programming

[rigid snow]

is it really that much better

[glass flower]

its fun for the first time you install linux but... i don't need my vscode terminal to show it cut off

[sick owl]

Yes, it honestly feels like a bigger upgrade for gaming than my 4090 was

[obsidian mantle]

i was shocked when i bought 240 hz monitor i thought difference between 60 and 140 hz is not visible

not when i look at my old 60hz second monitor it feels like its in 20 fps

[solid bough]

It's time

That I do the unthinkable

[sick owl]

Mine is 240 but I have it locked at 120 cause I cba with DSC

My old monitor was 360hz but I found that to be completely overkill

[clear sedge]

Update Readme.md

[obsidian mantle]

yeah 144vs240 is also much less visible than 60vs144

[solid bough]

I un-code the code, I unsort the sorted, Even worse: I scramble the unscrambled

Okay, my messages might have been a bit unhinged at the start...

[clear sedge]

even 60 vs 100 is very noticeable

[sick owl]

The improvements are non linear

The higher the numbers climb the smaller the difference

50 vs 60 is noticeable, 90 vs 100 you'd really have to squint

[clear sedge]

the inverse of decibels

[solid bough]

What the filtered filtered

[obsidian mantle]

trained

[sick owl]

Also worth noting that panel tech plays a role

[solid bough]

It now goes down though, but that rise

[sick owl]

120hz on OLED will beat many 240hz LCDs in input latency and motion clarity

[solid bough]

That is an unhealthy learning rate I guess

[obsidian mantle]

what do "test" and "train" actually resemble

im not sure i get it right

[solid bough]

train_RMSE and test_RMSE.

[obsidian mantle]

yeah but where do train and test come from

[sick owl]

Albeit lower fluidity

[maiden geyser]

a victim of the burnt pixel

[solid bough]

Basicall: py train_loss = model.evaluate(x_train_mini, y_train_mini, batch_size=8, verbose=0) train_rmse = np.sqrt(train_loss) print("Train RMSE:", train_rmse) test_loss = model.evaluate(x_test, y_test, batch_size=batch_size, verbose=0) test_rmse = np.sqrt(test_loss) print("Test RMSE:", test_rmse) train_score.append(train_rmse) test_score.append(test_rmse)

[solid bough]

Oh, I was gonna send the part that gives the values

[obsidian mantle]

i comprehend nothing

you are training your model, so i guess it goes to "train"

but what is test

[solid bough]

We are in the same boat, I have no filtered idea what that means

I could just DM the maker

no really

[obsidian mantle]

so it got it's parameters adjusted, then it tries to work and we have this rmse value

here we have train rmse after each "training" session

but who is test

[solid bough]

It is basically doing a loop for X epoches: ```py
for i in range(num_epochs):
print(f"Epoch {i} of {num_epochs}")
log_info(f"Epoch {i} of {num_epochs}", logg)
print_vram_usage()
print(x_train.shape, x_train.dtype)
print(y_train.shape, y_train.dtype)
model.fit(x_train, y_train, batch_size=batch_size, epochs=1)

# Evaluate using mini subset for VRAM-friendly training metrics
train_loss = model.evaluate(x_train_mini, y_train_mini, batch_size=8, verbose=0)
train_rmse = np.sqrt(train_loss)
print("Train RMSE:", train_rmse)
test_loss = model.evaluate(x_test, y_test, batch_size=batch_size, verbose=0)
test_rmse = np.sqrt(test_loss)
print("Test RMSE:", test_rmse)
train_score.append(train_rmse)
test_score.append(test_rmse)
model.save('Model.h5')
print("Model saved.")
plotscrores(train_score, test_score, 'Scores.png', True)
gc.collect()```

filtered

[obsidian mantle]

batch_size=batch_size mhm

[solid bough]

Which is rn 20

Just for kicks n giggles I wanna put one of the Sketches as a pre-selected sketch when the code starts

[obsidian mantle]

what "batch_size=batch_size" even does when you pass it as an argument

[solid bough]

On "evaluate" I have no frickin idea

I'm mostly just doing trial and error and that train session was not even the right data

[obsidian mantle]

[solid bough]

Just for kicks n giggles I tried... Uh... Images that I can't show here and it failed hard

before going to sleep

[rough bloom]

sets the parameter batch_size to the value of the batch_size variable
if you don't specify it'll use the default value of the function parameter instead

[obsidian mantle]

but cant you just send it "batch_size"

[rough bloom]

no, it would have to be a positional parameter then

[solid bough]

Idk, it was like that in the original Python 2 Code

[rough bloom]

Python has both positional and named/keyword parameters for functions

[obsidian mantle]

so function doesnt have 4 arguments

its like foo(a,b)

[solid bough]

I might just sift through another 3K images

[obsidian mantle]

but we call it with foo(a,b,batch_size=batch_size)

or what

[solid bough]

And I tried as good as possible to have good amount of variation but also good data

[obsidian mantle]

or it has million of parameters and we dont need to specify each and here we tell it which is which

so in batch_size=batch_size on the left we have funcion's batch_size and on the right our batch_size we have here, not the function's internal one

[solid bough]

I'm real close to the VRAM max I set in code VRAM: 25.2 MB / 6095.4 MB | (2016, 5, 192, 144) float16 (2016, 3, 192, 144) float16 57/101 [===============>..............] - ETA: 4s - loss: 0.0316

[rough bloom]

maybe I explained badly
function parameters have names and sometimes also default values
if you don't specify the names in the function-call it is assumed that you are passing the parameters in order
but you can also specify them by name (like with the batch_size thing)

[obsidian mantle]

what the fuck is this

something is broken in my bot

quantum chess

[solid bough]

I might need to be more strict on what faces I use

oh my god

And that are the sketches

I might as well spin up Francium again

shuni.ex wanna know what trips me up real good?

