programming

[olive sable]

[tender river]

the reason i dont like these concepts as much anymore is because i realized how poorly imperative code composes

but its the best thing we have for bringing C programmers out of the gutter

[olive sable]

i was wrong on this btw
small triangles are bad cuz of wasting quad samples while rasterizing.
this doesnt mean you should have no small triangles, this means you should maximise triangle area

so left bad, right good

[trim valve]

*except for nanite meshes

[olive sable]

we need to go on a pilgrimage to NN

fumes?

is NN toxic?

[prime ridge]

a lotta rage baiting in NN

Can't last more than 20 seconds there

[olive sable]

[stark needle]

i introduced bwaa, konii introduced neuropoghd

[olive sable]

branding

it would be funny to brand game-jam 3 game under "bwaa studios"

i need to buy that domain ngl

im poor tho so nah

[rigid snow]

bwaapoghd

[olive sable]

enub is just konii in my head

cuz short

[tender river]

just checked wikipedia and apparently my height is a couple cm above average

[olive sable]

like regioanl average? or just average global?

[tender river]

regional

(i'm below 170 but idr the exact value, something like 168 or 169)

[olive sable]

according to wikipedia the average male height here is 176.8cm

so my 183cm is above

[tender river]

well i assume async wont be supported right?

[trim valve]

[tender river]

epoll is a beast

and theres io-uring too how fun

async doesnt require threads

its just polling the kernel for new data for multiple files at the same time

yeah

i think just copy the posix api open read write, maybe creat

will file handles be platform independent?

or will ableos use some special thing for representing files

on posix-compliant systems its just an integer

just write a File object that wraps a file handle, write (write data from a slice, returns how many bytes it actually wrote and may have to be repeated), read (read data from the file into a slice, returns how many bytes it read), open (flags like read, write, append, create should ideally exist, idr much else)

i think for an actual good lsp i might have to use epoll to read files and process requests in parallel but the good thing is i just want a working lsp not a good lsp for now

also paths are their own can of worms but lets just work with unix paths for now

also stdout/stdin should ideally have the same read/write API as files (should be natural since stdin/stdout are just files internally)

oh also flush is important

and the bad thing is, all of the io operations can fail

and i dont want the lsp to crash when it fails to read a file

oh and a rename (move) api might be useful for atomic filesystem writes

[trim valve]

hm do I need to prepare a june variant of my cat pfp

[tender river]

yeah i saw something about linking to C

its cool but wheres the fun in that

[olive sable]

i have mobile detection now i guess

[tender river]

you can at least optimize by copying in 8 byte chunks but not much more i guess

or or or

need a way to jit code

then you can just write the assembly

actually

doesnt hblang have function pointers

if so you can already do it

yeah but you have to find the right one

find a compiler bug that emits wrong assembly

[olive sable]

grass and hair is pretty similair if you think about it

[tender river]

can you get a function address at least

how do i compile for x86

wait i just realized i can probably overrun the stack and execute arbitrary code

this is clearly the best solution for optimized memcpy

well, time to derust my binary exploitation skills

i think ROP is the most realistic approach here

[olive sable]

works

tablet still gets recognized as not a phone so i might need to fix that idk

with shadowmapping the tablet gets 70fps, and the phone gets 0.5fps. so for phone it needs to be turned off, but on tablet its fine to be on imo

this depends on the tablet tho so i might have to fix this someday

[tender river]

oh no... this code sets up a []u8

i think it might be fine at the end

i just want a buffer on the stack that i can overrun...

no it has to be on the stack

i already did it but i just want asm thats easier to work with

[olive sable]

i was wondering why my heightmap didnt work on mobile

In Google Chrome on mobile, the maximum texture size allowed for WebGL is generally 4096x4096 pixels.

why must mobile always cause extra issues?

https://tenor.com/view/he-cant-keep-getting-away-with-it-gif-19335672

[nocturne olive]

Because SMOL

[olive sable]

@ bwaa fix chrome please

[tender river]

progress

[nocturne olive]

You may just have to create a tiled texturing system specifically for use on mobile devices

[tender river]

i successfully jumped to the shellcode but it segfaults on first instruction

probably have to make the page rx (or rwx)

[opaque sigil]

fun fact, mobile firefox maxes out at 2048x2048

just checked

guess it depends on the gpu and/or driver

or maybe not actually, chrome says 4096 idk

[olive sable]

FUCK FUCK FUCK

https://tenor.com/view/throw-out-rage-parks-and-rec-nick-offerman-ron-swanson-gif-17688925

why must things be this way

i just want heightmaps to work

do i just make a lower quality version that gets used on mobile? i guess that works...

[opaque sigil]

is cutting it into multiple textures not option?

[olive sable]

kinda

[tender river]

please back me up i'm bullied by the compiler

o you think uint is a type

[olive sable]

damn, google has AI code boxes now. and the code is not too bad

[olive sable]

idk why it feels the need to split the colour channels tho

[tender river]

this is all i did

yeah i tried a bunch of things but they didnt work but now that i return @syscall it works

thx

(i tried assigning the result to a variable before, that didnt seem to work)

grr it still segfaults

why wont the cpu just let me have fun

[opaque sigil]

Chrome & Firefox dump courtesy of gpt, idk how useful this is but here you go @olive sable

[olive sable]

thanks

[opaque sigil]

my phone is a bit unusual since it has an amd gpu tbf

[olive sable]

wdym webgl 2.0 not supported?

what

thats crazy

[tender river]

i guess i just cant execute that part of memory

i should copy it somewhere

[opaque sigil]

the test page could also just be wrong idk

[stark needle]

Mannnn there was an australia ahh looking ~8cm arm span spider under my bed😭😭

[opaque sigil]

okay yeah i think something went wrong with the webgl2 check

desktop firefox should support it but it also says unsupported

[tender river]

it worked when i allocated a separate chunk of memory

it just didnt work when i used the binary's static memory

time to strip the size of the shellcode executor down and like make it actually return control to the caller

[olive sable]

it seems to say webgl 2.0 not supported on my side too
eventho it should be supported on desktop chrome

[stark needle]

https://futurism.com/computer-science-majors-high-unemployment-rate

I love massive unemployment

[opaque sigil]

[tender river]

memcpy performance is more important than whatever useless code the users are writing

[opaque sigil]

i'm getting different results now

[olive sable]

it seems to be chrome is limited to 4096, and firefox should be fine
so on firefox my terrain should be rendering and chrome it shouldnt be

[deft pike]

Does anyone know what pathfinding algorithms games like StarCraft, League, or other MOBA's use?

