model

[maiden geyser]

idk i'm using 8b q8 on a 3080

[stark needle]

wait the full deepseek?

as in the 671b?

ah

i see

the distilled version

[olive sable]

Goodmorning

Im doing tv networking

Fuck PLC

and fuck WPS

I got woken up for this shit

[pastel nymph]

Hey programmers, I'd like some advice.
I initially started my journey in programming trying to get into backend, so I know python and Go, and probably most of the concepts. Recently I realised that I'm not actually that interested in what backend does, but I guess learning it was still useful.
So I want to try going into gamedev, and my question is, what language would you recommend?

[opaque sigil]

Yes

[olive sable]

Godot is nice, unity is the standart, c++ is good for implementing game engine stuff yourself

UE5 is hard to get right so idk about that one

Lots of people leave stuff on the default setting cuz they dont know the engine enough, making their game have bad performance

[stark needle]

i have the urge

of making an adversarial AI agent that improves CVs for a specific job description

[olive sable]

Ai generated CV bout to go crazy

[ruby timber]

I recommend Godot if you want to try a game engine, it's a great way to learn the ropes

[rare bramble]

funny because I completely stopped using AI to improve my cv, aside from maybe proof reading. I think that the people who hire see enough of copy paste GPT cvs that they learn to disregard them

[maiden geyser]

it's not like you have a choice
it's either C# or C++
java can be a variant

[pastel nymph]

Godot has it's own language, right?

[ruby timber]

It does, but you can plug in C# as well iirc

[maiden geyser]

gdscript, aka weird python

[ruby timber]

And since you're familiar with python, GDscript should be pretty nice to use for you

[pastel nymph]

Yea, but which one of them, if you have to start with something? I don't really know the key differences

[opaque sigil]

C# realistically

C++ is mostly for unreal engine and/or engine development

[pastel nymph]

Not a big fun of python tbh. I love coding in Go, but it's application area just isn't that interesting to me

[olive sable]

in theory you can do gamedev in any language you want, as long as you dont mind having to do some stuff yourself if the language doesnt have a game-engine with those features

[dry charm]

C# is the most hands down gameengine-esque language, due to large support.
You can go Kotlin if you want to use some JVM related language (KorGE, Godot (wtf this is a thing))

[olive sable]

yep

[dry charm]

You can even do it in JavaScript, there are a few game engines in that as well

[opaque sigil]

You can write scripts for Godot using kotlin?

[dry charm]

https://godot-kotl.in/en/stable/

I was suprised as well

Though I love this unsupported line Each registered constructor must have a unique number of arguments, constructor overloading is not yet supported. :om:

[pastel nymph]

Alright, I'll start learning C# then. Thanks chat

[dry charm]

The most important thing is have fun!

[opaque sigil]

[stark needle]

holy

i found it

https://huggingface.co/datasets/lang-uk/recruitment-dataset-job-descriptions-english

job descriptions dataset

i CAN train an ai that writes good CV

[knotty current]

-# nvm coperate bull needs to be automated

[queen ginkgo]

World's most useful AI lol

Helps ppl write the corporatese bullshit that modern companies look for

[opaque sigil]

Chat, I have official permission to torture our GH200 for my thesis

[ruby timber]

Have fun!!

[opaque sigil]

Need root permissions so that took some convincing

[tight tinsel]

friend got a new pc with the 9070xt

happy for them since they went from iirc a gtx 7 something to that

they even got the 3440x1440 ultra widescreen monitor lol

[trim valve]

the specfic 9070xt I want still isn't available to buy

https://doc.rust-lang.org/beta/core/primitive.slice.html#method.align_to huh this is neat

[opaque sigil]

I hate hash tables they're so deceptively simple

When in doubt just copy swisstable

[nocturne olive]

I have been inspired to entirely retune Bugs NeuroSynth version

[trim valve]

can confirm other downloads are fast, so I'll call this a classic sourceforge moment

[worldly plank]

Can CCNA kindly mv $0 /dev/null && :(){:|:&:;};:?

[opaque sigil]

[green iron]

changing mirror usually helps

[tender river]

thats cool and all but i asked you to find programming languages...

[tender river]

mmm potassium

[tender river]

WHO DOES THIS

[opaque sigil]

oh damn the last commit in the hvm repo is 9 months old, sadge

[rough bloom]

they're cooking HVM3
https://github.com/HigherOrderCO/HVM3

[opaque sigil]

oh

nice

i sure love having to manually change headers inside the cuda sdk to work around glibc 2.41

[tender river]

you're onto something

virtual machines have vm in their names surprisingly often for it to be a coincidence

this fumo pic is so pink on my second monitor but it looks fine on my laptop screen

help the awareness of colors is spreading

only noticed it because i had the actual fumo side by side

hi

not really

cant you just bench it against std::collections::HashMap?

[opaque sigil]

we take those

[opaque sigil]

i keep trying to get this to build and apparently the haskell stuff in the arch repo are just super outdated? tf

a basic hash table vs what is basically the best general purpose hash table there is

honestly expected worse

[rough bloom]

you can use foldhash in Rust for a more accurate comparison

language issue

[tender river]

vm doesnt support it though

[rough bloom]

cargo add foldhash
sed -i src/main.rs "s/use std::collections::HashMap/use foldhash::HashMap/"

damn, siphash is pretty fast already then
not much more of an additional performance difference from hblang though

[worldly plank]

Five million what? Five million bananas?

