Rosy

[echo crystal]

ye that's what im using

[elfin cape]

lucky you

[echo crystal]

that's the plan ™

[elfin cape]

I use it at work because thats the only choice. Its nice but some things are not

I hate that you cannot open a file in raddebugger without doing the text tab file:"...".data thing

[echo crystal]

i like how they wanna explore visualisations more

[elfin cape]

also the file path mapper is nice

because I work on 2 computers at the same time and sending binaries to other computer it is really handy

[cloud rivet]

do you also use rad debugger?

I use it

[cloud rivet]

what is gpu mode? https://github.com/gpu-mode/lectures

a YT channel?

https://www.youtube.com/@GPUMODE/videos

feels like some kind of organization

[astral hinge]

a four letter organization

[cloud rivet]

no graphics content

what is a four letter organzation

hrmmm I see amd mentioned a lot

there's nvidia and apple gpu content though

idk

mysterious

I joined the discord

[astral hinge]

like a three letter organization but with an extra letter

so it's 33% spoopier

[cloud rivet]

it just seems a little AI pilled

boring

looks like there's a bindless extension for wgpu https://docs.rs/wgpu/latest/wgpu/struct.Features.html#associatedconstant.SAMPLED_TEXTURE_AND_STORAGE_BUFFER_ARRAY_NON_UNIFORM_INDEXING

[broken fog]

yea

unfortunately just an extension and not on webgpu

i could really use bindless right about now

[cloud rivet]

alright I'm going to add a slider to my ui

and a button

and then I can start on stuff

real 3D stuff

I think some of this server's activity has moved to other servers

that's fine

I shouldn't make it 4x4

it should be 2x4

32 * 4 * 2 = 256

I don't have 512 stuff

probably fine though actually

[astral hinge]

are you curious about using webgpu?

[cloud rivet]

hell no

I tried it once

it's actually cool, except for the bindless part, wgsl syntax is weird

I just got everything I need with

I love being able to send a giant buffer without describing anything about it other than the offset, a usage flag and how many bytes there are

[astral hinge]

btw you can simplify it even further because the usage flag has no effect on desktop GPUs (except for descriptor buffer)

[cloud rivet]

oh I mean memory usage

[astral hinge]

ah

memory type

[cloud rivet]

yes

[astral hinge]

yeah that's handy

vulkan is nice

[broken fog]

if by weird you mean ass then yes

[cloud rivet]

it's kind of interesting in how different it is, but ya

idk I made some spinny colored cubes with it iirc

[cloud rivet]

I was working on a sameline component

like imgui

and I have decided instead that it is easier to just have to specify next line

I like that it is explicit

otherwise it has to kind of be an undo operation or a weird state management

and I rather have an explicit API and simple logic without weird state or lookahead or other bs

and I want these things to be nodes and not attributes

and it can serve as a horizontal rule tbh

hrm

maybe that should be its own node_type

I kind of want tabs also

[astral hinge]

SameLine() could just set a flag and widgets only increment the line if the flag isn't set

and if the flag was set, they unset the flag

[cloud rivet]

but then every widget has to advance the line

[astral hinge]

I think explicit is fine though

[cloud rivet]

and do logic on that

[astral hinge]

yeah

[cloud rivet]

that is nicer than anything I was thinking of though

yes I like it explicit, nice

I think I will use a function pointer for button callbacks

hrmmmm

thinking

I think so

or I could return a bitmask

or you can give it a bitmask as an arg that can be read

oh no the bikeshed possibilities are growing exponentially

fuck it function pointer was my first idea

[cloud rivet]

@broken fog do you attend Universidad Nacional del Noroeste de la Provincia de Buenos Aires ?

argentina is so pretty

every time my daughter advances into some new stage of her life I completely change how I feel about other people in that stage. Now I see young people attending college and it makes me happy to think about their future and what they're going to learn and do in their lives. Like I feel hope. I did not used to think about college kids in this way.

[dark saffron]

how did you think of them

[broken fog]

nope, don't think i've heard of it either, why?

[cloud rivet]

Someone from there on linkedin added me and I was wondering who it would be

[cloud rivet]

Mostly just didn’t, felt nothing if I did I guess

[cloud rivet]

is nvim telescope dead

it has tons of deprecation warnings and no updates in months and tons of github issues and prs ignored

https://github.com/nvim-telescope/telescope.nvim/issues/3524

ah

the latest release is from 2024 and switching to the master branch fixed the issues

[cloud rivet]

my ui is getting a bit complex with multiple components having interactivity so switching to a unique ids and adding helper functions with safety bulit in, was just indexing into arrays yolo like before

[cloud rivet]

I'm also going to have nested components hrm like tabs and trees and whatnot

[cloud rivet]

#software-rasterization is weird, in api channels people state facts about how to use APIs and refer to specs, in that channel, it's just opinions, arguments and negativity and a weird gpus are bad take?

maybe that's harsh

it's just not my vibe

I just looked and see a huge argument about OOP lol

voxel's videos are really cool though

[bronze socket]

you might just be unlucky

also you say "in API channels" but you have not had your soul crushed by #opengl like deccer