[rough bloom]

[obsidian mantle]

so now it just finds transposition of this position from previous search and returns it

[solid bough]

That line of the code does NOTHINGpy dots_per_images = 60 #default is 60

It's used nowhere

[rough bloom]

yes, I know
I already told you so like a week ago

[sage crag]

ye

for osu 120hz oled is better than 240hz lcd

[solid bough]

Oh, That went over my head

[sage crag]

because less ghosting and better contrast

[rough bloom]

[solid bough]

dw, Francium is gonna keep me in check

Because since there is always a fluctuation of vowels which sort of forces me to clear the context I have in my head enough to not make me go wild.

In other words, I don't linger around with that thought that I need a long while on the sorting

Well, loads of images are getting yeeted if ANY amount of Background edges are detected (Dieter Bolen Get's a pass)

or if the face is not clear

[midnight sigil]

think of function(a, b, c, d), where a, b, c, d are all optional parameters, and you want to pass a, b and d only. You call function(a, b, d), but the interpreter would think you're calling function(a, b, c) instead, so using function(a, b, d=d) specify d(the argument passed) should be positioned at d(function argument intake)

[solid bough]

Caibi am I insane that I listen to Francium while sifting through the roughly sorted images again to do fine sorting?

[obsidian mantle]

okok since they are optional it all makes sense

[midnight sigil]

miku dayo

[rough bloom]

except they don't have to be all optional
you can do this with any combination of parameters AFAIK

[midnight sigil]

yuh

[solid bough]

Well in a weird way Francium (Neuro cover) keeps me more sane than usual when looking closer if an image is suitable for training

[trim valve]

i personally make all my functions in python like this ```python
def my_func(/, a, b, c):
...

to make sure nobody uses those pesky keyword arguments

[solid bough]

you mean kwarg?

I loosely use them in my Text Covnerter

[midnight sigil]

you should put it in the back no?

[trim valve]

uuuh

[midnight sigil]

/ works for arguments before the keyword

[trim valve]

am i misremembering

[midnight sigil]

yep I'm right

[solid bough]

Jesus christ that was a jumpscare for me

The base image looks normal though

[manic jacinth]

Looks straight out of an analog horror

[solid bough]

Yeah, and that is the base image:

like wtf

[solid bough]

And that is the Edges image

[stray dragon]

also why is it still smoothed

matplotlib smh

[solid bough]

Idfk, every time at the start it is a simple P2P connection and then turns to that

[trim valve]

I can never take the phrase "analog horror" seriously anymore lol

[solid bough]

But that image got me good ngl

[midnight sigil]

that's over-fitting

[solid bough]

How the frick can it overfit within the first like 5 epochs???

[midnight sigil]

[solid bough]

or like 10

sure I used only 630 with flipped version (aka doubling that then) but damn

and that with like 4 samples of that image

[midnight sigil]

[solid bough]

Well, it's then time to maybe even sift through 10K images

RIP my evening

[sick owl]

This screen fills me with irrational hatred

He even has a fedora on

The bastard

[solid bough]

A white hat pen tester I guess

[proud ivy]

God. Maybe it is overfitting.

[midnight sigil]

basically if you see your test and train loss not decreasing in the same kinda curve, it's being wrong

[proud ivy]

Somewhat scary.

[solid bough]

Yeah, that besides the face is not really what I was aiming for, and now I re-filter it again:

But Francium (Neuro cover) keeps me somewhat in check

And since I have quite a lot of images like that, I loose a ton of images then due to that

That is my mood rn

[solid bough]

Well, I'm done carefully checking the images, and the VAST majority were deleted (Now I only have 163 images left)

of the 630

But finally that kind of curve stopped happening:

[solid bough]

Oh yeah right, I frogot to say that the code uses matplotlib

[stray dragon]

i mean, it has a pretty distinct look

[stray dragon]

oh so it's NOT smoothing the line at all

[solid bough]

https://tenor.com/view/correct-power-chainsaw-man-csm-agree-gif-2034489833296620644

[stray dragon]

still seems very weird that it doesn't have any noisiness on the plots here though.

[solid bough]

And that was with learning rate 0.008

Like pretty agressive I guess

[stray dragon]

i guess old machine learning thing, so old optimizer

[solid bough]

Well, for kicks n giggles I also wanna try to Re-train a model

All that potentially could happen is that it goes horribly wrong

[stray dragon]

that'll lead to catastrophic forgetting probably

not sure what happens past that though

[solid bough]

That's what I'm doing rn

Just trial and error pretty much

And the line turned to a saw like shape initially

[stray dragon]

now that's what i expect!!

[midnight sigil]

maybe there're no randomness introduced in weight eval

[solid bough]

Basically what I did was putting the "Model.h5" as "pre_model.h5"

[stray dragon]

perhaps

[solid bough]

Idk why I changed it... OH JESUS

[midnight sigil]

[solid bough]

That is a huge spike

[nocturne olive]

Machine learning moment

[stray dragon]

it happens

very common tbh

[solid bough]

True, I rarely do ML stuff, I just recently dedicated liquifying my brain via sifting through Thousands of images to end up at 162 images

[midnight sigil]

you should make the loss plotted on a logarithmic scale, I think it would be better to visualize

can't read linear

[solid bough]

Either the faces weren't clear enough, there was stuff in the background, Or the quality like is like an Atari

[nocturne olive]

There was a NeuroSynth run that had a graph like this

[solid bough]

Wait, so that mans that is good?