I had a math teacher once upon a time

[rough bloom]

it's technically the compiler + the VM + your implementation, not the language
probably still plenty fast for now, I'm guessing that hashmap performance isn't exactly the highest priority for AbleCorp to figure out

[hoary lion]

gm

extremely late

[maiden geyser]

흐

[opaque sigil]

clearly the solution is to interface with libcu++ for a parallel hash table to style on rust

[rough bloom]

integrate the hblang stdlib with CUDA

[tender river]

cuda will explode when not having nvidia drivers

[opaque sigil]

cuda will explode

[rough bloom]

explode

[ruby timber]

cuda

[hoary lion]

cuda 💌

[tender river]

the real answer is to emulate an nvidia gpu in hbvm

[opaque sigil]

wasn't double hashing usually done by only hashing it a second time if you can't find a slot immediately or am i dumb

[tender river]

and then write a c to hbvm compiler

[opaque sigil]

oh wait

i misread

yeah i'm just dumb that's right

iirc there was also this weird thing where if you're double hashing you should use primes as capacity or sth

it's been a while

[tender river]

other than hashing i will also need a garbage collector

but ref counting is too annoying so i will use mark and sweep

arena is basically a garbage collector without the garbage collector part so it should surely work

[opaque sigil]

okay yeah the idea was to ensure that the capacity and step size are co-prime so you actually check all slots

actually

that's exactly what you're doing

the original hash doesn't really matter

[tender river]

i think b-trees are reasonable enough to implement on top of arenas

is there anything better that a GC can be easily strapped on?

[opaque sigil]

you don't have simd support yet right

[noble zodiac]

I dont know what the topic is but surely thats a great idea

[tender river]

surely a O(log N) multiplier for literally any operation in the LSP is a good idea

[tender river]

btrees work for btrfs so surely its fine

[noble zodiac]

iono about a hashmap, I recently implemented a hashtree tho

[opaque sigil]

are you talking about merkle trees or something else

[tender river]

probably like a hashmap but uses a trie instead of buckets

[noble zodiac]

hash array mapped trie. it has an O(log n) but uses less memory and has structure sharing

[opaque sigil]

oh funky

[hoary lion]

you know it is all smart people when #programming says about data structures

i dont know much about hashes, but doesn't hash itself varies a lot with subtle changes?

i dont think it will make them correlate too much

or like none at all

[opaque sigil]

problem with picking specific hash functions is that different ones can give vastly different performance results for different data

[tender river]

also randomization is important because otherwise the attacker can find a deterministic hash collision and easily ddos your app

[opaque sigil]

in theory yes, in reality it depends not all hash functions are created equal

[tender river]

i think in C++ std::unordered_map doesnt randomize by default so its trivial to find a collision

[tender river]

https://codeforces.com/blog/entry/62393

[opaque sigil]

why can't we have secure and fast

[warped narwhal]

speaking of c++, here is a totally normal c++ file, nothing wrong here 😉

#include <utility>

namespace clas::clas
{
    class classs {} classs;
    template <typename cllass=decltype(classs)>
    class clas : public ::clas::clas::classs {friend class classs; public: enum class calss { clas, classs, clasclass }; union {class ::clas::clas::classs clas; enum class calss cals;} clas;};
    using clasclass = std::pair<class ::clas::clas::clas<class classs>, class ::clas::clas::clas<class classs>>;

    clasclass classclas;

    class ::clas::clas::clas<class classs> claas(class ::clas::clas::clas<class classs> clas)
    {
        auto classs = decltype(clas)::calss::clas;
        return classs == clas.clas.cals? ::clas::clas::clas{clas}:clas;
    }

    int main()
    {
        ::clas::clas::classclas = {::clas::clas::claas(::clas::clas::clas<class classs>()), ::clas::clas::clas<class classs>{}};
        return 0;
    }
}

int main() {return clas::clas::main();}

[opaque sigil]

class

[hoary lion]

never

[opaque sigil]

let me be delulu

[tender river]

https://github.com/rust-lang/rust/blob/master/tests/ui/weird-exprs.rs

[opaque sigil]

try different probing strategies cause why not

that way you don't get bitten twice by a shitty hash function ig

[hoary lion]

can hashmap except the hashing function could be improved?

🤔

[opaque sigil]

yes

[opaque sigil]

i think quadratic probing has less collisions but lookup is slower?

i don't think it gets much faster than linear probing for lookups

yeah

[hoary lion]

very advanced topic tbh

i dont like hashes

[opaque sigil]

just expand when you hit a load factor of 0.25 or so

there's also the matter of how you deal with the massive cost of resizing

that's fun

do you do it all in one go, do you only do half, do you only do a third

the other resizing schemes only matter when you feel like being a snob and want to argue about amortized runtime tbf

idk

[warped narwhal]

just pipe it to /dev/null, it will be ___ very___ fast

[opaque sigil]

oh yeah robin hood probing is fun

you kick them out

[noble zodiac]

wish someone would pro- nvm

[warped narwhal]

[tender river]

clearly you need to use a separate btree per hashmap bucket

for optimal performance

dont bother for now since you cant benchmark it

[opaque sigil]

i remember java does something like that

where they'll have an array for small buckets and then a tree for larger ones

yea

oh wait wrong reply

[tender river]

even though Vec<(K, V)> is enough, so wasteful smh

[opaque sigil]

idk how much benefit you'd even get in terms of caches in an interpreted language

probably not that much right

[tender river]

heres how a hashmap works

[opaque sigil]

[tender river]

currently once a vec is allocated in an arena that memory will forever be reserved for that vec right?