[cloud rivet]

Yeah that’s a special case due to opengl popping up in gamedev beginner searches

[broken fog]

oh yeah no idea

i don't use linkedin like ever

[broken fog]

from what i've seen it's been taken over by a bunch of grumpy retro enthusiasts who screech any time someone asks about gpu sw raster

instead of just like

letting the convo happen

idk why

[cloud rivet]

I just don't find that a useful channel to participate or lurk in

[broken fog]

whatever this is the cooler sw raster channel 😎

[cloud rivet]

idk if you've seen dodo's his way cooler

Lucerna

[broken fog]

i haven't actually

👀

[cloud rivet]

#1311466891415519302 message

[broken fog]

neat

makes me want to do some sw raster

oh well i'll get the chance to improve the one in my kernel thing for uni soon

though i was also thinking of implementing a smol desktop environment

[cloud rivet]

how does the Argentina university system work? are they private or public or a mix of both?

just curious

[broken fog]

we have both

public unis are either completely fucked because they have zero budget or great but small and very hard to get in (hard cap on how many students are admitted every year)

[cloud rivet]

I see

[broken fog]

private ones you have all kinds of stuff, from meme unis to very good ones

and they go anywhere from affordable to pretty damn expensive

not us expensive though

[cloud rivet]

the military paid for my uni after I got out, I went to an inexpensive state school

[broken fog]

like they are "your family is well off" expensive but not "get in lifelong debt" expensive

[cloud rivet]

I see

[broken fog]

afaik we don't have such a thing as student loans

it's also a very different culture

[cloud rivet]

student loans are very bad imo

[broken fog]

like people don't move out to go to uni, unless you live in some province and are coming to buenos aires to study

but if you live with your family in ba you most likely stay there

[cloud rivet]

I think if student loans were restricted to majors that led to job outcomes it might be better, but also more unfair I guess

I knew people with 6 figure debt and musical degrees

[broken fog]

cause students just can't afford rent

[cloud rivet]

you can't declare bankruptcy or wipe your student debt

it has to be paid

[broken fog]

6 figure debt is absolutely insane regardless of the job you get

[cloud rivet]

yes

it's same as a house purchase basically

my school tuition was like $2k a semester back in 2002

[broken fog]

you can get through all of the most expensive private unis here and pay tuition + some basic rent without ever hitting 6 figures total money spent

[cloud rivet]

I didn't pay it though

[broken fog]

but then again our income is like 20% of the us

at least in software development

[cloud rivet]

it depends where you work in the US, if you work in the bay area you are doing really well

if you're working in oklahoma you are basically same salary as any professional service

[broken fog]

yeah here swe has pretty good salaries and benefits overall but like, all salaries are shit

i think we have like 40-50% of the coutry below the official poverty line or something like that

also uhh

[echo crystal]

hola

[broken fog]

(that wasn't even our worst inflation in the last decade)

[cloud rivet]

oh things got worse recently?

oh this is a long timeline

1825 days

[echo crystal]

https://tenor.com/view/pou-sad-cry-shower-plush-gif-9035760200285102054

[broken fog]

yes

the line getting flatter is worse actually

that's the usd being kept kinda artificially low

[cloud rivet]

I thought that chainsaw guy was fixing things

[echo crystal]

😂

[broken fog]

it's uhhh complicated and very schmolitical

at this point i honestly don't know if it's better or worse than the alternative

[cloud rivet]

oh ok, idk anything about politics or economics

[broken fog]

the thing is 2020-2023 we had massive inflation and the usd shot up

if you had savings in ars you were fucked ofc but no one is that dumb

and if you are in software engineering you were very likely being paid in usd so you could shrug off inflation and cost of living in usd was very low even though salaries were also much lower than in us/eu

but more recently the exchange rate is stuck, but inflation just kept going just a wee bit slower

so cost of living in usd/eur is now higher than many european countries

like, some stuff is more expensive than austria

anyway yeah the economy is fun

at one point we had 9(!) different exchange rates for ars-usd

[cloud rivet]

wow

[echo crystal]

wow

[broken fog]

a reasonable reaction

[bronze socket]

arbitrage opportunities abound

[cloud rivet]

I have to keep a window cursor to track inline positioning I realized, so I can detect the hover area for a button

buttons are a bit more complex than I expected

that'll apply to the slider too

every component has to update the window with how it impacts the cursor's x and y in the window so the next thing is positioned correctly

not just in terms of how how components are drawn

but also for handling events

this is made easier by being an explicit api

nothing wil be determined at draw time

my ui really had to grow up to add a button

unique ids, positioning cursor, new lines, helper functions

I thought I'd just draw a box and detect mouse coordinate and call a function

[cloud rivet]

idk how I raised such a talented kid, she sounds like Evanescence. she's got a performance coming up I'm going to record it and put it up if she let's me

wasn't me though, was her choir and guitar teachers and all the work she's put into it

almost done with this button

[cloud rivet]

something is making my frame rate like half

idk what

I'm going to work on the slider tomorrow

and then this weekend I"ll be into rasterizing

probably for a bit

[cloud rivet]