[nocturne olive]

No that's just a normal thing

[stray dragon]

noise is expected in the curve

[solid bough]

Honestly, I have no idea how to do it at a log scale

Well, I just yeeted the pre-trained model through the learning process again

Like I train it the 2nd time now

[stray dragon]

stochastic gradient descent is an algorithm that adds random offsets to the weights to test how the performance of the model changes in response, and then it adjusts the parameters based on that data

since the changes are random, you should expect that at least some cause a lot more loss

[solid bough]

Time to check the model mids training

Holy, finally some sharp ish edges

[stray dragon]

i've seen some papers where instead of randomly initializing the weights of a large model and training from scratch, they train a smaller version of the model (with randomly initialized weights) to a certain point, and then upscale the weights to full-size somehow and train the full-sized model to completion

[midnight sigil]

just plt.yscale("log")

[stray dragon]

speeds up training process but final result was a bit worse than the model trained from full-size at the start

[solid bough]

Okay

[midnight sigil]

replace plt with your plot graph obviously

[solid bough]

Oh, I appreciate your help

Btw, is that supposed to happen every loop?

[midnight sigil]

you can also use early stops in pytorch to fasten and stabilize the training progress

[obsidian mantle]

???

[stray dragon]

i see

[obsidian mantle]

is it for safety

[stray dragon]

no, you forgot a )

[obsidian mantle]

no i didnt

[midnight sigil]

std::max ?

[obsidian mantle]

wait

[stray dragon]

max() terminates properly, but then evaluate_moves( isn't ended

[midnight sigil]

[stray dragon]

wait no

[obsidian mantle]

next line

[stray dragon]

next line

[obsidian mantle]

right right i didnt notice myself too

[solid bough]

That is now the scale, but idk if it chaned the form...

[obsidian mantle]

so max doesnt want int and uint64 because if int is negative it will give weird results

they are always positive i want tokens

[midnight sigil]

max works on all data types which support <

this shouldn't be an issue

did you specify namespace std

[obsidian mantle]

this compiles

no im using namespace std lets try

nuh

it was std

i can specify std:: even if i still have "using namespace std" right?

this compiles it really doesnt want to compare uint64 and int

[midnight sigil]

[obsidian mantle]

come on g++ i know what these numbers are dont limit me

anyway, i just thought, this is a bad solution.. i need it to be max (time, 700) and its too much

[solid bough]

Imma restart the code up to 5 times on train.py rq

[obsidian mantle]

it has to be depth limit

[rigid timber]

but what if you don't

[midnight sigil]

clang would give this a warning at compile time

[solid bough]

caibi should I have at least 500 images?

[midnight sigil]

I guess max can't do compare with unsigned and signed

[midnight sigil]

yes

[solid bough]

oh god

I could try to increase the samples per images

[obsidian mantle]

the int will never be negative, i meant to give him constant value in miliseconds to limit the bottom

the int which could cause problems while being negative, will never be negative in this specific situation since im constanting it

max should check what arguments its getting smh

[rigid timber]

well it's C++, what do you expect

[solid bough]

Hi toast.🔥

[opaque sigil]

sup

[solid bough]

I'm perfectly sane

[olive sable]

Surely

[solid bough]

Toootally

[midnight sigil]

why don't you make it unsigned int in the first place then

[solid bough]

Well, I can look what I can do

[obsidian mantle]

i want to type just 500 for example

but it defaults to int

[midnight sigil]

if you're on O2 and that variable is a constant, I think the compiler would recognize and hardcode that constant

[solid bough]

[obsidian mantle]

im doing this

[solid bough]

It's after the second run still like someone that lost their glasses

[obsidian mantle]

should i do g++ -O2

[midnight sigil]

do 0u

500u

[obsidian mantle]

no way that is 1 token right?

[midnight sigil]

so on

[midnight sigil]

well casting it takes even more

also use clang smh

faster binary

my python can only do 5000 nodes per second

I'm thinking of moving to C++ or do CPython magic

[solid bough]

Remember that?

[opaque sigil]

[solid bough]

I have a lower loss with 163 well sifted images

[opaque sigil]

nice

[solid bough]

as said, the VAST majority of the 630 images had bs in the background

[olive mango]

man:
WA test 1
MLE
WA test 23
kill me 😭

[solid bough]

How many images did you have to sift through?

[olive mango]

no, im doing algorithms

havent got the skills to code like yall

im only grade 9

[solid bough]

Lucky, I was sifting through a total of 5K images, and ONLY 163 were left

[solid bough]

And I based it of already existing code btw

[olive mango]

i mean in competitive programming

[olive mango]

i thought you wrote every single letter

[obsidian mantle]

clang

[solid bough]

Oh, nope... It was based of Deep Doodle

[obsidian mantle]

or is it o2

[olive mango]

wait, why not do in python

[solid bough]

aka: I downloaded the code and then made it run in Python 3

[olive mango]

or for a tournament, just bring the fish

[opaque sigil]

-O2

[solid bough]

Well, Python is not for everything...

[olive mango]

guys when can i upload image

[midnight sigil]

1. Token limit
2. No I/O access

[olive mango]

i mean, aint it better for AI?

[solid bough]

Idk

[obsidian mantle]

clang doesnt want dynamic range arrays

[olive mango]

uh, i mean online tournament

[opaque sigil]

iirc that's a gcc extension and not actually part of the standard

[obsidian mantle]

WHY IS IT NOT

[olive mango]

aintnoway

[obsidian mantle]

i know its dangerous

i know its absolutely safe in my implementation

[opaque sigil]

You'll have to declare it then memset

[midnight sigil]

oh no not the heap

🥀

[solid bough]

Welp, Imma be answering late for a long long time, Because Imma just not look if the 10K images look to the camera and do it straight on filtering on the edges filtering

[obsidian mantle]

??? does it allocate memory for array before the function who creates the array is called

what if i call it 5 billion times

[opaque sigil]

Nobody said anything about the heap

[obsidian mantle]

does it allocate 5 billion arrays

no it doesnt

[sage crag]

wrrrr

wrr

wrrrrr

wrr

wrr

[obsidian mantle]

its just stupid or im not seeing something

[midnight sigil]

[sage crag]

[obsidian mantle]

konii hi explain why i cant use variable to initialize array

[sage crag]

hi show me

[solid bough]

7Zip is choking on the big 7z file

[opaque sigil]

You can declare but not initialise

Go ask the committee why

[sage crag]

do you mean like:

int a[] = {1, 2, 3};
int b[] = a;

?