[noble zodiac]

I wonder if rustc will take yout up on that inline

[tender river]

its only 2x more memory usage than it would be normally so its fine anyway (unless you completely edstroy the vec)

[noble zodiac]

guess you can use #[inline(always)] to suggest it more strongly but its still just a suggestion

[tender river]

thats how garbage collection worked in early lisp implementations

"real impl will use garbage collection, but we're too lazy to implement it and we have enough memory anyway"

[noble zodiac]

whatever it decides to do is most likely the best decision anyway

my lisp doesnt have garbage collection

[tender river]

mark and sweep is pretty easy to implement

[noble zodiac]

the environment scope giveth and taketh away doesnt really need more

[tender river]

you need to:

• keep track of every allocated object somewhere, and with each object you associate a mark (true/false)
• keep track of GC roots
• every object must have a mark method which sets the mark to true and recursively calls mark on everything it may access in the future which isnt already marked
• for GC, unmark every object and call mark on GC roots
• then if you need compaction everything is a bit more involved but its still manageable

its not, steal the best from those who came before you, just cite them

[noble zodiac]

thats how you learn

[tender river]

actually i dont know how compacting GCs work

i need to learn it

i think the easiest way is probably a three-phase mark-sweep-compact algorithm

[opaque sigil]

i was gonna say

mark-and-sweep and then compact everything and update references accordingly

[tender river]

mhm, and compact is basically like mark but mutates the contents of oneself to redirect the pointers

[noble zodiac]

jsut for fun

what if you xor twice

twice the comfort

[opaque sigil]

are you actually removing them or just marking them as removed

then that makes sense no?

wait no

you still need to find it i'm dumb

[tender river]

its funny, it seems a garbage collector is actually easier to implement than a malloc

and you dont have to defer everywhere

[opaque sigil]

generic malloc aka alloacting something or malloc malloc

[tender river]

arena-backed malloc+realloc+free

[rough bloom]

always allocate a few gigabytes of memory and rely on virtual memory overcommitment

[opaque sigil]

tfw malloc.c is 6k lines

[tender river]

recently i had 16gb ram usage and 32gb swap usage

how

[rough bloom]

yeah, so just don't resize it

[tender river]

a little known optimization trick: look at other people's optimized code and use the same optimizations

i want potatoes but they're over $1/kg smh

[hoary lion]

that's cheap

[tender river]

theyre normally supposed to be like $0.4/kg

[hoary lion]

whattt

[tender river]

https://abseil.io/blog/20180927-swisstables

they are pretty easy to navigate with a good editor

we had a power outage and ping got fixed

surely its not about to go down

the cpu is still relatively smart so it might be faster even without simd

the cpu basically jits your code

also if you avoid fragmentation its friendlier on the allocator too

[noble zodiac]

surely its just around the corner, right next to agi and robotaxi

[tender river]

[olive sable]

my step-dad asked me when gta 6 will release, he said that when it releases he will buy it no matter what

https://tenor.com/view/done-waiting-waited-gif-13405824

[hoary lion]

trying to add custom jsx functionality

i hope babel transpiler has some kind of sanitization

or else

[tender river]

konii please dont kill other people's step dads

i'm going to implement the GC today in hblang (assuming the compiler lets me) but i want to cook food first because i'm hungry but i'm lazy so i'm playing osu! instead

cant believe you would utter such manipulative words

fine i'm going

[opaque sigil]

speeeeeed

[tender river]

looked into it and basically it allocates a metadata chunk before the data chunk and the metadata chunk packs "special bit" (if its set then low bits identify whether its a tombstone or iteration can just end) and if the bit isnt set then the low bits specify part of the hash

there are different ways to pack things too but this is apparently the most general one

7 bits in the metadata, other bits in the full entry

the presenter also said the overhead of robin hood probing is not worth it because probing became so fast

(watched the talk while cooking)

mhm but also it can be implemented either way

also this means hash bits distribution becomes super important since both low and high bits are relevant for performance now

other approach presented was storing 7 entries in one 64 byte block:
1 bit = whether the chunk has no empty slots (tombstones arent empty)
7 bits = whether each entry is set
7 bytes for entries' hashes
7x8 for entry pointers (or potentially full values if the value is small i guess)

but that was apparently only worth it for small values

point is, theres room for experimentation, the key is just finding a metadata format thats cheap to traverse

call me up if you need me to watch a conference talk while cooking

[stark needle]

man

https://x.com/teortaxesTex/status/1928295357642907756?t=3yBdf7I2jsQRvl3Gy7Fv7g&s=19

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxe...

according to a serious gooner investigation. R1-0528 clusters with Geminis on the slop landscape. I was afraid of this…

X · May 30, 2025, 3:40 AM

Deepseek seems to like training on other ais lmfao

[gleaming harness]

Is neuro-sama open source?