I should read through how box2D uses SIMD

https://github.com/erincatto/box2d/blob/main/src/contact_solver.c#L578-L674

see this makes sense it's using immintrin.h types and functions throughout and not converting back and forth, I think maybe my piecemeal approach was an issue

I'm not worried about it right now

just something to learn from

typedef __m256 b2FloatW;

hrmm

[cloud rivet]

it is done, I have a good enough UI to start rasterizing

it has edge cases and small issues but it's good enough to proceed, I'm not building a high polished UI here right now

[hushed creek]

What are you doing here lately tho I'm not 100% sure

Is this your own imgui?

[cloud rivet]

Yes

I am currently working on a cpu software rasterizer but I blit to a vulkan swapchain

This all otherwise cpu though, not gpu

[hushed creek]

O ok

[brisk chasm]

re #general or you coerce gobi to implement the missing feature in vichichi [veetcheetchee] 😄

im sure he is getting some motivation to work on it if more people use it and find missing things for him to fix/impl

padme-to-anakin-right?.gif

[cloud rivet]

we need to convince some mega corp with too much money to fund gob

[brisk chasm]

if he gets more exposure, then this could happen

[cloud rivet]

if someone pressured me on my open source project I would delete the project, which is why I don't have one

so I'm probably not the person, but I do want to support vcc

and Shady

#general message

lol

I should shut up

[brisk chasm]

hehe

[cloud rivet]

let me tell you about this thing I really like going over everything wrong with it first is totally how my brain works

[brisk chasm]

thats a totally legal thing to do

just like rolling all aspects of a game yourself, without firdparty libs

[dark saffron]

financially ? i have pretty much a dream job in this uni and unless you have enough money to hire more PhD students (and for that we have to make it a proper official thing through the uni, or maybe DFKI) there's little I can do with donations

I'm always up for for undergrad students/volunteers though, and we do have some budget to hire HiWi part-time (if you're studying at Saarland University, or possibly elsewhere in Germany)

[cloud rivet]

well if I know anyone like that I think might be talented enough to do a good job I'll try and send them your way

some resolution scaling, I refactored the code also now to accept meshes with vertices and indices

it's just a mesh with 3 vertices and 3 indices though still

I'll work on adding transforms tomorrow, and a perspective matrix, and after that I can add more triangles

I simplified the background also

[brisk chasm]

perhaps you could tap into government aids? (ah i didnt expect to come out like this but its also applicable i suppose in that context)

some game development fund bs they keep talking about

just make a game, call it vichichi [vee-tchee-tchee] (jaker may or may not tm'ed it already, hes 🇺🇸 after all) some cool monchichi jumpnrun which makes use of the tech

[dark saffron]

yeah we could try to get a DFG proposal

but it needs some sciency goals

[brisk chasm]

then you just found a company with your uni hiwis, and let the prof think about a goal

[dark saffron]

these proposals also take paper levels of time to write

and i kinda need to graduate 🐸

[brisk chasm]

that shall be your graduation project then 🪄

[dark saffron]

nope

[brisk chasm]

https://tenor.com/view/star-wars-obiwan-the-force-jedi-mind-trick-gesture-gif-5119850

[dark saffron]

i already have a thesis topic

[brisk chasm]

ok 🙂

then market it wherever possible on this server

[dark saffron]

you need to be a PhD to submit a grant proposal too, so if I get a post-doc I can write and submit one on my own, not relying on my prof or another post-doc

[brisk chasm]

ah

[dark saffron]

i'm actually going to keep the lid on what i'm working on next

i went public far too early with vcc

[brisk chasm]

vchichi 2 incoming

[astral hinge]

vcc 2

[brisk chasm]

https://tenor.com/view/my-man-together-smile-gif-20682410

[dark saffron]

opengl 5 obviously

[brisk chasm]

which has vccShaderSource, got it

a monchichi is probably a cool mascot for vcc

nevermind its probably copyrighted to death like ninendo 😄

[cloud rivet]

I'm going to solve my shader issues by just not having any

[brisk chasm]

i can see you speedrun ARB shaders already 😄

by temporarily going back to gl 2.0

[dark saffron]

gl 2.0 had glsl already

ARB assembly is from earlier

dx8 era hw

[cloud rivet]

I'm likely just going to have functions that have a builder state, and have builder functions that emit just the SPIRV I need, I'm not building a generic library for anyone to use, just going to solve problems I have

I'm sure anyone looking at whatever I come up with will go "but why" and that's fine

"but why" could be the tagline for my project

[brisk chasm]

ah i thought it was first introduced in gl 2, oki then

[cloud rivet]

I'm busy refreshing my understanding of LA and game math, rereading the first 7 chapters of essential math and first 3 chapters of lengyel's foundation book 1 before I get into doing anything

[tight torrent]

LA?