[opaque sigil]

VLA

[sage crag]

its because arrays can only have static initialisers

[obsidian mantle]

len_submoves does not change in this scope

[sage crag]

doesnt matter

[obsidian mantle]

its literally constant in this scope

[sage crag]

any array can only have static initialiser

doesnt matter

[obsidian mantle]

constants are invented by pussies

cosmetic shit

[sage crag]

a = b is a nontrivial operation for arrays

it requires a memcopy and an allocation

[obsidian mantle]

i can write into any constant anyway whats the point

to make me waste tokens

[solid bough]

I'm doing more filtering now

[sage crag]

its only trivial for things that fit in a single register

[olive mango]

chatgpt syntax( jk)

guys

brainrot words arent allowed right?

[sage crag]

yes

[olive mango]

+100000 social credits

[obsidian mantle]

so it allocates array of the size i wanted

then fills it with 0

[opaque sigil]

standard says no

[sage crag]

yes

well

it does

[opaque sigil]

[obsidian mantle]

whats the problem then

[sage crag]

depends where you declare the array

[opaque sigil]

Must initialise manually

[midnight sigil]

"Because that's dangerous tehe 😉"

[sage crag]

int erm[];

int main() {
  return 0;
}

erm here is a global and zeroed

[obsidian mantle]

so this is not allowed but 1/0 is allowed

[solid bough]

nvm: Traceback (most recent call last): File "e:/temp/datagen.py", line 88, in <module> main_datamaker() #for investigating the sketches File "e:/temp/datagen.py", line 33, in main_datamaker x_data = np.empty((num_samples, num_channels, image_h, image_w), dtype=np.uint8) numpy.core._exceptions.MemoryError: Unable to allocate 256. GiB for an array with shape (3313464, 3, 192, 144) and data type uint8

[sage crag]

int main() {
  int erm[];
  return 0;
}

erm here is not a global and therefore is uninitialised stack memory

[obsidian mantle]

its not global

i tell it how much of a size i want

len submoves lives the same life as the array itself

[sage crag]

having a size doesnt change whether its global or not

[sage crag]

only a const can be used as a length

lifetimes dont exist in c

[obsidian mantle]

its deallocated when function exits is it not

[sage crag]

doesnt matter

only const can be used as length

because lifetimes dont exist in c

[obsidian mantle]

[sage crag]

its deallocated, but c has no concept of lifetime

[opaque sigil]

int main() {
  int erm[variable] = {0};
  return 0;
}

committee angy

int main() {
  int erm[variable];
  memset(erm, 0, variable*sizeof(int));
  return 0;
}

committee happi

[sage crag]

also its only deallocated if its a stack array

[obsidian mantle]

tell me ONE example of where construction like this can cause any problem at all
int a = 5;
int b[a];

there are none

cosmetic shit

[sage crag]

this is just wrong

lol

its unclear whether a is const or not

so should the compiler initialise a and then do a computation on the stack?

or should it alias it and do it at comptime?

there are implications to everything you write in a systems language

dont blame it for not having syntax frills that dont make sense

[obsidian mantle]

whats the difference between 1 and 2

in this scenario it doesnt matter

pick any

mr compiler

[sage crag]

the compiler has no way of knowing this

[obsidian mantle]

just pick any and give warning

like g++ does

[sage crag]

that's stupid

[opaque sigil]

a sufficiently smart compiler should

[sage crag]

if it can be a footgun it should be an error

[sage crag]

you have no way of knowing how it will be used in ABI

[obsidian mantle]

compilers do random shit anyway all the time

i mean different compilers

[sage crag]

they do it according to specification

there is a specification

and for your knowledge,

#include <string.h>

int main() {
    int a = 5;
    int b[a];
    memset(b, 0, a * sizeof(int));
    return b[0];
}

this works just fine in c

[obsidian mantle]

is it true that clang gives faster binary than g++ as caibi said

[olive mango]

C

[sage crag]

no

[obsidian mantle]

no?

[sage crag]

clang and gcc are different compilers

[obsidian mantle]

they both give identical output?

[sage crag]

gcc will perform better on some programs and clang on others

[sage crag]

no

clang is more modern and will do better with complex programs than gcc but gcc has the benefit of being mature

[obsidian mantle]

i get it , so if anything is not specified in specification they do it as they like
if specified - strictly follow

[sage crag]

doesnt matter which one you use but it does matter if your program is so awful it doesnt work on both

[obsidian mantle]

except g++ doesnt follow it by compiling my a[b]

[sage crag]

you should always compile with -Wpedantic -pedantic to make sure your program will work on any c compiler

[olive mango]

just kill me, i use python

[obsidian mantle]

what do you mean "work on any compiler"

its compiled

no compiler needed anymore

[sage crag]

program not binary

[opaque sigil]

pedantic restricts it to the feature set guaranteed by the specification

[sage crag]

pedantic means you arent using some weird quirk of your c compiler in your program

[opaque sigil]

so it should work with every compiler that implements the spec properly

[solid bough]

toast to increase my speed of filtering the images, I do it in passes on one specific thing only now

[obsidian mantle]

ooh you mean so i see if my program will 100% compile on any compiler.
regardless of current successful build