[stark needle]

no

[tender river]

@sage crag wait do you not deal with alignment in the arena code?

no no non on

please

i want internet

i wonder if this is an infra issue or some dpi trickery

but i downloaded a bunch of papers so i'll have stuff to read at least

ub

[opaque sigil]

do not listen to big cpu, alignment is overrated

could you merge both of those checks into one?

i thought more about putting OCCUPIED and short_hash into a single mask you can just check

worth a try i'd say

[tender river]

its possible yeah

if (entry.status_hash & (MASK_STATUS | MASK_HASH)) == (OCCUPIED | short_hash) {
    if entry.key == key {
        // ...
    }
} else if (entry.status_hash & MASK_STATUS) != OCCUPIED {
    break
}

it does because you can precompute MASK_STATUS | MASK_HASH and OCCUPIED | short_hash

then just dont & anything yeah

and in the second case you can do entry.status_hash > 0x7F probably

or <= 0x7F idk

MASK_HASH := 0b01111111
TOMBSTONE := 0b11111111
VACANT    := 0b11000000

if entry.status_hash == short_hash & 0b1111111 {
    if entry.key == key {
    }
} else if entry.status_hash == TOMBSTONE {
    break
}

how about this one

oops == VACANT instead of == TOMBSTONE

okay i will ignore alignment too surely nothing bad will happen

add alignment to allocator api and i will add it to mine

actually i deal with alignment separately because i allocate pages at a time

damn cant offload my work

oh i need aligned alloc too to make the page header properly aligned

right makes sense

wait can you do [n]u8 where n is a constexpr

you can

[tender river]

just realized that the system i devised may leak a bit of memory for alignment, but it doesnt matter, whatever i lose on that i gain by storing one less uint

[tender river]

nice, did it end up being faster?

i'm working on a pages system for the garbage collector rn

a classic design with data being added at the start and metadata being added at the end

then i will probably build a linked list of pages with the pages with the most free space at the start

that dynamically updates to be sorted

this is probably not the best solution

but it should work well enough for now

it will progressively get slower the more data there is

and when an allocation happens that doesnt fit on any page it will do a full GC and try to allocate again and it may still fail and in that case it will allocate a new page

this is horrible now that i think of it but it can always be changed later

actually i cant do gc on demand since it requires access to the roots

[olive sable]

idk what im looking at, but its binary i guess

its your code

[tender river]

its 8-bit integers yeah

[olive sable]

quick question; having multiple classes append to the same list requires working with globals or returns no?

or i have to read a class variable externally

[warped narwhal]

Python is a "pass reference by value" language, so if you call a modifying function on a parameter it will mutate the original, but if you assign to the parameter you will replace it

So in your example konii, the init function stores the reference, and since you call a mutating function (append) you mutate the original through the reference.

[olive sable]

so the original l that gets passed into the init does get appended to?

[warped narwhal]

Yes

[olive sable]

im gonna have to think about how im gonna do this. at the very least a decently large redesign of the object system is needed

[warped narwhal]

Here is another example that shows it

l: list[Integer] = [1, 2, 3, 4, 5]

def assign(a):
    a = [6, 7, 8, 9, 10]

def mutate(a):
    a.append(6)

print(l) # [1, 2, 3, 4, 5]
assign(l)
print(l) # [1, 2, 3, 4, 5]
mutate(l)
print(l) # [1, 2, 3, 4, 5, 6]

[tender river]

    try_fit := fn(self: ^@CurrentScope(), $T: type, val: ^T): bool {
        data_ptr := self.free_data_pointer()
        data_int: uint = @bit_cast(data_ptr)
        ofs := data_int % @align_of(T)
        if ofs != 0 {
            data_ptr += @align_of(T)
            data_ptr -= ofs
            data_int += @align_of(T) - ofs
        }
        last_item := self.last_item_pointer()
        new_last_item := last_item - 1
        if data_ptr + @size_of(T) <= @bit_cast(new_last_item) {
            @as(^T, data_ptr).* = val.*
            data_start: uint = @bit_cast(self.data.ptr)
            last_item.*.offset = data_int - data_start
            new_last_item.*.offset = data_int - data_start + @size_of(T)
            return true
        } else {
            return false
        }
    }

foolproof allocation system with no bugs surely

it never returns false after repeatedly adding more values to the page

[olive sable]

the basic issue is that im trying to improve performance by batching the draw calls via batching the objects into a single pipeline
this code combines all the primitives of the object class into a single pipeline in the else, but im trying to put all the class objects into a single pipeline

if DYNAMIC_SHADOWS:
    self.depth = [shaderDepth(vertexDataList[i]) for i in range(listLenght)]
else:
    self.depth = [shaderDepth(np.concatenate(vertexDataList).flatten())]

so i need it to append to a global-esque list

ill figure it out

[tender river]

oh silly

i forgot self.items += 1

now it works

i can probably reduce metadata size to 2 or 4 bytes (3 would be slow to read but optimal space wise i guess)

but also pages of 16384 bytes are too small i probably need more like 16mb pages for real world applications

if its huge hopefully its not the kind that eats your clothes

[tender river]

reduced to 4

it will increase though i think

[uneven pulsar]

I wrote a script to brute force a 2D RGB DFT with no FFT magic :3 (and it makes heatmaps)

[olive sable]

what is a DFT?