[brisk chasm]

LA is a dangerous place

[tight torrent]

oh linear alg

[brisk chasm]

yea 🙂

[cloud rivet]

these are 2 of the 4 books I read last year when I was working on my foundation project

I might add some more UI while I reread these, but focus is probably to just read for the next week and refresh it all

when I did this last year I wrote my own math library in zig, this time I will write my own math library in C

[broken fog]

heh i had to write my own tiny math lib in c for the sw renderer

kinda fun, only did the absolute necessities tho

not having the stdlib will do that

[cloud rivet]

I don't need or want a std lib

I like doing it myself

I got a profiler I can improve whenever I want, I have a UI I can hack on, I didnt' need Tracy and I didn't need Dear imgui and I don't need a glm or a stdlib

or an stb or anything else

[echo crystal]

you arent using the C std lib ?

or u mean u dont need the cpp std lib (which is why u are using C)

[cloud rivet]

there's no C std lib

there's the CRT

oh libc

I was thinking like std in C++

hrm

yes I use libc

it would be pretty hard to do anything without it

[cloud rivet]

yes

[echo crystal]

i see

[echo crystal]

maximum bikeshed

or the kernel stuff bluescreen was doing

[cloud rivet]

I guess it is called the C standard library, ok I didn't know that

I guess I do need it

https://en.cppreference.com/w/c/header.html

[echo crystal]

hehe

[broken fog]

you're not nihing until you have to roll your own printf

oh and no malloc cause you didn't implement a memory manager

[tight torrent]

i love syscalls

asm by beloved

[broken fog]

i actually get to work on that assignment again next week, gonna be fun

want to see if i can get a little gui desktop going

[broken fog]

i made a syscall to draw gouraud shaded trongles

[tight torrent]

[cloud rivet]

well, no I didn't actually

since I still used libc formatting

I see what you mean

[cloud rivet]

so this is actually creating your own character representation from integer and floating point types?

like looking at the bits of the sign, mantissa and exponent in a float to write some characters into a buffer?

[broken fog]

the custom printf? yea pretty much

i didn't add float support tho

didn't need it

and it's way more complex than anything else lol

chars and strings are trivial, ints are pretty shrimple

floats are where it gets complicated

i actually have fp math in the kernel which is kinda sus

but i don't care i want those 3d giraffics

[cloud rivet]

this runs in a virtual machine?

I just don't want to link third party code, I am not going to reinvent C or the operating system

[broken fog]

ye, on qemu

tried booting it on real hw but it didn't work on my old thinkpad

iirc it did boot on a uni laptop but it was unusably slow for some reason

but still seeing a thing you wrote actually boot on a real computer is cool

(ok we didn't write the bootloader part but still,,,)

[brisk chasm]

inb4 🟦 📺 coerces bjorn into BjornOS development

[cloud rivet]

nope, boring

:P

[cloud rivet]

in december I am going to pause on working on Palinode and just try to make a small game with it in what ever state it is in by then, maybe a small snake game or something

I could already do that now, so should be even more capable by then

copy paste the repo and change whatever I have to make a small game with what's there

that should be fun and instructive

so 2.5 months of changes should maybe have some 3D and lighting and controls, probably just on the CPU software raster side though, GPU stuff is blocked by not having shaders anymore and needing to generate spirv

[cloud rivet]

it's great re-reading the math, I lose nuance over time and forget things I'm not using, but I'm not using them because I forgot them. each time it's also less of a lift and I find I know some things now more intuitively

[cloud rivet]

writing a math library with tests now, this is going to take a bit as I read and write code probably a couple of weeks, nihmaxing

[cloud rivet]

I think what I am going to do is spend two months per quarter working on the renderer and one month on a small game, that’ll be 4 games a year

Tiny games time boxed to a month

Using this project

[astral hinge]

so what game are you making this month?

[cloud rivet]

In December

Just like a snake or tetris game

[cloud rivet]

depends on what palinode can do by then

probably will just be cpu software graphics since I don't have shaders

[astral hinge]

nice, very doable

[broken fog]

gta 6

[cloud rivet]

I'm glad I did all that hard work last year making my own math library and doing all those math visualizations, reading 4 different math books and writing tons of notes. It's all coming back to me quickly. It's kind of interesting how I forgot about some of the details, but just reading is a quick refresher. I should probably do this like once a year until I just got it down. Then I can dive more into some of the complex things I always have to look up and maybe someday it all becomes intuitive. I have to look up rotation matrices and reread the perspective matrix math every time I want to touch one of those matrices still

[pseudo dock]

What sorts of things does your math library do?

[cloud rivet]

I don't really like glm

[pseudo dock]

I've never used glm so I don't have an opinion 🙂
I always just end up writing my own stuff too, like you.

[astral hinge]

Impressive cube

How are you culling triangles?