[tender river]

A* is the one you should know

you might also want to separate "global" pathfinding and per-chunk pathfinding

[deft pike]

Thanks

Yea dijkstras seemed too slow

[tender river]

its faster if you use a pqueue

[opaque sigil]

Take new dump courtesy of claude, no guarantees though @olive sable

[tender river]

the canonical implementation uses a vector but you can replace an O(N) with an O(log N) if you uses a pqueue

[deft pike]

Pqueue?

[tender river]

priority queue

[opaque sigil]

fib heap

[tender river]

a collection where you can add an element with a priority and remove the element with the lowest (or highest if you want) priority

[opaque sigil]

the file

[olive sable]

seems to check out ye

so 16K on firefox, 4K on chrome

[opaque sigil]

[olive sable]

This is new

so good news, i can see the terrain being rendered on firefox

bad news, why is it so SMOL????

[nocturne olive]

SMOOOOOOLLLLLL

[olive sable]

this is very interesting data, but i think none of it matter too much until you hit the cap

[olive sable]

Went away on its own somehow

weird

oh well

[opaque sigil]

If it's good enough for musl it ought to be good enough for me is my motto

[nocturne olive]

Code: Un SMOLifies your SMOL

[safe path]

wow unemployment

hello chat

[olive sable]

hi

[safe path]

i am here to once again complain about how perfect the linux sound ecosystem is

it's so good, i love it so much

explain to me how it is possible for sound to travel large distances, to twitch servers, then down to some random persons computer faster than it goes through my sound hardware then into my ear

i am literally just monitoring the audio directly how in the world do i have 8 seconds of latency compared to somebody who is across the planet

how

raaaaa

[opaque sigil]

8 seconds of local delay is crazy

[safe path]

bro i wish i was joking i actually have video proof

[rough bloom]

just switch to JACK

[tender river]

i'm about to add simd to hblang without even modifying the compiler you cant stop me

[safe path]

hblang does not have SIMD yet

unusable

[tender river]

and it wont

[safe path]

konii explain this

[opaque sigil]

Wasn't it 30% yesterday

wtf happened

[tender river]

i successfully made the shellcode return control... to libc

[safe path]

konii have you considered implementing SIMD anyway

it would be funny

but

fast

speed

wow the efficient local inference LLM i was going to write in hblang is no longer possible

this is true and not 100% made up, but it doesn't change the fact it is no longer possible

[opaque sigil]

c interop surely comes to save the day

[safe path]

@import("tensorflow")

can i link hblang to the python runtime and then run pytorch

[opaque sigil]

Inline asm/bytecode when

Do you?

[tender river]

i succeeded

i returned from the shellcode

[faint sandal]

what the hell are you guys making

[tender river]

mov rax, 0x3C is to demonstrate that the shellcode does something

and then it just returns from the original function

[rough bloom]

doesn't hblang have fancy comptime stuff?
just launch the assembler

[rare bramble]

solve the halting problem in comptime

no need to worry about runtime execution speed

[tender river]

ideally i'd make it so after running the shellcode overwrites the original function with what it's supposed to be and jumps there

[rough bloom]

hblang x86 emulator...

[tender river]

if they do then what i just did is kinda pointless

i kinda felt like that while implementing the gc

[rough bloom]

I know that it runs natively on x86
I mean embedding x86 programs to run at comptime or in hbvm so that you don't need to rely on native FFI or IPC

[tender river]

hm

because they only exist in a meaningful fashion at comptime probably

isnt return code just the lowest byte

okay i guess all i have left to do is make it actually do something useful rather than just write to eax

[low cedar]

hello can anyone help me with a python project im working on? its basically a website where i can calculate pp using oppai-ng based on inputted mods, acc, and # of maps wanted. however, i can't seem to get the files to download properly

[tender river]

i guess you can do something evil if you know what order the compiler puts symbols in

[hoary lion]

😠 I am not getting the concepts of flax

[tender river]

i wrote assembler code that adds two numbers

well thats easy but it runs in hblang

code:

    mov rax, [rsp - 0x30]
    push rbx
    mov rbx, [rax + 8]
    mov rax, [rax]
    add rax, rbx
    pop rbx
    jmp [rsp - 0x20]

func := fn(arg: ^uint): uint {
    a := arg
    f := &a
    // run_shellcode didnt change
    return run_shellcode(@bit_cast(f))
}

main := fn(): uint {
    arg0: [2]uint = .[1, 9]
    ret := func(arg0.ptr)
    return ret
}

[opaque sigil]

[olive sable]

[tender river]

and if i change it to return ret + 1 it actually returns ret + 1

[hoary lion]

atlas
10M tokens probably not scalable like previous titans but still

[tender river]

so uh do with this what you want
simd, syscalls, blowing up the computer, vcvtne2ps2bf16, whatever you want

you are now free

❓

ah depends on known function address

my solution is less prone to breaking

it still can break easily

my solution doesnt depend on known addresses it depends on stack offsets staying the same

they can change across compiler versions though

i dont really think i can do better without the ability to call arbitrary addresses

actually i might

i just dont want to

because the solution is looking for the nearest integer in the stack that points to executable memory

and then patching the shellcode to return to the correct address and use the correct input address (both are pretty easy, in fact you can just make the input be a global variable since everything is single threaded)

wait theres a better solution

its adding one more indirection, func2

then then idk i lost chain of thought

oh right

then you overwrite func2's return address to return to the shellcode instead of func, but the func caller's address will stay in the stack so you can just ret

but this depends on the function not being optimized to a jmp

anyway this is cursed and pointless so i'm done

gonna actually do some work for once

right doesnt it do dfs from the entry point

i mean, if you can run stuff on program startup everything becomes a bit easier

one of the ideas i rejected was making the shellcode rewrite its calling function with itself on initial execution and then jump back to main to restart the program

it would work but might execute some code twice so i rejected it

but it would be worth it for the sheer cursedness

the code i wrote for addition in asm is enough to implement literally anything, no additional code required

because it passes a []uint to the assembly code

which can be used to encode arbitrary data like first item being the function to run and other items being arguments

thinking of ways to use it for evil

wait i just realized i dont need static linking if i have dynamic linking and a shellcode to call arbitrary functions

const char**, probably somewhere in libc

found this code

char **get_argv()
{
  static char **preset_argv=NULL;

  if (!preset_argv)
  {
    extern char **environ;
    char ***argvp;
    void *p;
    asm ("mov %%esp, %0" : "=r" (p));
    argvp = p;
    while (*argvp != environ)
      argvp++;
    argvp--;
    preset_argv = *argvp;
  }
  return preset_argv;
}

extern char **environ is key i guess

oh environ is on the stack too

oh here https://github.com/phip1611/linux-libc-auxv

[olive sable]

apparently im able to get context info like max samples and stuff, but not the max texture size

[gleaming harness]

“Awww dammit this system is too hard to crack”
“Your fired”
“What?? Why?!”
“This man is running wfw 3.11”
“… you’re joking”
“Nope”

[olive sable]

what does this mean?