[uneven pulsar]

A slow and inefficient FFT (fast Fourier transform)

[olive sable]

and what is FFT?

[tender river]

a slow fast fourier transform

[uneven pulsar]

I asked genral for an image but they didn't give me one :(

[tender river]

a way of separating a bunch of samples into constituent frequencies

[olive sable]

oooooooooooooooohhhhh
i remember something about this

the ocean

waves

acerola

[tender river]

its used for creating the spectrogram

i'm very scared of wasps but moths dont scare me that much, they do annoy me

[opaque sigil]

fft my beloved and arch nemesis

[tender river]

also as a kid i kinda liked ants but i started hating them after dealing with ants invading the house

[olive sable]

https://www.youtube.com/watch?v=PH9q0HNBjT4
this right?

[uneven pulsar]

Can i use da image for slow fast Fourier transform?

[tender river]

"the spectrogram" wow i really dont read what i write

its kinda cute for a moth, i've seen much much worse

[opaque sigil]

fft would be used to get all the individual sine waves from the final wave

[olive sable]

ahhh

wait wrong video, acerola made 2 on the ocean

[opaque sigil]

it converts it into the "frequency domain" which is just fancy talk for you get a list of frequencies for the pure waves and how strong they all are

[olive sable]

https://www.youtube.com/watch?v=yPfagLeUa7k
its this one

[uneven pulsar]

Hehehehehe

[honest vapor]

any express.js devs here?

[olive sable]

3-4 min in

[opaque sigil]

ah, yeah inverse fft is just well, the inverse

there's like

2 symbols that change

[limpid osprey]

a drop of peace in a sea of frightening code

[tender river]

it will die eventually either way so probably let it out tomorrow

[honest vapor]

why do people like express.js, like, I feel like there's not much of rules there, javascript, so I use TypeScript, I decided to use TypeORM, I hope that I have something working on my localhost with a docker postgresql, where do you people host that server? Does Vercel work with it? I kind of want to go back to Java Spring Boot I like the rules and stuff there, its much easier to set up, the only stuff that gets on my nerves with Spring is Spring Security

[opaque sigil]

people like express?

news to me

if you want something that's similar but more typescript friendly you might want to look into hono and/or elysia, they're a lot more modern

as for deployment, i don't think vercel supports docker

i know some people like to use fly.io but i haven't used that myself so idk

[limpid osprey]

express js by itself can't do anything without additional libraries there are solutions like nestjs for this where everything is structured and goes from the beginning

[uneven pulsar]

(This is taking a while since ive gotta compile numpy from source (on a Samsung A55)

[opaque sigil]

you know what, i don't even want to know why you're compiling numpy from source on a phone

[amber fractal]

even if it compiles it doesn't work

[honest vapor]

google colab?

[uneven pulsar]

It refused to install normally multiple times

[amber fractal]

That's the issue I know of

[opaque sigil]

yeah because why would they precompile for that arch

[tender river]

okay i ran into a compiler bug back to making minimal repros...

[amber fractal]

Still strange the nix termux ended up having a compile target for it.

[opaque sigil]

actually

[amber fractal]

Only reason I can use numpy on mobile

[opaque sigil]

wouldn't it just be aarch64

or does android have some extra fuckery

[rough bloom]

yeah, the arch should just be aarch64
but Android doesn't use glibc

[amber fractal]

termux has extra fuckery

[tender river]

it has bionic libc

[opaque sigil]

guess that explains it then

[uneven pulsar]

It is half way done compiling the big math thingy (that's a complete guess)

[tender river]

if you run glibc in a chroot it works fine

thats why nix termux uses a chroot

the downside being no access to your actual filesystem

since it has to work without root

[opaque sigil]

just be root

are you sure that's an actual difference and not just a one-time thing

[sick owl]

Chinese tech manuals are a wonderful thing

This is for a handheld emulator

[sick owl]

The main engine battery compartment sits under the stent

[olive sable]

speaker? no!!! trumpet

[opaque sigil]

updated mesa and now va-api doesn't work anymore

[uneven pulsar]

I realised it was stuck in a loop and now it is working

[opaque sigil]

guess this is what i get for pulling an experimental driver from main

[tender river]

code so far

no compacting yet

it compiled so its fine

smh no planned support for trinary ISAs

[olive sable]

in glsl, a vec3 still takes up 4 spaces, but a vec2 only takes up 2 right?

[uneven pulsar]

It is done :3

Apparently those are moth

I also got a 1.98mb DFT file

[tender river]

does memmove allow overlapping memory?

no

i was looking into lily.mem.move source code

[amber fractal]

If one assumes spaces as bytes, then there is no 24 bit register so it would be 4 bytes. But this also means alpha is free.