[cloud rivet]

just via frustum and winding order

I added triangle clipping but it is slow af

so I turned it off

I need some kind of near plane clipping since the triangle will just disappear otherwise it will cause a division by 0 or appear upside down if I don't clip the full triangle whenever it crosses the near plane

[astral hinge]

Guard bands will make it so you don't need to clip on the sides most of the time

[cloud rivet]

can't guard band near plane as far as I understand the math

[astral hinge]

Yeah

[cloud rivet]

I have really learned to appreciate GPUs

I currently don't have a depth buffer, I think I will work on that next

then I'll add a torus, and then some lighting

[astral hinge]

Do you have an obj loader

[cloud rivet]

not yet

I will write one though

I think it would look horrible without a depth buffer, so I thin that should come first

[astral hinge]

Yeah for sure

[cloud rivet]

I really like the torus shape for working on lighting

[astral hinge]

How will you implement the depth buffer?

[cloud rivet]

was thinking just a f32 array

from 0.f to 1.f values

have a little debug visual so I can copy the values to the bitmap

[astral hinge]

What about the depth test?

I mean it's trivial with single threaded code

But if you want to process triangles in parallel, it's more complex

[cloud rivet]

it's still single threaded, but I was thinking I'd always be rasterizing one surface at a time?

my rasterizer is uh, a bunch of functions and some shoe strings tying it all together, I think I will pass the z around and have the depth buffer just on the context object

[astral hinge]

rasterizer should be independent of fragment tests imo

[cloud rivet]

well I have to know at pixel write time

that's happening in the rasterizer

[astral hinge]

in hw, the rasterizer generates a bunch of fragments and then you have early fragment tests, fragment shader, then late fragment tests and blending+srgbisms

[cloud rivet]

yeah

[astral hinge]

the tricky part is that all of it is an atomic operation basically

AND it has to happen in triangle submission order

or at least appear as though it did

[cloud rivet]

I put a wrong if statement anywhere and my frame time increases an order of magnitude

I'm going to think about the concept of fragments

[astral hinge]

frogments

[cloud rivet]

probably would be a rewrite to do that

it'll be interesting how all this changes once I start on the gpu software rasterizer

[cloud rivet]

added support for rendering multiple meshes, now ready for adding a depth buffer and depth test

[cloud rivet]

I can probably make some changes and reuse that "resolution factor" (happens at the end) as a generic quantization effect to use on the GPU, it looks kind of cool.

[broken fog]

perf is looking great, massive improvement from before

is this still single threaded?

oh ig it's also not rendering at 4k anymore lol

[cloud rivet]

It’s still 4k

Yes single threaded

The perf got a little better, but it depends on how many pixels I draw. The paint pixel function is a bottleneck at per screen pixel resolution

[cloud rivet]

I need a specialized draw function when it's directly to a screen pixel I think

[cloud rivet]

I will need a proper pipeline eventually, it's just a big for loop that descends into a function call stack in a single thread

[broken fog]

oh right

way fewer pixels covered than with the big trongle

[cloud rivet]

before:

now:

I used a win32 thread pool https://learn.microsoft.com/en-us/windows/win32/procthread/using-the-thread-pool-functions

I'm rasterizing 24 rows at a time

cut my frame time in half

haven't started on the depth buffer yet

I could batch the work differently

it ain't much, but it's better

see that frame time approach 40ms and sticking around 15ms in the second

i32 num_threads = NUM_RASTER_THREADS;

  for (i32 y = y_min; y < y_max; y += num_threads) {
    raster_arg rargs[num_threads];
    PTP_WORK work[num_threads];

    for (i32 yo = 0; yo < num_threads; yo++) {
      rargs[yo].t1 = t1;
      rargs[yo].bitmap = bitmap;
      rargs[yo].dimensions = (uint2){scale, scale};
      rargs[yo].raster_bounds = bitmap_dims;
      rargs[yo].bitmap_dims_f = (float2){width_f, height_f};
      rargs[yo].mins = (int2){x_min, y_min};
      rargs[yo].maxs = (int2){x_max, y_max};
      rargs[yo].pitch = pitch;
      rargs[yo].y = y + yo;
      work[yo] = CreateThreadpoolWork(raster_surface_pixel_row, &rargs[yo], &actx->crctx->thread_pool.cleanup_env);
    }

    for (i32 si = 0; si < num_threads; si++) {
      SubmitThreadpoolWork(work[si]);
    }

    CloseThreadpoolCleanupGroupMembers(actx->crctx->thread_pool.cleanup_group, FALSE, NULL);
  }

my debug build is also as fast as my release was prior to adding worker pools so that's nice

putting my CPU fan to work though

I should sleep the thread to cap max frame time to 16ms or something

I can reuse this thread pool for other things tbh

neat

[broken fog]

nice, threadpools are cool

[bronze socket]

have you tested against your own DIY threadpools?

I feel like you could get some improvement with some dedicated threads you feed with simple work queues

[cloud rivet]

Yeah I am going to wrap the windows thread pool into a job system where I can iterate on it.