[sage crag]

no, it means if it compiles on one c compiler it will compile on any c compiler

pedantic means:
compile here => compile everywhere

[solid bough]

I go through like 1.2 images per second on average

[sage crag]

no pedantic means
compile here => dunno if it works anywhere else

[obsidian mantle]

does it work cross platform too

[sage crag]

platforms are a different matter

they arent defined by the c spec

[obsidian mantle]

then pedantic only guarantees that it will compile on any compiler on this specific platform

[sage crag]

as long as your program uses platform agnostic libraries it shouldnt be an issue

[sage crag]

it guarantees that your program follows the c spec

it doesnt guarantee anything about platforms

the c spec doesnt mention platforms

its a language spec only

[obsidian mantle]

so if i compile my program with g++

and use pedantic

it will not compile

[sage crag]

g++ is gcc

g++ is gcc for c++

[obsidian mantle]

do i use both or pick 1

wpedantic pedantic

[sage crag]

i dont remember

use both

[obsidian mantle]

it gives warning

but still compiles

like it should

i want a warning not an error 😡

[sage crag]

[obsidian mantle]

if its a warning it should be a warning

anyway

how do i fix it

cast into const or smth?

[sage crag]

dont cast into const just define the length as const

int len_submoves ❌
const int len_submoves ✅

[obsidian mantle]

its a function input

which is a variable

[sage crag]

then you cant use it????

[obsidian mantle]

it will change between function calls

[sage crag]

thats a variable length array

make it a const parameter or

use the heap

[obsidian mantle]

im literally using it

so you suggest.. to define it as array[500] for example

[sage crag]

yes because g++ understands your garbage program probably meant to use const

[obsidian mantle]

and throw error when its above 500

[sage crag]

what are you on about

[obsidian mantle]

i want array

i want the array to have unlimited potential size

[sage crag]

you cant do that on the stack

[obsidian mantle]

but i dont need it to be unlimited in every time

[sage crag]

you will overflow it

use the heap

[obsidian mantle]

is it slower

[sage crag]

allocating on the heap is negligibly slower

better than running out of stack

[obsidian mantle]

then fuck the police im not doing that

why would i run out of stack

[sage crag]

you may have stupid

[sage crag]

its not unlimited

[obsidian mantle]

if i could do this
int a = 5
int b[a];

i would not have problems at all

[solid bough]

Where is the neck?

[sage crag]

you would

that is an array allocated on the stack

[opaque sigil]

just increase the stack size

[obsidian mantle]

and deallocated in 30 lines of code

[sage crag]

do you use any recursion in that function?

[obsidian mantle]

no

[sage crag]

do you call any recursive functions in that function?

[obsidian mantle]

no

[sage crag]

default limit is 16kb

[sage crag]

then you may be OK

[obsidian mantle]

i used to have recursions though in previous versions

how would that cause problems

new function call allocates new array

stack is 16kb???

[sage crag]

depends on the system, some have as little as 1024 kb

even less

[obsidian mantle]

i thought its ram

[sage crag]

you as a programmer dont deal with RAM

you deal with program memory

stack or heap

stack is the preallocated memory your program gets when it is created

barring some situations the stack never grows or shrinks

[obsidian mantle]

so wait, if i declared it as const, it would still be in stack?

or no

[sage crag]

...

its an array in the stack

regardless of what you use, int a[] = {} or int a[100]; or whatever

its an array in the stack

[obsidian mantle]

so if i have int a[100] and function called recursively 1000 times i will run out of stack and crash

[sage crag]

yes, for a simple example

the only time stack is freed is when the function returns

firstly, dont allocate or create arrays in recursive functions

[obsidian mantle]

they are small

[sage crag]

doesnt matter

[obsidian mantle]

then i cant allocate any variables at all

[sage crag]

secondly, if you do, either do it properly or use the heap

[sage crag]

incorrect

[obsidian mantle]

lets say its not int a[100]

but 100 different ints

[sage crag]

typical variables are stored in registers

not on the stack

[opaque sigil]

You can't stop me

[sage crag]

you are a competent programmer

[opaque sigil]

See that's where you're wrong

[sage crag]

[obsidian mantle]

why is stack so small

[sage crag]

because every program gets the same amount of stack

[obsidian mantle]

why

[sage crag]

and you may have thousands of programs running at once

[sage crag]

because that's how operating systems work

[obsidian mantle]

just use ram

[sage crag]

...

stack is ram

[obsidian mantle]

both heap and stack are ram are they no

[sage crag]

the kernel manages ram

[obsidian mantle]

i have 64 gb ram

[sage crag]

ok?

[obsidian mantle]

use it man why 16 kb

[sage crag]

...

ok so

you give a program all 64 gigs of your ram for stack

[obsidian mantle]

no

[sage crag]

oops you can no longer open any more programs

[trim valve]

[obsidian mantle]

i give it as much as it asks

[sage crag]

there is no "asking" for stack

[obsidian mantle]

when it allocates it takes some ram

deallocates - frees

[trim valve]

resizeable stack, pls invent konii

[sage crag]

stack is allocated before the program starts

[sage crag]

ye IT CALLED HEAP

[obsidian mantle]

can it cause problems with fragmentation?