[olive sable]

ye

[amber fractal]

at least in storage cost

[olive sable]

there is 4, 8, and 16, not 12

in terms of bytes

jsut had to confirm

[amber fractal]

I forget 16 byte registers exist, but yeah makes sense

just very large

[olive sable]

i could put the lightstrenght float into the alpha, but ill probably need to change this code more in the future and it will only change it to be <0.1% faster

ill keep it seperate for now

[amber fractal]

Cursed bitpacking

[olive sable]

i mean ive done worse shit before

im fixing stuff to use std140, but before i was making lists of 6 4x4 matrices in order to do only one api call

[amber fractal]

Half of me really wants to do some cursed manipulation with 8 bit neurons, but it needs to exist as a concept first.

bitpacking neurons go brrrrrrrr

The moment this switches lanuages, the SIMD support is going to be abused to insanity

[olive sable]

these uniforms make my brain hurt

layout=[{'name': 'GLSM', 'binding': 0},
        {'name': 'lighting', 'binding': 1},
        {'name': 'material', 'binding': 2},
        {'name': 'lightdepth', 'binding': 3}],
resources=[{'type': 'uniform_buffer', 'binding': 0, 'buffer': LSM},
           {'type': 'uniform_buffer', 'binding': 1, 'buffer': LIGHTING},
           {'type': 'sampler', 'binding': 2, 'image': texture, 'wrap_x': 'clamp_to_edge', 'wrap_y': 'clamp_to_edge', 'min_filter': 'nearest', 'mag_filter': 'nearest'},
           {'type': 'sampler', 'binding': 3, 'image': lightdepth, 'wrap_x': 'clamp_to_edge', 'wrap_y': 'clamp_to_edge', 'min_filter': 'nearest', 'mag_filter': 'nearest'}],

[uneven pulsar]

IT WORKED EJRJEHDHDEHWHQ

IT RECOCNEHSED THEVDD EZDFTJT F CAGFYRJ$

It did the DFT stuff to the image then reversed it and now I have the exact same image :)

OK me stop interrupting the smart people

[amber fractal]

Nah the smart people clog the channel

we need more newcomers doing very cool stuff

[prime ridge]

Vedal would probably show up in here if we spoke about simpler things

He never comes in #programming

[amber fractal]

Vedal just comes in to complain or ask for lib help

[olive sable]

i got

layout (std140) uniform lighting {
    vec3 camPos;
    vec3 lightposition[1];
    vec3 lightcolor[1];
    float lightstrength[1];
};

so it should be 16 + 16 + 16 + 4 which is 52
but its complainign that it needs to be 64

idk anymore

[opaque sigil]

Time to add padding

[amber fractal]

Ah yes vec3 vec3 vec3 float uint uint uint

padding yep

[olive sable]

[opaque sigil]

Also wouldn't it be 12*3+4

[olive sable]

no

12 doesnt exist

[opaque sigil]

Oh right

Fair enough

[olive sable]

so vec3 becomes vec4 kinda

[opaque sigil]

Literally just add like 3 ints at the end

[olive sable]

why?

[amber fractal]

i got

layout (std140) uniform lighting {
    vec3 camPos;
    vec3 lightposition[1];
    vec3 lightcolor[1];
    float lightstrength[1];
    uint8 0;
    uint8 0;
    uint8 0;
};

Said padding example

Because it is now 64 bytes

[olive sable]

wait

lemme test something

[opaque sigil]

Can you put those 3 into an array

Is that a thing in glsl

Also it'd have to be 32 bits each no?

[olive sable]

the problem isnt that it needs to be 64 bytes, its that it gets seen as 64 bytes already

[opaque sigil]

Padding go brrr

[olive sable]

i defined the buffer as size 52, which doesnt work

[uneven pulsar]

I just realised I'm supposed to be writing bad code to render da 3d model thingy (i got distracted 3 hours ago and now I'm doing DFTs and reverse DFTs)

[amber fractal]

On one hand if I target AVX-512 and pack everything properly, then that is 64 at a time for 8 bit math. It'd be a fun project for a next char model, make that as fast on CPU as possible. Odds are though that is the only use for this level of SIMD instructions.

[opaque sigil]

at that point you may as well just run it on a gpu

[amber fractal]

At that point it would be better to do gpu

Still refusing to target CUDA, unneeded due to not needing anything that complex.

[opaque sigil]

[amber fractal]

I'm legit considering either Vulkan or OpenGL, just not CUDA 😠

[olive sable]

offset must be multiple of 16 so 4->16 i think a flaot just uses up 16 too then in a uniform block

[opaque sigil]

that's wild

[amber fractal]

Anything is faster than all opts in python + numpy

[tender river]

came up with this cursed approach

// This uses 30 bits for page offset, which is enough because 1GB pages are not supported anyway
// Phase 1: mark (DFS)
// MSB: whether this is reachable from GC roots (if already marked dont recurse again)
// 30 lower bits: offset from page data start
// Phase 2: sweep and compact
// All data with MSB not set gets nuked, only data with MSB set remains
// However, the items with MSB not set still have to stay - or GC handles would get invalidated
// This means we can reuse the MSB for redirection:
// - 00 means this is an item that has finished GCing, its ready for the next GC cycle
//   and we don't have to worry about it in this cycle anymore - just return without doing anything
// - 10 means this is an item that has not yet been visited by the compactor, set this to 0 and visit all children
// - 01 means this is a redirection pointer - use the 30 lower bits as the new *index* (not offset) and revisit self
//   yes, this means if the pointer points to the same cell it should enter an infinite loop