[gleaming harness]

It means that no firewall matter for me because I’m only safe because hackers just forgot about me

||And also because I made a huge lan brick with a firewall antivirus and htTPS to HTTP encode and decoder ||

[uneven pulsar]

I have to learn rust :(

[gleaming harness]

Nope

It’s not on Windows 3.11

Hell, they just barely still support Windows 7

[olive sable]

why do we have people asking to play roblox now here?

[gleaming harness]

Don’t ask me I’m so tired from getting this application running on Windows 3.11

You’re just unnecessarily toxic

It really isn’t

[olive sable]

sure bro

[gleaming harness]

And I bet in everything that you’re a nine-year-old who asks your mother for fruit snacks every morning

Surrreee

[olive sable]

https://tenor.com/view/plankton-eating-popcorn-popcorn-plankton-meme-plankton-yes-gif-18251234252356731757

[gleaming harness]

Are you drunk?

Like bro, you literally just joined

And you just want to ask about Roblox???

And your discord account has existed for three fucking days

[open copper]

i killed him

ignore

ill start deletiong

[hard shale]

Lol

[olive sable]

i was wondering why it said no mutual servers lamo

[gleaming harness]

LFGGGGGGGG

Let’s be honest he probably was under age

Like who goes into a programming channel in an ai server.. goes up to someone running fucking Windows 3.11 and says “ wanaa plaiy rooblx”

[hard shale]

Doesn't matter. Too late now

[olive sable]

as long as you're 13+ its fine, just dont be a an asshole i guess

[gleaming harness]

He was acting like a 5 year old 💀

[real sierra]

was a throwaway account

someone just flexing their unemployed status

pay it no heed

[gleaming harness]

I think it was both

They’re probably just gonna come crawling back and ask to play roblox again

[olive sable]

just move on

dont give them attention and they'l stop

[gleaming harness]

Or leak the fact that they live at Celtic point worksop S81 7AZ (look up before ban or mod ping and you’ll see the joke or just click here ||its a fuckin daycare 💀||)

[olive sable]

damn

you had me in the first half ngl

[gleaming harness]

Lol

[olive sable]

[gleaming harness]

Ngl they probably do go there every day for about five hours before getting picked up by two deeply in love heterosexual adults with opposite genders that are related to them

A.k.a. a very in-depth description of a mother figure and a father figure

And them going to daycare

[olive sable]

dude

stop

what even is that supposed to mean?

[gritty dust]

what happened?

[nocturne olive]

Thoughts on how this is going so far?

[tender river]

stdout works, and division will probably make x86 basically 100% usable

not counting potential compiler bugs

[gleaming harness]

It’s a very in depth and detailed explanation of a child going to daycare and being picked up by their parents after five hours of being there

[olive sable]

was a rhetorical question but ok

[nocturne olive]

-# Hopefully I'm getting better at this tuning thing

[prime ridge]

This is just ridiculous for a single problem

35 sections to a single question...

[olive sable]

blender shit

[tender river]

interesting

it probably will

wait i guess i have to support \u in json

[gritty dust]

screw it close enough its better than before at least

[real sierra]

ive tasked myself with implementing right bit shift

with subpar hardware support

why is it so much worse than left bit shift

[olive sable]

just left bit shift enough till you right bit shifted

[real sierra]

if i had left rotate, that would actually work

could left-rotate 15 times, and then apply a mask that zeroes the MSB

unfortunately for me, left rotate is very hard when you have uhh

...two registers

[olive sable]

2 registers huh....

[real sierra]

[olive sable]

[real sierra]

so instead

im completely leaning on the stack for this

going to do what's effectively a copy of the upper 15 bits of the input to the lower 15 bits of the output

but loops are hard (two registers!!!)

so im doing it recursively

stack frames are basically free memory anyway right

[tender river]

one = stack pointer
one = instruction pointer
and hopefully two more?

[real sierra]

instruction pointer is hardware so i dont have to worry i guess its technically a third register

stack pointer is not hardware, its just at a fixed address in ram

the other two registers are for general stuff

[tender river]

sounds slow...

[real sierra]

you would be correct

just wait till you see the instruction set

[tender river]

i dont expect multiplication from what you said

[real sierra]

the ALU supports

• addition
• subtraction
• increment
• decrement
• AND
• OR
• XOR
• NOT

[tender river]

though multiplication isnt that hard, its just bulky, a lot of bit shifts and adders

[real sierra]

pfft

nice try

i already implemented multiplication recursively

[prime ridge]

depends on the ISA

[tender river]

i mean yeah you can just use the binary exponentiation algorithm for multiplication too

...but you'd have to have bit shift for it to actually be fast

or, well,work

[real sierra]

yeah...

lack of bit shift is hindering a great many things

which is why im currently implementing it

left shift is easy enough, just add something to itself

right shift...

[tender river]

multiplication of A by B is basically:
for every 1 position in B, shift A to the left by that amount of bits and add to the result

child's play compared to division

though admittedly i don't actually know how division works

[real sierra]

the naive way is to subtract the divisor from your value until the divisor is negative

the number of times you can do that, minus one, is your answer (if we're doing integer division)

[tender river]

i mean yeah or you could bin search the answer if you have multiplication

[real sierra]

that seems kinda annoying to work out the logic for

if you're doing something like 15/2

there isnt an exact answer

actually i guess you know how many steps your bin search will take

shouldnt be an issue

[dense cosmos]

What are you working on, that you have to implement such low level details, if you don't mind me asking?

[tender river]

its easy to make a bin search algo for lower or upper bound

[real sierra]

i remembered i had a game that lets you assemble logic circuits and put this together

[dense cosmos]

Y not add a module for right shift, if you need it?

[real sierra]

not sure i have instruction bits to spare

[dense cosmos]

Ohhhh

[tender river]

okay i could read this if i wanted to but i dont https://pages.hmc.edu/harris/research/srtlong.pdf

[dense cosmos]

hmm

-# add more bits tho

[real sierra]

JUST DOUBLE THE BITS

the ripple carry adders already hurt to watch

dont do this to me

[opaque sigil]

Replace them with carry-lookahead-adders ez, how bad could it be

[tender river]

if you implement cool stuff like division i'd love to hear you talk about it nothing like the feeling you perfectly understand something without putting in the effort because someone else talked about how they did it

[real sierra]

the bar for talking about this is division? it's that high?