[sage crag]

its not possible

the way stack allocation works is with a pointer incrementation

it requires contiguous memory

[trim valve]

but Reddit said stack fast heap slow

[obsidian mantle]

YESSSS

i get it

[sage crag]

ye reddit glueless

[obsidian mantle]

noow it all makes sense

[trim valve]

so everything must go on stack

[sage crag]

i dont think you do

[obsidian mantle]

this is what i meant by fragmentation

it cant use 2 chunks in corners of my ram

[sage crag]

yes this is the most basic aspect

[noble zodiac]

it as nothing to do with fragmentation

[obsidian mantle]

the arrays must be contiguous in memory

[sage crag]

well, that's not true

[opaque sigil]

Is the stack even guaranteed to be contiguous in physical memory

I don't remember

[obsidian mantle]

doesnt matter

[opaque sigil]

I guess it's 4 pages

[sage crag]

nope

only virtual memory

[tender river]

technically stack can grow through the kernel lazily mapping its pages

[obsidian mantle]

virtual memory != physical memoroy

[sage crag]

physical memory is your RAM

virtual memory is what you deal with

you never use "RAM"

you use heap or stack

[obsidian mantle]

ok then what happens when i declare a[50000] (const)

anywhere

[opaque sigil]

You get that

[obsidian mantle]

stack is 16k i want 50k

ohh

wait

[opaque sigil]

well nothing is going to happen until you actually use it

[obsidian mantle]

what about "all programs get equal"

[sage crag]

nothing is going to happen until you initialise it

[obsidian mantle]

so i literally cant have int a[50000]

[sage crag]

you can, it will just overflow the stack and crash the program if you write in it outside of the stack bound

[obsidian mantle]

i want to initialize it of course why would i make it that big otherwise

i will test it

[sage crag]

[opaque sigil]

There's nothing stopping you from writing into uninitialised memory as long as it belongs to you

[sage crag]

correct

[sage crag]

be careful, because compilers are smarter than you

[opaque sigil]

What about -O0

[sage crag]

it obviously work on -O0

[obsidian mantle]

nothing happens program works

[sage crag]

i literally just siad

[sage crag]

pls

[obsidian mantle]

explain i dont speak assembler

🥺

[opaque sigil]

It returned 0

[sage crag]

it optimised out the stack allocation

[trim valve]

it decided your function was cringe

[noble zodiac]

the compiler will ignore what you write if it does nothing of substance

[obsidian mantle]

WHAT

[sage crag]

basically your program was garbage

[noble zodiac]

its called optimization

[sage crag]

and the compiler decided to erase it

[obsidian mantle]

??????

no way

[sage crag]

yes its called optimisation

[noble zodiac]

shit guys, the jig is up, he called our bluff

[obsidian mantle]

no warning no nothing?

this is so funny

[trim valve]

why would it warn you

it still works

[obsidian mantle]

rebel of compilers, he doesnt want to compile what i told it to

[trim valve]

it's just faster

[trim valve]

don't you want fast?

[tender river]

no warning the compiler is allowed to erase code that wont do anything or does behavior that cant be relied upon

[obsidian mantle]

what do you mean it works

it erased my code

[trim valve]

it does what you asked it to do

[noble zodiac]

yes because your code did nothing

[opaque sigil]

It's the same outcome

[sage crag]

it erased your code because your code does nothing

[noble zodiac]

so why keep it

[opaque sigil]

Semantically equivalent

[trim valve]

its a function that does nothing and returns zero

so it made a function that does nothing and returns zero

[tender river]

c isnt assembly

[obsidian mantle]

ohhhh you mean if i do something useful it will crash stack overflow?

[trim valve]

tias

[sage crag]

oh

depends

[obsidian mantle]

i want to test a[50000] causing stack overflow

[sage crag]

here is what happens if i compile it without -O0

[sage crag]

here i did it for you

[obsidian mantle]

does -O stand for "optimization"

[opaque sigil]

I need that font

[sage crag]

SIGSEGV (Address boundary error) -> we wrote outside the memory the kernel gave to us, therefore program is kill

[sage crag]

Annotation Mono

[opaque sigil]

Oh right

That one

[sage crag]

i like it

similar to fantasque sans mono

and comic code i guess

[opaque sigil]

Is nice

[obsidian mantle]

it still works

oh wait

it wasnt compiled

[opaque sigil]

Less stylised if that's the correct word

[obsidian mantle]

because its clang

[sage crag]

[sage crag]

ye

also it has a font package on nixos

fonts = {
  enableDefaultPackages = false;
  packages = with pkgs; [
    annotation-mono
    noto-fonts
    noto-fonts-cjk-sans
    noto-fonts-emoji    
  ];
};

[opaque sigil]

I just saw

[sage crag]

[obsidian mantle]

it still works if i compile it like that

so what do i even do? use vectors?

[noble zodiac]

use the heap

[obsidian mantle]

how

[noble zodiac]

malloc

[obsidian mantle]

use_heap(please)

[tender river]

its not c

its c++

[opaque sigil]

vector

[tender river]

so yes vectors

[obsidian mantle]

so vectors are slower than arrays

[noble zodiac]

how dare you disrespect malloc, kinda rude

[obsidian mantle]

konii said heap is slower than stack

vectors heap arrays stack

[sage crag]

no

allocating on the heap is slower than allocating on the stack

[noble zodiac]

allocating on the heap is slower

[sage crag]

the heap is the same speed as the stack

[obsidian mantle]

yes im going to do many allocatings

[sage crag]

thats a skill issue

[olive sable]

[noble zodiac]

no, you allocate your array once and then only assign values within the already allocated memory

different things

[obsidian mantle]

actually i can just declare a[500] array hoping i will never get 500 legal moves

however...

[opaque sigil]

Wasn't there a hard cap

218 or sth

[sage crag]

their program allocates the array inside the function

which

is bad

[obsidian mantle]

how big is this i dont understand

[sage crag]

24

[obsidian mantle]

bytes or

[sage crag]

24

[tender river]

if you're optimizing for tokens you gotta use globals either way

[obsidian mantle]

24x64?

bits

24x8

[sage crag]

24Ui64 just means its the value 24 stored inside a uint64

[obsidian mantle]

array[500] will take 100 kb its over

oh

so 24 bytes (its sizeof(Move))

[sage crag]

why dont you just calculate it

500 things, each of size 24

do the maths

[obsidian mantle]

i thought its not 24

i thought its 24x8

+int which is 4 bytes i assume

28x500

12k ...... alot

ALSO I DONT WANT TO ALLOCATE THAT MUCH

usually its just 10-20

[opaque sigil]

I found it

[sage crag]

[obsidian mantle]

i cant this is stupid

[sage crag]

then dont

[opaque sigil]

218 is the max

there you go

[obsidian mantle]

max moves per board in chess?