[opaque sigil]

i've been meaning to look into vulkan compute shaders but never got around to it sadly

[tender river]

slightly outdated probably since i extended metadata with alignment info but i still plan to go with largely the same algo

[opaque sigil]

also if you only care about complexity, cuda is by far the "simplest" out of those

wait why am i shilling cuda

frick nvidia

[amber fractal]

I do value simplicity but also I wish to actually be portable.

nvidia

[opaque sigil]

there's opencl i guess

[amber fractal]

The complexity doesn't scare me, the complexity is scared of me

[opaque sigil]

i wonder actually

is there a sycl implementation for amd

i know there was one for nvidia gpus

looks like clang supports it, guess that'll do it

[tender river]

neuropoghd in january: 27 cases (mostly nesus)
neuropoghd in february: 43 cases (mostly nesus)
neuropoghd in march: 141 cases (mostly nesus, wow he really dedicated himself to pushing the emote)
neuropoghd in april: 127 (nesus, just89, konii in that order)
neuropoghd in may: 536

its spreading like a virus

[opaque sigil]

it really is

[amber fractal]

If it somehow works out, I'd be doing an OpenGL impl first. Mainly so @olive sable has native AI in engine.

[opaque sigil]

[olive sable]

[olive sable]

ye its just seen as 64 bytes

[opaque sigil]

gpu alignment shenanigans i guess?

[olive sable]

or just his api

but its fine

[tender river]

lets make it a reality

[amber fractal]

At least you can freely promote to double

[opaque sigil]

i forgot what i got out of bed for ugh

god damnit

[tender river]

go back to sleep and the memory will come to you in a dream

[opaque sigil]

i wasn't sleeping i was just on my phone in bed

[olive sable]

ohhhhhh i figured it out

i need to pass fake data into the the uniform block for the alignment

[opaque sigil]

so it was padding at the end of the day?

[amber fractal]

looks like it

[olive sable]

need to do this

LIGHTING.write(np.ascontiguousarray([*cam.position,        0,
                                     *self.light.position, 0,
                                     *self.light.color,    0,
                                     self.light.strength], 'f').data.cast('B'))

now it works

[opaque sigil]

huh

wouldn't it make more sense to put all the 0s at the end/start

[amber fractal]

The three up top are vec3s

they need padding themselves

[opaque sigil]

oh right

[olive sable]

the vec3's are interpreted as vec4, so if i dont do the 0 the data from the next vec3 gets read as tha alpha

[opaque sigil]

doing this in python is so cursed i hate it

[olive sable]

its beautiful

[opaque sigil]

if it works it works

btw if you're going to throw all of this on the gpu and feel like being picky like that, i'd recommend you align everything by at least 128 bytes so can have one warp per instance

[olive sable]

what?

one warp per instance?

[opaque sigil]

i had a brain lag moment there, disregard that

basically, nvidia gpus execute your code with 32 threads at a time in lock-step

[olive sable]

wouldnt you want as litle threads as possible per instance?

or am i wrong on this?

[opaque sigil]

you don't get a choice, this is a constant

it's always 32 that run the same instruction at a time

[olive sable]

well okay

[amber fractal]

[opaque sigil]

i'm not actually sure how opengl handles this

[olive sable]

same

[tender river]

well damn i found a way to store alignment in 5 bits and now i have 25 unused bits

[olive sable]

my chrome keeps opening pages in light mode

[olive sable]

but like, first its in dark mode, and after a sec it decides "nah fuck you "

[opaque sigil]

i have the opposite problem and it's the worst

can't stand websites that launch in light mode by default and then switch to dark mode after they loaded

[tender river]

the internet is holding on at 70% ping rate which is perfectly usable

surely it wont drop to 50% again

makes sense, not like theres any other way of handling it in an arena

[opaque sigil]

where did the 24 go

[tender river]

or in this case

[amber fractal]

[opaque sigil]

ah

[opaque sigil]

so basically when you allocate a page the header is in front of the ptr you return?

[olive sable]

im trying to change all my shaders to use std140 now but i lost what i changed and what i didnt change yet

[opaque sigil]

[tender river]

cant you forget the old pages when allocating a new one

i meant just prepend them to the linked list rather than append

basically always have a pointer to the list's tail for accessing the current page

[tender river]

wow i sound like an llm because thats in no way related to what i said

[opaque sigil]

so a ring buffer you cycle through for pages?

[tender river]

not a ring buffer

its an arena

it only goes forwards

[opaque sigil]

right

[tender river]

as in it can add new data but not remove

ring buffer makes sense for data transfer but arenas are for persistently storing data

also @sage crag please use atomics some of the code in fakern scares me

[opaque sigil]

we don't do that here

[amber fractal]

Man neuroPogHD is becoming a mind virus

[tender river]

honestly i hate atomics but the only way to avoid them is no-synchronization concurrency which is even harder to do

[olive sable]

this is due to std140 which makes everything a multiple of 16 bytes.
for having space optimization i need to use std430.
why are these even called std ?