[tender river]

no you can talk about much easier stuff too

[real sierra]

i feel like i might've set your expectations a little high

[tender river]

no this is very cool already

[real sierra]

i spent 4 hours trying to figure out why i could only do recursive calls 2 stack frames deep

[tender river]

i know how time consuming even simple stuff can be

[real sierra]

turns out i was trying to restore something after throwing away the stack frame it was in

and this just didnt manifest as an issue unless you went deep enough

huge props to this game tho

very cool

[tender river]

i remember some online learning tool that made you write various functions in verilog and checked their outputs

[real sierra]

the built-in assembler lets you define your own tokens as binary values

here ive used them to map out my opcodes

then there's the macros

[tender river]

i implemented basic logic elements and adders but didnt go further

[real sierra]

and now im putting together basic math functions in a little library

[tender river]

most of the bit tricks i know are from competitive programming not from any hardware related stuff

[real sierra]

good way to learn verilog

i was thinking about learning it or VHDL at some point

but it seems painful and theres no way i could ever use it

[tender river]

there are some cases where you can trivially solve a problem for every bit and then merge the results

[tender river]

when are you implementing a lisp

https://raw.githubusercontent.com/kspalaiologos/malbolge-lisp/master/MalbolgeLisp.pdf

[real sierra]

malbolge

this pdf is a crazy read

[tender river]

yeah i really like how something like this was possible

one thing is saying something is turing complete and another is making it run semi useful programs

[tender river]

but a forth is even easier to implement than a lisp

its almost purely stack based

[olive sable]

First attempt at touchscreen controls

Its seems to bedoing something, but not what i was trying to do

[amber fractal]

where fps?

I don't see many frames here

[olive sable]

in terms of fps, we have no fps

idk why tbh

mobile shit

i think i linked touchscreen to the movement, but its also seen as mouse input by pygame so thats also something

[real sierra]

watching the blinkenlights

[hoary lion]

i am witnessing advanced sorcery

[real sierra]

https://gyazo.com/ba010541a579295b76ff642077d6febd.mp4

s l o w

[olive sable]

what in the

[hoary lion]

so that's like

i dont know how much clock i have

[olive sable]

Tablet has enough power to do >10 fps so here's tablet

[hoary lion]

like million times slow than real time right

[real sierra]

1 instruction on this thing takes 120 ticks with the current clock settings and fetch cycle

so at the current tick rate thats uhh

[gritty dust]

Really nice man

[real sierra]

13 instructions per sec

[olive sable]

its tap based apparently, idk why

i guess il just plop some arrows on the screen and call it a day?

[real sierra]

dude this looks sweet

[olive sable]

thanks

[hoary lion]

so this is built with python and then wasm compiled right

[olive sable]

yep

[hoary lion]

🔥 🔥

[olive sable]

https://github.com/samvanmaele/game-engine

the link in the readme has the current version, should run fin as long as you're on desktop i think

[real sierra]

right shifting isnt working

catastrophic failure

it returns to a wrong address

who knows what kind of stack poisoning might be happening here tbh

16 stack frames * 10 bytes per frame

i might need to make the stack bigger

tomorrow prolbem

[olive sable]

[amber fractal]

I swear I remember a video on binary opts

Let me look through a playlist rq

[amber fractal]

https://youtu.be/rDDaEVcwIJM?si=w48j3r4FKMiGezO8 40:18
The entire video is a masterclass (I'm looking at you human method of square root) but this the division section.

This probably doesn't help with the current issue but it is fun

[tender river]

it's all coming together, the language ideas i had but didn't know what to do with are somewhat integrated in my head now

it's no longer the question of "what" but "how"

[amber fractal]

@olive sable Engine mobile test at 165hz xdx

[olive sable]

at least it seems to be running well

[amber fractal]

Besides movement being yes

it runs well

[olive sable]

the tablet gets 70fps with a snapdragon 870, but it doesn't get recognized as mobile so the shadow mapping is on there

what does your phone have?

[amber fractal]

let me check

https://m.gsmarena.com/asus_rog_phone_6-11647.php

snapdragon 8+ gen 1 (we love tech naming)

I'm gonna miss dual networking when I have to end up switching phones

[olive sable]

so 4/3 faster

[amber fractal]

I love having enough ram in this day and age

I knew the phone was cooking with video, but not that much

didn't know I was paying for a slo-mo camera

[olive sable]

my A52 has this

[amber fractal]

Surely you can get decent 8k photos but software isn't allowing it

[olive sable]

64MP is a lot more than 8K suposedly

so its probably just software

but i wouldnt actually use 8K from this phone's camera, cuz sensor is SMOL

[amber fractal]

The #1 reason I switched to RAW photos for high rez

but yean SMOL

[olive sable]

guys im trying to testrun the LLM stuff apollo sent, does this look correct?

[amber fractal]

it looks like it is running and training

[olive sable]

that does seem to be the case

i meant more like the model reporting as 0.4M params

is that correct for those settings? i wouldnt know cuz im just the guy with the gpu for this project

[fast pagoda]

i like that training log output in the terminal, what is that

from

and maybe it just means the amount of trained parameters?

like so far, or if it's a lora, the size of the lora params only

[olive sable]

idk

i think its supposed to be total?

it started at 0.4 and never changed

[fast pagoda]

how big are you expecting

[amber fractal]

I think this version is just smol

[olive sable]

im not expecting anything, since idk how this works

[amber fractal]

params add up to being 0.4B

[olive sable]

so its just a typo i guess?

[olive sable]

i think its from "rich"?

python lib apparently

[fast pagoda]

i mean it's definitely rich output

but like

what tool are you running that has that built in

[olive sable]

idk, not my code

[fast pagoda]

but yes that is definitely rich log style

im using that in my batch prompting thing

[amber fractal]

it probably isn't built in, apollo probably spent a lot of time doing it right

[fast pagoda]

https://tenor.com/view/snoop-dog-who-what-gif-14541222

i feel like i recognize or probably would recognize who "apollo" is in this context

[olive sable]

my pc is laggging every so often from this lmao

[olive sable]

apollo is the guy that made the program, and im the guy with the 3090

so im running his code cuz he doesnt have a 3090

[fast pagoda]

apollo from this disc? isnt there an apolloiscool

[olive sable]

that is the guy yes

aight im gonna let this run while i sleep

hopefully it wont go over 24GB and quadrupel the needed time

the checkpoints take up 12GB each?

im gonna run out of starage space in half an hour

12.5GB every 2 min is 6.25GB/min, or 1TB every 2.5hrs

how am i supposed to keep this running for 150 hours like this?

[amber fractal]

Deleting old checkpoints probably

[olive sable]

i have school and stuff, i cant babysit this the whole tim

[amber fractal]

Need a script for that if it doesn't already include one

If apollo names the checkpoints based off of what checkpoint it is, it should be possible to automate

[olive sable]

this

[unkempt citrus]

Anyone read AI 2027?