[opaque sigil]

Not that it'll ever come up

[tender river]

just use globals

[sage crag]

correct

[obsidian mantle]

globals yes nice idea i'll do that

[opaque sigil]

max number of legal moves you can have

[obsidian mantle]

will it allocate this array somewhere else not in stack

[solid bough]

What is the chance that the right arrow key is gonna die first on my keyboard?

[sage crag]

yes, that is a global

[obsidian mantle]

[sage crag]

well

you are allocating another one

[obsidian mantle]

i feel programmism flowing through my veins after this conversation

[sage crag]

so you have the same problem still

[obsidian mantle]

oh

[noble zodiac]

that is, as the kids say, truly unfortnite for you

[obsidian mantle]

yes there are 2

[sage crag]

do they

[noble zodiac]

sure do

go ask them

[sage crag]

wrrr

god i hate programming

[tender river]

neuro cat mode on

[sage crag]

https://cdn.discordapp.com/attachments/1421123912720187512/1425566113755959386/neuroCatModeOff.webp

[opaque sigil]

[sage crag]

https://cdn.discordapp.com/attachments/1421123912720187512/1425568062551232512/neuroCatModeToggle.webp

[obsidian mantle]

is this good

[sage crag]

no

explode

[obsidian mantle]

whats wrong

it doesnt work without -c

[sage crag]

why are you using it

-c outputs an object file

[obsidian mantle]

this is why

ooh wait

[sage crag]

what are you using to link it

[obsidian mantle]

full command

im linking to c

.o

[sage crag]

linking to c isnt .o

thats just linking

why are you even using object files instead of the c src files

[obsidian mantle]

thats how shiro is going to do it i think

i forgot why

[sage crag]

g++ -O3 -Wpedantic bitboard.c chessapi.c cpp_bot_test11_negamax.cpp -o cpp_bot_test11_negamax

[tender river]

wont necessarily work since shiro uses c extensions not supported in cpp

[obsidian mantle]

interesting

i know why

because i guilt .o

with gcc first

because something didnt work

[sage crag]

why isnt shiro using extern c

[obsidian mantle]

and now i use g++ to link

[sage crag]

[tender river]

thats only for headers

not for c files

i think

[sage crag]

er

maybe

dont rember

[tender river]

c++ is best considered a separate language from c

less surprises that way

[sage crag]

true

i hate programming

[tender river]

cppreference says

Every function type, every function name with external linkage, and every variable name with external linkage, has a property called language linkage. Language linkage encapsulates the set of requirements necessary to link with a program unit written in another programming language: calling convention, name mangling (name decoration) algorithm, etc.

[sage crag]

[tender river]

extern "C" only used to change "linkage"

[sage crag]

g++ is gcc in a trench coat anyway

[tender river]

https://en.wikipedia.org/wiki/Compatibility_of_C_and_C%2B%2B

[olive sable]

[opaque sigil]

Isn't it just gcc with some extra params passed

[obsidian mantle]

my bot gets super bad move on depth3 iteration and stops depth4 because takes too long
what is proper mechanism to stop iterations

so bot manages to finish game in 1 minute

i give it "timeleft/60" per iteration

[tender river]

different binaries on my system

dont ask why -d0

[obsidian mantle]

if i give it more time per iteration it just wastes time in early game

[tender river]

wait why

why is cpp a separate binary from g++

[obsidian mantle]

so is this good or no

what is o3

optimisation?

[tender river]

it works just keep it

yes

O3 not o3 though

[obsidian mantle]

they are different?

[tender river]

important since -O is optimization options and -o is output file name

[obsidian mantle]

it takes both

-o3 is output file?

[sage crag]

purity or something

reproducible build

[tender river]

-o3 means output to the file "3"

[sage crag]

[tender river]

i refuse

[obsidian mantle]

what if i have -o3 and then -o file.app later

it ignores -o3?

[tender river]

maybe

[sage crag]

weliv

[obsidian mantle]

i dont have "3" file even though i built it with -o3 before

nvm it might have been error-build before

[tender river]

weevil

[sage crag]

[opaque sigil]

Hmm maybe it changed, I could've sworn I've seen g++ and clang++ as symlinks to gcc and clang before

[solid bough]

vituha did you ever wonder what images I consider as bad images for traingin? Something like that is not good:

[opaque sigil]

Unless it's a nix thing that they're separate?

[obsidian mantle]

because of hand?

i did wonder yes

[solid bough]

It's worse quality than the average images

[obsidian mantle]

and the guy on background

[opaque sigil]

I'm talking about something else

[solid bough]

For example, that image is better on quality:

[obsidian mantle]

are there no "ai training database faces" out there

[solid bough]

Sorry toast

[tender river]

stackoverflow

As far as I remember, GCC (not clang) has always shipped separate executables for gcc and g++. Early versions even shipped separate source archives for the C and C++ compilers. In some versions, g++ was a shell script calling gcc with additional options, but I’m not aware of it ever being a link to gcc.