It’s nice to see this lead to an improvement

I am unsure though why I if I wrote it myself it would be faster though

I am able to yield and wait these work threads

I think I need to break up the work better and think more on cache use across threads to get better performance right now

But it would not be a lot of work to implement myself

I just wrote all this last night after a full day of work so haven’t tested it against anything yet other than just yolo threads which ground everything to a snail’s pace

I want to make progress on depth testing now

[bronze socket]

it basically just depends on how much work you're doing to keep the threadpool fed, maybe the windows API for it has a negligible cost comparable to DIY thread pools

[cloud rivet]

I just kind of want to do my own thing anyway

[cloud rivet]

I think early tests to discard unnecessary work is next, which just checking the barycentric coordinate values of the bounds of the work group, and depth testing the pixels at those bounds will help with also

like if all four min max coordinates are not part of a surface, or occluded, I can avoid a lot of work

although maybe not in the latter actually

like maybe there's small surfaces occluding just the four corners nevermind, that won't work

I think occlusion culling will have to be its own thing via path tracing or something

if the bounds of one mesh are occluded fully by a single other mesh

[cloud rivet]

I like how people are able to use "Scene views" framed by a ton of unnecessary UI empty space to excuse having a smaller resolution render, I will use this trick and nobody will notice #showcase message

actually I just like the UI

the smaller render is just a bonus :P

I'm kidding anyway about it being an excuse

[pseudo dock]

This screenshot is from one of my favorite games, The Immortal. But even as a kid I noticed that all that space around the window wasn't doing anything (except for the health bar above)

(Some of the Ultima games also had small game windows, but at least the rest of the screen had other UI and served a purpose)

[cloud rivet]

looks so cool though

looks like original Diablo

[pseudo dock]

It also had great music (which I still listen to today). It was Will Harvey's studio.
If you've never played it it's probably hard to recommend today because it involved lots of trial and error and unfair deaths, but at the time it was amazing!

[cloud rivet]

yeah Blizzard came out with a remake for Diablo II and I loved that game so much when it came out, but when I tried the remake (which I bought) it was unplayable

I played so many hours of that game

when it first came out

I couldn't play it for 1 hour when the remake came out, it was so frustrating lol

Diablo II Resurrected

our standards for playability have changed, or at least mine have

[pseudo dock]

I've never played any Diablo games, believe it or not 🫣
I do like the vibe, but the gameplay never seemed like my thing.

What was frustrating about the remake?

[cloud rivet]

it's very difficult to see where the loot dropped, the navigation is really clunky and the pace of the game is very slow

it was just not any fun at all

I don't know, maybe I was having a bad day and I should try it again

I paid money for it and never even got past the entry level

[pseudo dock]

That must be frustrating... Maybe they've released patches to improve it

[cloud rivet]

I think gob left the server?

[broken fog]

yea

[cloud rivet]

trying to break thread pools

I set the minimum number of threads to 1, and max number to 500 and just submit each pixel in a row to a thread which is thousands

my frame time for max draw dropped by 5ms

let me up it to uh more?

I want to know where it breaks

it doesn't break it just doesn't get better

oh interesting, the more work I give it the better it performs, increasing the number of threads actually has less of an effect

I was being conservative just giving it a tiny bit of work at a time, but and waiting

I have 12 cores, but just giving it 6 threads do to the same amount of work just increased frame time by 1 ms

from 10ms to 11

I'm going to try to submit the entire number of pixels of the bitmap and wait for it

I'm doing science 🧑‍🔬

I've peaked

this went from 40ms frame time to < 9ms

just submit the number pixels in the window work items and that's best perf

with 2x number of cores

hrm

I switched from stack memory to heap, because I got a stack overflow trying to create too big of an array for all the args for the thread functions and it got < 1ms faster

anyway

this is good for now

time for depth testing

[cloud rivet]

I'll guess I'll work on a torus tomorrow and start on lighting

also I was stepping around the disassembly in a release build it is definitely vectorized, tons of SSE instructions

I think using the clang vetor types helps, I had align my arena memory allocations to 16 bytes, although I know nano said I can specify an alignment

let me see what that does

nothing

typedef f32 float4 __attribute__((ext_vector_type(4), aligned(4))); still needs to be 16 byte aligned like this

I don't want that anyway

[cloud rivet]

rendering is slow again

[astral hinge]

time to get a threadripper

[cloud rivet]

I'm no match for 500 trongles

it's actually not drawing that's slow now, it's just the number of triangles, it's slow at any resolution

[astral hinge]

so the vertex processing is slow?

[cloud rivet]

well

slower

drawing is adding up

hrm

damn

everything is slow

[astral hinge]

time to profile harder

[cloud rivet]

ya

[bronze tendon]

at least the time use was polite enough to bunch up mostly in one bin

[astral hinge]

sampling profiler time? or are you gonna instrument more

[cloud rivet]

going to try and do my own sampling

get some % of times, then average the times and then multiply by number of runs or some janky strategy

so like 100k iterations, sample like 1% of that, get 1000 times

average those

multiply by 100k

idk

I should read about it

tbh I think just that would be useful

I need to do that thread local for the stuff in the jobs

I can have them write to their input and once they've joined I can read from them

idk

[bronze socket]

you don't need to use thread_local, it can be slower than just using normal memory

just operating data-parallel without threads stepping on each others' toes and requiring fine grained sync should be enough

nvm I might have misread

[cloud rivet]

I don't actually qualify as thread local

the variables are just thread local because they're scoped at function scope in the job function

maybe I used the wrong words

[cloud rivet]

ok I set up a good test scene, that has a ton of things that need fixing

performance, there's a weird bug where I'm not drawing pixels, I think at values close to zero or something? my y axis is inverted, and scaling is affecting frustum culling in a weird way

[broken fog]

81ms in raster trongle

[cloud rivet]

I actually need a UI change first, I am going to add tabs, because I don't have enough space to add more tracing info

[broken fog]

is the issue fillrate or vertex calcs?