I skimmed it a while back and thought it was astupid

but I just watched a video on it and its even dumber than I remember

[dense cosmos]

whats that

[amber fractal]

I'll just wait 2 years for it to exist irl

[unkempt citrus]

Its a "prediction of the most likely AI outcome"

Where fictional company OpenMind makes an AI Agent that achieves AGI in 2027 then destorys the world after an arms race with China

in a stupidly long tale thats really just the paperclip machine

but don't worry, it can be stopped if you gets some more AI safety researchers

Written by some AI Safety researchers

No conflict of interest there

And these people definitely have qualifications yep

[olive sable]

[unkempt citrus]

https://ai-2027.com

[dense cosmos]

well, tbh it does make sense that the people most paranoid about ai would want to be in the positions to stop said ai

[unkempt citrus]

2nd section

[amber fractal]

There is no preventing AI from having bad thoughts, just like how you can't force thoughts into people

[unkempt citrus]

yes they'll be able to build 1000x data centre in 3 months time

[dense cosmos]

lulz

[amber fractal]

1000% this can only exist if we get very sparse, very fast

[unkempt citrus]

Also hive mind AIs

and two evil AIs Agent 5 from Open Mind and DeepCent merge together to take over the world

Cos AIs merging is a thing

[amber fractal]

Man people really gave up on new archs huh

last time I checked you can convert somewhat but not entirely merge models

[dense cosmos]

was wondering the other day, even if the capabilities of ai scaled with parameters and therefore their ability to self improve, resulting in exponential growth, it's still going to be stymied by the curse of dimensionality... right?

[amber fractal]

nvm I forget image models (as long as it is same base model) can do thus

[unkempt citrus]

With the amount of training data, double descent kicks in so its not that big of an issue

[amber fractal]

but my first point still stands

[unkempt citrus]

but data is more the limiting factor than hardware usually

[unkempt citrus]

Yeah but are they twice as powerful

[amber fractal]

I feel like eventually one will need to incorporate synthetic data as part of training larger models

I mean some methods already exist, but probably a lot more so

[amber fractal]

Ah yes the trenchcoat method

[olive sable]

how do i code it to delete the file? is there like a torch.delete()?

or is via os.remove() accepttable?

[unkempt citrus]

I think some already are, but I think they'd hit the same barrier as genetic algorithms did

[amber fractal]

I'd be doing os.remove()

[olive sable]

aight

old_checkpoint_name = f"discord_llm_step_{self.step - 200}.pt"
old_checkpoint_path = os.path.join(self.config["checkpoint_dir"], old_checkpoint_name)
os.remove(old_checkpoint_path)

code

[amber fractal]

Ik neurosynth has something like this, so it is being heavily used already if anyone can impl it.

Though I'm not 100% if it is synth data or just a method to use data better

[olive sable]

the code for removing the old checkpoint seems to work

im going to bed now

[amber fractal]

@prime ridge thought it'd be good to ping you

[amber fractal]

Gn sam

[hoary lion]

let's make a DSL based on JSX

[amber fractal]

I have no idea what this means

[hoary lion]

[molten pollen]

whats that

heard the name tho

[hoary lion]

domain specific language

[molten pollen]

what dataset is it trained on

[hoary lion]

am i contextually right

[amber fractal]

Something, he's clueless

[molten pollen]

oh is it a private one?

[amber fractal]

I can see if the owner opened training data

[molten pollen]

so what did u find

[amber fractal]

still looking, had to boot up laptop

[molten pollen]

anyways are like 10million twitch chat msgs enough to train a 124M gpt2 model

or i gotta go smaller on the model size

[amber fractal]

Seems like it is not open at the momemt

[molten pollen]

oh

i love ur profile pic btw

whats it based on

[unkempt citrus]

https://ai-2027.com/

[molten pollen]

what is it like a horror fiction story

tf is openbrain

[amber fractal]

Based on this meme

[molten pollen]

what did u censor?

[amber fractal]

I censored because I needed to censor

[molten pollen]

real

[amber fractal]

server is 13+ after all

[molten pollen]

https://arxiv.org/pdf/2502.15840 funny paper

[maiden geyser]

why do ai people always have to use emojis in everything they create

[hoary lion]

see

little do we know is that that figure is made with ai

[hoary lion]

isn't this the one where claude panicked and called the FBI

[molten pollen]

yea lmao

[molten pollen]

a lot of emojis mean the man who made it is either

1. AI
2. AI pretending to be human
3. a man with autism
4. a kid

[hoary lion]

please use uhh react icons

[maiden geyser]

5 is doesn't exist?

[unkempt citrus]

less AI-bros and more hustle bros

but theyve adopted AI a lot

[molten pollen]

i got distracted

[molten pollen]

hustle bros?

what they hustling

anyways did anyone implement transformers 2 or squared

whatever its called

[dusky jackal]

I see so much shit like this. Just wait until the people eating up these articles learn about Neuro-sama.

[ruby timber]

Good morning

[mighty bane]

Any of you nerds play Space Engineers? 😄

[maiden geyser]

morbing

[nocturne olive]

Make it delete old checkpoints like any sensible trainer that makes big checkpoints

[nocturne olive]

NeuroSynth's current datasets are entirely synthetically generated using Neuro RVC and existing compatible datasets

This is done because making a full vocal synthesizer dataset from scratch (which we will do for 1.0) takes a lot of time

[maiden geyser]

reinstalling windows was too easy this time wtf

i didn't even have to do regedit or oobe\bypassnro

why can't microsoft invent Microsoft Basic Display Adapter 2 or something

[burnt aurora]

Meanwhile :
👉 ⬇️

➡️ 💻 ✅

[burnt aurora]

It'd just be MBDA1 but with a new icon in device manager

And more telemetry

[maiden geyser]

they can at least upgrade the resolution to 1280x720/1920x1080, 800x600 isn't real anymore

[burnt aurora]

Too much work

The only real way is to ai generate the resolution and image and everything with Microsoft Copilot Display Adaptor