[opaque sigil]

Hmmm

[opaque sigil]

Maybe I'm just misremembering

Or it's a clang thing only

[sage crag]

clnang

[obsidian mantle]

so is clang giving faster file or no

compared to g++

[sage crag]

we already talked about this

[obsidian mantle]

i forgot

you said no

[tender river]

doesnt matter in practice, wont be too different

[obsidian mantle]

[sage crag]

between gcc and clang the performance is very similar

[obsidian mantle]

i get really inconsistent depth search durations

depth1 is 10 ms
depth2 is 50 ms
depth3 is 1 ms
depth4 is 2500 ms
depth5 is 30 ms
depth6 is 30 ms

[sage crag]

sound like bug

[obsidian mantle]

it finds transpositions

i assume

like, wtf is this..

max depth was 8 but it stopped at 7 ply??

yeah must be a bug

is == two tokens

im talking about
x == 0
x < 1

[trim valve]

is "idk check" an ok answer

[obsidian mantle]

basically i dont care if its 0 or below 0

theoretically it cant be less than 0

[trim valve]

if you really care about tokens wouldn't if (!x) be better

[obsidian mantle]

it would

[sage crag]

if (!x)
  ...
else
  return;

if (x)
  return;
else
  ...

you can also see if you can do something like this swap

[trim valve]

true

they should invent reverse guard clauses

[sage crag]

rrr

[worldly plank]

Don't think so?

[trim valve]

int func(int arg) {
  if (check(arg)) {
    goto func_check_ok;
  }
  return -1;

func_check_ok:
  /* ... */
  return 0;
}

[opaque sigil]

== should be 1 token

[sage crag]

maybe with toknt

it wasnt with whatever we were using before

[opaque sigil]

Hmm

Really

[solid bough]

Why did I stumble across 3 Steve Jobs images already

[opaque sigil]

Well he is a famous person

[solid bough]

Well, Steve Jobs got a pass on glasses, but not the iPod

[sage crag]

int func(int arg) {
void **out = check(arg) ? &&check_ok : &&check_not_ok;

goto *out;

check_not_ok:
  return -1;
check_ok:
  return 0;
}

[trim valve]

good idea

we need this pronto

[sage crag]

no this works

[trim valve]

can you pr that to the linux kernel

[sage crag]

ye ok

[trim valve]

it would save at least 1 byte somewhere

[opaque sigil]

Why the double pointer

Does 1 not work

[sage crag]

nope

[opaque sigil]

Interesting

[sage crag]

its a gcc extension, but also its logical

its a label not a value

[solid bough]

toast, I interrupt for now just that one time now while I sift through the images, I can't make a model with 10K images

[opaque sigil]

The label still represents an address

[sage crag]

just be careful not to type goto **out;

[sage crag]

the label doesnt represent an address

it represents control flow

[trim valve]

[sage crag]

otherwise you get something or other

[trim valve]

smh konii why isn't your stack executable

[sage crag]

just to be clear, gcc doesnt care if you type void *out = &&lbl; or void **out = &&lbl; because its an informal extension

[opaque sigil]

Sounds like you need a 2nd model that can filter out the bad ones

[sage crag]

the double ref is required tho

[opaque sigil]

Interesting

[sage crag]

this technique is called computed goto

its used a lot in interpreters

[opaque sigil]

I remember there was a way to force a segfault in the compiler with something like this

[sage crag]

void *ops[] = {
  [vm_exit] = &&exit, [vm_print] = &&print,
};

int pc = 0;
goto *ops[bytecode[pc++]];

exit:
  return 0;
print:
  printf("hello, world!\n");
  goto *ops[pc++];

[tender river]

that would generally be a bug in the compiler

[sage crag]

doubtful

computed goto is pretty size heavy

its faster than a switch in some cases but it requires a big ol table

[trim valve]

😔

[sage crag]

i use it in my vm

[sage crag]

https://git.ablecorp.eu/lily-org/hbvm-c/src/branch/trunk/src/lib/hbvm.c
read it

[trim valve]

4 frame anubis jumpscare

devious

[sage crag]

who even is this random girl

[trim valve]

very hard challenge right there

[tender river]

rrrr_mdr

[sage crag]

anubis presumably

[tender river]

🇫🇷

[tender river]

its not me

[trim valve]

its not me either

[sage crag]

true

not me either

[solid bough]

Huh?

What is that about?

[sage crag]

ye your fault that this will now take 10 minutes on some random low end laptop

[tender river]

Making sure you're not a bot!
Done! Took 16823ms, 1 iteration

[olive sable]

[trim valve]

mandatory 10 minute reconsideration period before viewing hblang

[sage crag]

hbvvmmmm

fastest hbvm implementation

in da world

[olive sable]

only a human would have the patience to wait 10 minutes, surely

[sage crag]

i challenge anyone to make a faster version

[tender river]

git rerere

[sage crag]

its not about stopping web crawlers

its about making them really expensive

[olive sable]

i see

[sage crag]

it has the effect of reducing the web crawler traffic because it costs more

but its more of a "im shooting you" than "please leave"

[olive sable]

i have to do my blender homework

[tender river]

konii american

[tender river]

cheese powder

[trim valve]

smh anubis too fast I can't get a screenshot of how fast it goes

scheese

[sage crag]

seese chamwich

[olive sable]

make it slower by reducing the amount of resources your browser is allowed to use

[trim valve]

[trim valve]

tbh I don't even know how much that'd help

ig chrome can throttle cpu

[olive sable]

i know that for firefox you can test different patato settings

[trim valve]

where is the settings menu on a potato

[olive sable]

idk

i dont have a patato with me

[trim valve]

[solid bough]

The potatos I know have no settings menu

[sage crag]

must not have cooked them properly

[solid bough]

https://cdn.discordapp.com/attachments/1067638175478071307/1417224748093542520/neuro-dude.gif

[sage crag]

https://tenor.com/view/meow-gif-24525483

unhand me

[solid bough]

-# 200 OK

[maiden geyser]

why &&? isn't a label already a pointer to the code?