[cloud rivet]

nfi, everything

then I will work on a sampling tracer

[broken fog]

does fps get good if you zoom way out

[cloud rivet]

yeah drawing fewer pixels improves perf a little bit, but if I have just one giant triangle covering the full screen it's < 5ms per frame

so it's not drawing, but drawing is part of it

[broken fog]

ah not fillrate issue then

[cloud rivet]

oh I see what you mean, I didn't know what that word means

[broken fog]

oh ye it's like, how many pixels can you draw per unit of time

[cloud rivet]

I just have to measure and figure it out

[broken fog]

ig your bigger issue now is whatever you're doing per vertex is slow

[cloud rivet]

I like my hexagon floor though

I came by it by accident

I was working on making a sphere

and realized they looked nice tiled together as a plane

hexagons are cool

I think I could make a cool torus from hexagons since they tile so nicely

I should make all my shapes from hexagons including cubes

[broken fog]

they are the bestagons after all

[cloud rivet]

alright time to start on a tabs UI

I think the weird blank pixels is a bug in my pixel to screen space function or precision issue or something, it gets worse depending on the angle yea, just looking at it

hrm

[astral hinge]

maybe the rasterization rules you're using are not watertight

this page has nice illustrations
https://learn.microsoft.com/en-us/windows/win32/direct3d11/d3d10-graphics-programming-guide-rasterizer-stage-rules

[cloud rivet]

oh I think you're right since it's not a floatting point precision issue I can replicate with super huge pixels

lol

looks horrible lol

I'm going to read that thank you

I should share this ^^ in #software-rasterization to show off my skills

I don't have a concept of a pixel center

huh

idk what that means for my code right now

I have to think about it

pixels are basically points, like a thing that has no with or height, I guess that's a problem

having to do more math per pixel is not going to make my perf better lol

I'll figure it out

[cloud rivet]

fixed

I love those hexagon tiles so much

ok, now working on ui tabs

that was really bothering me so glad to fix those gaps

[cloud rivet]

TIL about clang blocks https://clang.llvm.org/docs/BlockLanguageSpec.html

I'm already all the way in with vectors and matrices

I'm going to use blocks and __block storage to power my tabs

I could have also used a static local variable, but those are kind of gross

idk

[astral hinge]

AKA gauge blocks
https://en.wikipedia.org/wiki/Gauge_block
||because you clang them together||

[cloud rivet]

I like those

Variables qualified by __block act as if they were in allocated storage and this storage is automatically recovered after last use of said variable. An implementation may choose an optimization where the storage is initially automatic and only “moved” to allocated (heap) storage upon a Block_copy of a referencing Block. Such variables may be mutated as normal variables are.

[cloud rivet]

oh

this is macos only lol

weird

I guess that was a waste of my time

static local variable it is

[cloud rivet]

I changed my colors around

it now cycles between a bright blue sky and a darker blue sky and I changed the UI colors and added tabs

I also figured out what was slow

I didn't even need to do sampling

I have to rewrite the entire pipeline

the problem is how often I wait

I just returned immediately from the jobs without doing anything and it was slow, it's just the job scheduling & waiting

I shouldn't have needed the thread pool, people are getting performance than I am with more triangles than I am just single threaded, I am going to start over on how it works

[astral hinge]

single threaded perf is the fundamental thing

[cloud rivet]

yeah

[astral hinge]

I consider multithreading to be a last resort unless it's super easy to add

[cloud rivet]

it's a fun problem though

[astral hinge]

rasterization does not seem trivial to parallelize

[cloud rivet]

I'm not frustrated or anything

[astral hinge]

yeah I didn't get that vibe

[vagrant musk]

I have made a horrible attempt at parallelizing rasterization using cuda and it uh
somehow a single triangle ended up with 30 ms render latency iirc

[bronze socket]

I wonder what strats are optimal for CPU software rendering, maybe some kind of tile based solution where you bin your triangles into tiles and rasterize each tile on a different thread

that would probably lead to minimal sync requirements

[bronze tendon]

I'd totally reach for tiling as a first option. Bet it works pretty nicely.