Who needs 3:1 ai frames if you can have all ai frames

[real sierra]

does anyone have advice on a good way to build a stack frame

I'm not 100% on what should really be put in it

right now what I'm using has the function args, return address and a pointer to the parent stack frame

feels like I'm missing something

I was mulling over stack canaries tbh

currently shit just fails catastrophically if I corrupt the stack accidentally

debug info might be smart too

but I'm not 100% on how to format that

idk where to put that info either

if you're unwinding the stack you're probably discarding stack frames as you go

so where to store the return addresses as you go

well I was thinking for a stack trace

it is yeah

yes I realize that

but if you want a stack trace from within nested function calls

your return address in the current stack frame only tells you the immediate caller

you would have to inspect the previous stack frame for the information on the next caller up

which seems non-trivial to me

yeah I don't quite have that

what I called a frame pointer earlier actually points to the first stored argument in the frame

but the information for how many arguments a frame has isn't stored

well I should probably get this fully working first before adding features

I have a function that recursively calls itself 16 times and it's causing issues

my debugging tools are hell

I get to stare directly at ram

it's not as bad as it seems, this is all in a circuit designing game so it coddles me a bit

but not enough

coddle me more

I may as well be

don't get me wrong, the only saving grace is that the instruction set is simple

and that stuff isn't compressed in memory or anything

nor are addresses randomized or stuff like that

the ALU is severely lacking though, as evidenced by my trying to implement a logical right shift in software

[mighty bane]

Also: A few days ago I uninstalled my AV.
Never ran better

[burnt aurora]

The best antivirus is to not be compatible with the virus

[maiden geyser]

it's like saying frogs are better than humans because they can't contract a human disease

[nocturne olive]

Did you know? There's a lot of Linux malware that's potentially significantly worse than Windows malware due to how widely Linux is used by servers for big companies

[prime ridge]

Yeah you can delete all the checkpoints. I switched model architectures so those old ones will be replaced with new ones. Also there should be an auto-delete functionality. These damn files are taking literal days to zip. I only expected it to take 30 ish hours to zip but it's just been going and going forever...

[prime ridge]

Compatible with targeted attacks, not compatible with generic attacks. What do you encounter more often

[nocturne olive]

The point is there's very much Linux malware

[prime ridge]

well yes ofc

[grave fractal]

I definitely recommend adding canaries ( it clears your mind about actually corrupting data. You'll focus on other things), and also checking for other issues can break recursive calling.

i’d love to dive into this with you tbh, but i’ve got an exam on caches coming up — L1, L2, L3, all that fun stuff @real sierra

[scarlet arch]

I love these kinds of videos. They're so adorable.
https://youtu.be/ta99S6Fh53c

I love that whole channel actually

[olive sable]

Goodmorning

I am alive

For now

[knotty current]

i regret ever legging chatgpt write numba code

damn that was not type safe

[loud thicket]

Both woke up rn

[knotty current]

no

i was schzoing over at the ARG thread for the whole day

[olive sable]

only did 2% while i was sleeping

this is gonna take a while huh

[real sierra]

cache exam

every exam question is a cache miss on my brain

[grave fractal]

i'm just ---->

[grave fractal]

wait , you are not using HDL ?

[rare bramble]

to me caches are just black magic, being able to massively boost performance just be reordering some operations is kinda nuts

[grave fractal]

i really liked/messed when they were "black magic" for me

[opaque sigil]

I like caches, they're not my problem to implement

[nocturne olive]

LLM pretraining be like

[grave fractal]

implement ? , nah ez , study for exam ---->

[olive sable]

my GPU voltage keeps dropping from 1.050 to 1.010 and then goes back up

i think my 750W PSU is struggling

[nocturne olive]

It has nothing to do with the PSU, that's just what the GPU voltage regulator is doing

The GPU core voltage is not coming from the PSU

[olive sable]

ive never actually seen my disk util go to 100% before

[real sierra]

nope

the processor is something I rigged together in this random circuit design game over the course of 2 days

16-bit

[grave fractal]

GPU VRM is . . . . dying

[grave fractal]

risc-v ? or custom ?

[real sierra]

i put it together because someone dared me, so I couldn't be asked to learn and implement a semi-reasonable architecture
instead I just adapted the nandgame architecture

[grave fractal]

oh boy

[real sierra]

16-bit system with two general-use registers, a program counter register and uhhh

that's about it

separate ram and rom

[grave fractal]

good start

[real sierra]

two registers definitely complicates even simple tasks

[grave fractal]

i made a 32bit core in risc-v (in hdl {verilog} tho)

[real sierra]

impressive

verilog and vhdl scare me

[olive sable]

try not to flex challange

[real sierra]

seems tough to learn

[grave fractal]

just stay away from floating point alus , you will be fine

[real sierra]

oh don't worry my ALU can't do floating point

[grave fractal]

[real sierra]

it's missing some really vital operations

like a logical right shift

turns out this is agony to do in software

[olive sable]

then do it in hardware

make a pcb

[grave fractal]

soooo

no division

[real sierra]

division will probably just get binary searched

[grave fractal]

[real sierra]

wait can I even do that

[rare bramble]

how can you binary search without a division

[real sierra]

[grave fractal]

good point

[real sierra]

I guess we're doing repeat subtraction then

even stuff like this has been extremely challenging though

[grave fractal]

12341 / 2 , alu pov : 🦥

[real sierra]

both registers are needed for most arithmetic so there's nowhere to really store stuff like loop counters

[olive sable]

just make more registers

[real sierra]

as such I've restored to recursion for even stuff like multiplication

[burnt aurora]

Ofc there is, but finding it is definitely more difficult

[real sierra]

no storage? use a stack frame

[grave fractal]

the same stack frame that can overflow ?

[real sierra]

reeeeaaallly big stack

make sure you put the smaller arg second

[nocturne olive]

Silly

[rare bramble]

for division, could you do something like:

6/2:
2x1 >= 6: false
2x2 >= 6: false
2x3 >=6: true

6/2 = 3?

discord formatting

[olive sable]

that works with certain ints i gues

[real sierra]

does such an algorithm work with two's complement negative values

[olive sable]

but if you try to do 1/2 you're fucked

[rare bramble]

ye but if it cant do floating point, it should be fine right?

like 1/2 would be 0 (or 1?)

[grave fractal]

7 ÷ 2:
2×1 = 2 → false
2×2 = 4 → false
2×3 = 6 → false
2×4 = 8 → true

[olive sable]

probably yes

[real sierra]

I was about to say "floating point would get a software implementation" but if bit shifts are already software that might be intractable

[burnt aurora]

Also, genericism
Windows only has 10 and 11, and most win10 stuff works on 11, so malware can target pretty much all consumers and businesses at the same time (win7 can be ignored atp)
Linux has so many different base distros (ubuntu debian arch gentoo etc), which themselves get made into even more distros (ubuntu alone has too many to remember them all) with so many ways to customise everything that writing generic malware is much more difficult, so i'd assume that most of it is targeted at distros that get used by money makers, like server distros

[grave fractal]

for A / B , it should be B × Q ≤ A

[real sierra]

you need to start your trial multiplication with N×0 to catch the case where a/b with a<b

but also uhh

there's no hardware multiplier

[nocturne olive]

Maybe

[grave fractal]

[burnt aurora]

But yes, the linux malware can go much deeper

Which is why you make backups and partition home away from the rest, reinstall, and plug in your backup

[stray dragon]

whatever, go my windows defender

[burnt aurora]

Ngl if i was actually someone who gets targeted by malware I'd have shit like bulk reinstall scripts
Format the whole drive, reinstall linux from scratch, and run bulk-install-programs.sh

[real sierra]

proprietary software user

[burnt aurora]

🍷

[real sierra]

yeah, my software comes from a company with hundreds of paid professionals maintaining it

[burnt aurora]

My software runs at more than 10fps tho

[nocturne olive]

Very silliness

[burnt aurora]

It's incredible how shitty the fusion 360 ui performs
Especially because the 2d ui somehow runs slower than the 3d cad view

On windows

[nocturne olive]

That seems a little crazy

[burnt aurora]

(On linux it's all kinda slow because wine isn't great for performance)

But that's why I'm learning freeCad

[olive sable]

what about proton? is that any better?