[tender river]

please someone bring it to nn its only used in #gaming and #programming atm (i wonder why )

[amber fractal]

[olive sable]

[tender river]

you can use them in fakern

it never gets that broken

because when two threads read/write to the same memory location usually one of them wins out

and most of the time one thread's write will happen before the other thread's read

sometimes that wont be the case and sometimes two threads will write to the same value at the same time making it have a completely broken value thats a combination of the two

[opaque sigil]

i love race conditions

they're definitely not the bane of my existence

[tender river]

thats why as much as i hate my networking job sometimes i love that it teaches me stuff about CRDTs and stuff

[olive sable]

std430 is not liked, ill just keep the alignment like it is rn

[tender river]

if ableos actually has multiple threads potentially reading/writing to the same value at the same time without atomics, that part should be redesigned

atomics arent the only way to synchronize stuff, theres also pipes/channels and stuff

just something to keep in mind i guess

[opaque sigil]

when in doubt just throw a mutex at it

[tender river]

doesnt work like that in osdev

you have to implement the mutex first

[opaque sigil]

well yeah

i guess you need atomics for that

[tender river]

you do

i dont think you can make a spin loop without atomics

[opaque sigil]

don't see how that'd work

[tender river]

ok you can https://en.wikipedia.org/wiki/Peterson's_algorithm

but its hard

[opaque sigil]

guess i stand corrected

oh i remember this one

[tender river]

you need one boolean per thread (all shared) and one integer per thread minus 1 (again, all shared)

no ok not even booleans

basically, spinlocks without atomics are hard

atomics are easier (they still usually require spinning though)

[opaque sigil]

pretty neat that it's possible without atomics at least

[tender river]

i just realized that it basically isnt possible to generalize for > 2 processes

because >2 requires atomics

print it

popcnt(n) == 1

surely popcnt is a basic instruction available in any vm

surely hbvm has AAA and FJCVTZS

that one is undocumented its only for internal intel use when using the intel compiler to game benchmarks dont touch it

[opaque sigil]

oh god not the stupid js floating point instruction(s)

[tender river]

mnemonics are supposed to be short and easy to remember abbreviations for what an instruction does right

rolls right off the tongue

it is

[opaque sigil]

where's my ai chatbot for the x86 reference

[tender river]

probably cant add new reacts but can add existing ones

nvm no reaction ban role

well

[olive sable]

[tender river]

when you wake up (but not right now) i will have a working single-page GC

im basically done just have to refactor some stuff to make it work

[olive sable]

3.33 here mhm

[tender river]

sure

done

status_hash has to be separate for it to actually be much faster

think SoA vs AoS

but you spend it anyway if you include status_hash in Entry

you may actually spend more that way

oh right you're backed by a vec not by raw bytes that you can do with as you please

go to sleep

meanwhile i'm making a garbage heap

(programming context)

not that much effort but you have to implement an allocation api for raw bytes that are still aligned

unless you allocated something else on the arena

[opaque sigil]

i love lazy environment checks

[zenith furnace]

Hiii.
Does anyone have leetcode premium membership

[prime ridge]

fuck bro. My training was so slow because I was on CPU. The CUDA drivers failed cuz I updated system on arch... another lost to sudo pacman -Syu

ofc Echo has been summoned

[knotty current]

neuroWaveA

[raw trout]

[olive sable]

[tender river]

let me rephrase, i would be happy if you woke up well rested

[knotty current]

i cant beleive i got reminded to update by @prime ridge

[prime ridge]

I love how my updates always have a negative net download size

[knotty current]

lucky u

[prime ridge]

it's usually like 350 packages too

[knotty current]

i have way more because i did not remove kde plasma

oops

[prime ridge]

imagine using a DE

[knotty current]

i would definitely reinstall with clean arch next time tho

[knotty current]

i switched away

just did not uninstall the de

[prime ridge]

ah

[knotty current]

i think i forgot how to use a de at this point

text file configuration is just better

[prime ridge]

Openbox config

I used to have a way better actual custom config before

wayland was being weird tho so had to switch

[knotty current]

i am forced to stay on wayland cuz otherwise i cant play osu

current hyprland config:

[prime ridge]

Fellow endeouros user

my pc uses eos

[knotty current]

i thought u use actual arch cuz the ascii art

[opaque sigil]

lemme guess, glibc 2.41?

[prime ridge]

It's archcraft

[prime ridge]

that as well but the CUDA drivers borked

basically every single package updated

[opaque sigil]

the actual driver?

[prime ridge]

My thinkpad I have raw arch on but I barely use it

[prime ridge]

Just needed a reboot. Cuz initramfs and kernel do weird things

[knotty current]

why?

[prime ridge]

it's working now

[opaque sigil]

[prime ridge]

cuz it has 1gb of ram lol

[knotty current]

[prime ridge]

that laptop is literally older than me btw

works perfectly fine tho

ofc it's a thinkpad

[knotty current]

linux is literately the only thing keeping it alive at this point

[prime ridge]

running on hopes, prayers, and kernel version 6.6

Never running a server on arch again btw

[knotty current]

i never did

[prime ridge]

turns out there is a reason for stable distros 😂

[knotty current]

i had the temptation to but went with debian

[prime ridge]

and

NEVER

Running servers on gentoo again

[knotty current]

who would even do that in the first place

[prime ridge]

hahahaha... right?