[opaque sigil]

proton is just wine isn't it

[olive sable]

idk

i dont use linux

[opaque sigil]

well, with some extra stuff on top

[uneven siren]

its better wine

[burnt aurora]

Proton is wine for games

And it runs great

Except if you use the wrong nvidia drivers

[stray dragon]

classic nvidia drivers

[burnt aurora]

I forgot which one it was but it made every polygon on screen take away a frame

[olive sable]

nvideoiae

[burnt aurora]

Satisfactory died completely

[uneven siren]

TBH if it wasn't for games I play needing kernel level anti cheat I would have ditched windows for it.

[nocturne olive]

60 polygons = 0 frames?

[olive sable]

even on windows satisfactory has graphical glitches due to the new nvidia drivers

[burnt aurora]

It's amazing, they even broke on windows long before i switched to windows
We almost mailed the pc in under warranty because it looked like the gpu was deadge

[burnt aurora]

Pretty much

[nocturne olive]

Windows 11?

[burnt aurora]

I think it may have been 570 but I'm unsure

[stray dragon]

yeah i was struggling with driver issues for months after getting a 4060 ti 16gb

came with graphical bugs day one

right out of the box

[burnt aurora]

3070QM here

[olive sable]

ye, but the drivers should be the same ones as for 10 so

[burnt aurora]

The top 3cm of screen were rendered above the bottom 3cm in windows, in a streaky almost transparent pattern

[nocturne olive]

Well, who knows
I've heard of Windows 11 24H2 breaking stuff that worked before

[burnt aurora]

Only win11 tho

23h2

[olive sable]

it wasnt windows update that was the problem

[burnt aurora]

Nah that's on nvidia completely

[olive sable]

happened only after updating nvode drivers

[real sierra]

could I get someone to verify the code for my right bit shift

I wanna believe it's correct but

it's crashing every time

[olive sable]

[stray dragon]

after trying to fix an issue with the gpu that was causing stutters as i left the pc on, and reinstalling the drivers at least 6 times, i gave up and switched the card out for the RX 6600 XT i had sitting around

[real sierra]

I'll send it when I get home lol

[nocturne olive]

That's crazy

[olive sable]

i will verify, but i will have no clue what im looking at actually

[burnt aurora]

Why do chips not have modulo in their ALU if it's doable via hardwiring input to output

[olive sable]

[real sierra]

its not too complicated, mostly just stack macros

[burnt aurora]

Lul

If only i could

But sadly it's a laptop, dad denied my pleads to just get a powerful desktop and mediocre laptop

And now the only thing that required a laptop is done by his phone

Idiot

[real sierra]

found a video I posted from yesterday #programming message

[burnt aurora]

Is that a chip

Emulator

[olive sable]

i have to read a book and make a project about it, and i have to make a motivation letter by tuesday.
so imma do that now
bye

[nocturne olive]

Self-triple-scammed

[real sierra]

it's a little niceified, it abstracts actual electricity away so you can just work with logic gates and high/low logic level

but yes

[grave fractal]

sure

i'll be studying tho , so hit me up dm or mention me

[real sierra]

shifts

[opaque sigil]

replace shifts with subtraction and it works ig

[real sierra]

not in hardware

oh wonderful

this architecture actually has a name

https://en.wikipedia.org/wiki/Hack_computer

[olive sable]

chatpgt is fine to know my home adress right?

they surely wont store it

[real sierra]

16

@olive sable @grave fractal i'm home now, here's the code i'd like reviewed:

# math_shr(x) returns x right-shifted by one position.
function math_shr 1;
    push_arg 0;
    push_value 1;
    push_value 2;
    call math_shr_recursive;
    push_retval;
    return;

# internal backing call for math_shr(x)
# math_shr(v, mask_out, mask_in) performs a right shift by copying bits using masks.
function math_shr_recursive 3;
    push_arg 2;
    pop_d;
    if_else_d JGE math_shr_recursive_if math_shr_recursive_else
    @ math_shr_recursive_if
        push_arg 0;
        push_arg 1;
        push_arg 1;
        add;
        push_arg 2;
        push_arg 2;
        add;
        call math_shr_recursive;
        push_retval;
        goto math_shr_recursive_end;
    @ math_shr_recursive_else
        push_value 0;
        goto math_shr_recursive_end;
    @ math_shr_recursive_end
    push_arg 0;
    push_arg 2;
    and;
    pop_d;
    if_else_d JNE math_shr_recursive_shift math_shr_recursive_ret
    @ math_shr_recursive_shift
        push_arg 1;
        or;
    @ math_shr_recursive_ret
    return;

not sure if there's better highlighting i can provide

[olive sable]

so the thing is, i was joking about verifying, i only work in python really

[real sierra]

oh

[gritty dust]

Based

[olive sable]

[real sierra]

with division? no not really

easily calculated if you can do division

i always get these mixed up

my LSB is on the right

does that help

[rough bloom]

big endian is the one that makes sense and has the correct byte order if you just read it as a sequence of bits
little endian is the one that's fucked up

if the b in lsb stands for byte here, then yeah, lsb on the right is big endian

[real sierra]

then im BE

i think my one complaint about this game's assembler is

you can't nest macros

so i have to expand any macros i use in writing my macros down to their assembly

this leads to wonderful things like

# call <functionName>
macro call {LOAD ARGS; LD R; LOAD SP; LA R; LR D; LOAD SP; LR INC_R; pointer rv inline 0; LOAD rv; LD A; LOAD 12; unpoint rv; LD ADD_DA; LOAD SP; LA R; LR D; LOAD SP; LR INC_R; LOAD a?; JMP; LOAD SP; LR DEC_R; LA R; LD R; LOAD ARGS; LR D; LOAD SP_SWAP; LD R; LOAD SP; LR D}

[scarlet arch]

So based. To this day I still don't get why little endian is used most of the time. Network order is also big endian and it just makes so much more sense

[tender river]

also sometimes stack traces can get in the way of performance, sometimes jmp function does the trick better than call function

[real sierra]

this strikes me as a very radical take

you seem like the person who actually uses goto in java

[rough bloom]

something something it's better for CPU design because the lower bits are first and they're necessary for the higher bits to be useful in calculations
apparently it's better for casting integers to smaller types too because then the pointer doesn't change
idk I think it's largely vestigial now

[tender river]

floats are useless anyway

[real sierra]

no trust me my 16-bit floats are very precise

[opaque sigil]

iirc big endian for networking was chosen after intel had adopted little endian so i blame them, could've just gone with what intel was doing but nooooo, we can't have nice things

[grave fractal]

having it in "pseudo" made me , but it seems fine , does it work ?

[scarlet arch]

Yeah it makes no difference what the internal representation is if the CPU can just read them differently. I don't get it. I mean I understand why it was done once. But you'd think we'd find better solutions.

[tender river]

i dont use java but actually this approach is called while() in java so you probably use it

[real sierra]

oh im impressed, i was waiting for you to show up so i could explain what the stack macros meant

but you seem to have figured it out on your own

[tender river]

little endian is nice because a larger integer is just an extension of a smaller integer

[grave fractal]

i'm not that bad

still tho

[tender river]

truncation is trivial, no address mangling necessary

[real sierra]

so it does look like its working?

or at least, shouldn't be corrupting the stack

[tender river]

i do use big endian in networking

[real sierra]

honestly i was kinda hoping that wasnt the case, because it means there's a bug with my stack macros instead

while
no thanks

make it recursive and ill think about it

[grave fractal]

the wrapper is fine (math_shr ) , maybe the (math_shr_recursive) ,, it may act up

[real sierra]

[scarlet arch]

Okay and whatever that is (I'm not that deep into the CPU rabbit hole) can't be solved in a way to keep using big endian and have it be efficient?

[tender river]

getting the least significant bits is a common operation, getting the most significant bits isnt

so it makes sense to optimize for the former

[grave fractal]

man i didnt touch asmb like for 2 years

[tender river]

bigints are generally stored in little endian too i think

[grave fractal]

so rusty

[scarlet arch]

Why are they called least significant if they're so significant

[tender river]

maths

[scarlet arch]

I'm mostly joking at this point haha

Not like I can have any valuable opinion on this

[real sierra]

add, or, and pop 2 values off the stack, perform the relevant computation and push the result back onto the stack
push_arg <n> copies the n-th function argument and pushes it onto the stack
push_value <n> puts the literal value n on the stack
push_retval pushes the result of the last function call onto the stack
pop_d pops the topmost stack item into the D register
if_else_d <condition> <true_branch> <false_branch> checks the value in the D register against the jump condition and branches accordingly

my math library depends on you konii

[grave fractal]

ok i wasnt too far

i cheated tho

took a peak on the wiki you sent

[real sierra]

[grave fractal]

https://en.wikipedia.org/wiki/Hack_computer#/media/File:Hack_Computer_CPU_Block_Diagram.png

[real sierra]

didnt even know the wiki existed until today

much better reference than what ive been using

[grave fractal]

so yah it looks fine , from 30 second review

buut , as i said , i'm rusty

[tender river]

there should be a game about quantum ISA design

[maiden geyser]

if_else_d <condition> <true_branch> <false_branch>
why not do it like in x86

[grave fractal]

wait

you mean something like :
CMP EAX, EBX
JNE label_true
JMP label_false

[tender river]

whats the difference between push_arg and push_value, are those different stacks?

[maiden geyser]

but without caps, you're not writing SQL here

[tender river]

its a macro

[real sierra]

when a function is called, the args are stored in the stack frame
push_arg takes an index and pushes the arg at that index onto the stack for actual use

[grave fractal]

lol i just googled it , and copy paste it , dont expect form me to write it

[real sierra]

push_value accepts a literal and pushes that onto the stack

[tender river]

gotcha, would push_arg break if push_value was called before it?

[real sierra]

shouldnt do, there's a fixed address in ram that stores a pointer to the current function's first argument

so you can use push_arg at any time

and as many times as you'd like

[tender river]

and its pushed when calling a function?

or does it only work before the first call

[real sierra]

yes, before the function body is entered the value of the new args pointer is set

stack frames look like this atm

stack grows toward the bottom of the screen

when a function is called, the return address is pushed and the function is called to

then the current value of the ARGS global is pushed onto the stack so it can be restored later, and the new one is calculated and stored in the global

when returning from a function, the topmost stack value is popped and stored in a RETVAL global, then the current ARGS global is stored in a SP_SWAP global as it's used to restore the stack pointer, then the prev args ptr is popped and stored in ARGS, and then finally the return address is popped and jumped to

[tender river]

and what does if_else_d compare? (other than d register)

[real sierra]

in practice you feed it a jump condition, so it compares to zero

now i just need to port it to my language

actually

can i get a math check

the yellow args here

for each frame, the first one should stay the same, and the last two should double

[olive sable]

Realtek's equalizer sucks

[real sierra]

but i feel like 0x0600 is not twice 0x0100

is my math bad on that or

ALU?

[grave fractal]

what is that again

i was away

[real sierra]

it's the stack, with stack frames highlighted

[tender river]

so is it something like this?

math_shr_recursive(a, b, c) {
    if args[2] >= 0 {
        a = args[0]
        b = add(args[1], args[1])
        c = add(args[2], args[2])
        ret = math_shr_recursive(a, b, c)
    } else {
        ret = 0
    }
    v = and(args[0], args[2])
    if v != 0 {
        ret |= args[1]
    }
    return ret
}

[grave fractal]

yah there some "math" , not doing "math" stuff

[real sierra]

[tender river]

also buffered reader is important for reading until newline, good luck

[real sierra]

adding something with itself is a left shift, and once the bit reaches the highest position the number is interpreted as negative, which is why that first condition is args[2] >= 0

[tender river]

i dont see bugs major enough that they would mangle stack frames fwiw

[real sierra]

nor do i

which is why im worried

if the function returned the wrong result, that'd be fine

but it shouldn't be crashing horribly

if the ALU is having issues tho, that could explain it

probably the ripple carry adders again

ill slow the clock and see if that helps

[tender river]

hardware debugging, like software debugging but you have even less idea what's going on

[real sierra]

at least it's consistently wrong

[tender river]

not if you have multiple cores!

[olive sable]

but i made it myself

i have become the AI

[maiden geyser]

why do you even care anymore

[olive sable]

cuz i get graded on this shit

i cant do another year man

[real sierra]

felt that