[cloud rivet]

so I broke apart my rasterizer into different stages where each stage generates a flattened array based on the previous stage: meshes -> produce a flat array of triangles -> triangles produce a flat array of fragments

well these are more like potential fragments

everything is fast as can be

per frame raster time < 1ms

so now I am in this function that operates on each potential fragment, which is basicaly all the pixels that exist within the bounds of each triangle

it does nothing

I add a simple addition

just one addition

frame time increases by 1ms

that's with the addition

1.2M potential fragments

I can't do shit in this function

no addition

[astral hinge]

and yet the CPU can do tens of billions of additions per second

[cloud rivet]

ok

oh

[astral hinge]

maybe the compiler optimized out a bunch of stuff when you didn't do the add, or it pushed some heuristic over a threshold that made it do something else, idk

[cloud rivet]

I lied

I was commenting out and commenting the addition

but the thing I was adding was derived from a function that was optimized out

ok so I changed it

it wasn't the addition great

working on that function now that added a 1ms! it's a tiny function

[cloud rivet]

single adds are impactful though, I just added of a simple float value to see how it compared to type casting an int to a float and it was meaningful frame time

they're both single ops

single instructions

I should look at the disassembly at this point

this is like squeezing blood from a stone

if I had tracy I would know more

I need to clean this shit up

I just need to pass in scalar values

well scalar values and the vectors

it's kind of cool though how fast everything else is fast

and completely isolated from this

at this point it's just a big array of fragments which I want to turn into just arrays of a depth and color pair that then test and paint

I'll keep going on this tomorrow

[pseudo dock]

I don't remember, is there a reason that you can't use Tracy? Or have you just not set it up yet?

[cloud rivet]

it's a from scratch project

no libraries or dependencies outside of the OS, the compiler and the vulkan driver and headers

I think counting the fragments is creating a dependency

it's a single integer

that probably is creating an issue for the compiler

[astral hinge]

are you doing anything with it besides incrementing?

[cloud rivet]

no

yeah I need to stop guessing

[astral hinge]

look at the disasm

[cloud rivet]

yeah

[astral hinge]

I guess that means you need to write your own debugger

[cloud rivet]

:P

I use rad debugger

[astral hinge]

that's rad

debugger

[bronze socket]

it could be a lot of things but I'd bet a lot has to do with the overall memory access patterns

[bronze tendon]

cache is king

[cloud rivet]

I allocate memory enough space for all 12M fragment structs and then just iterate through the fragment structs to do calculate the barycentric coords, mix color and depth, and then draw, I'm not at home right now so can't share the code, but it's a small struct, should fit in the cache, and I am just iterating over the memory in the sequence it is written. If I do very little work in that loop it is < 1ms frame time

it's very interesting

this morning before my work commute I removed basically everything that loop does, and I will slowly add operations to it to see how it impacts the performance

there are no function calls, it is only working with the memory in the struct, although I do have a pointer to the bitmap and its dimensions in scope

it's hard to imagine that I am having any kind of memory access problems in this scenario based on what I understand, I will try and experiment more with it and work through the disassembly

if it does turn out to be a memory access issue I'll give in and add Tracy to investigate that

[bronze socket]

you should try to switch up the memory layout and see how it affects it

maybe try to make it a SOA

[cloud rivet]

hrm

that's interesting

there's probably some alignment padding with the clang vectors

since they are 16 byte aligned

so maybe that will be a meaningful change

[cloud rivet]

I was so motivated to work on this I didn’t want to go to sleep last night and I woke up this morning to hack on it a little lol

Gob pinged me on a change he made to vcc to add a DSL to configure/declare SPIRV intrinsics in vcc

It’s really cool

It’s still in draft but it’s pretty awesome

https://github.com/shady-gang/shady/pull/57#issue-3491858023

[cloud rivet]

AOS vs SOA

[bronze socket]

so SOA was like a 20x+ improvement?

[cloud rivet]

 p_fragment_t *frags = p_malloc(sizeof(p_fragment_t) * num_fragments);
...
frags[i].screen_coords = (float4){((f32)frags[i].pixel_coords.x), 0.f, 0.f, 0.f};

vs

 p_fragment_data_t frag_data = {0};
  frag_data.t1 = p_malloc(sizeof(p_triangle) * num_fragments);
  frag_data.bc = p_malloc(sizeof(float4) * num_fragments);
  frag_data.color = p_malloc(sizeof(float4) * num_fragments);
  frag_data.screen_coords = p_malloc(sizeof(float4) * num_fragments);
  frag_data.pixel_coords = p_malloc(sizeof(int2) * num_fragments);
  frag_data.depth = p_malloc(sizeof(f32) * num_fragments);
...
 frag_data.screen_coords[i] = (float4){(frag_data.t1[i].positions[0].x), 0.f, 0.f, 0.f};

[cloud rivet]

yeah, but this is just test code in my loop of 12M things, it's not rendering yet, I confirmed it was just accessing data that was slow

well it does actually render

but to one pixel

all 12M fragments are currently painting the same pixel just so the loop does something

let me slowly start adding instructions back in to see what happens

[pseudo dock]

Which of those fields in the struct were actually being accessed?

[cloud rivet]

  i32 num_fragments = 0;
  {
    for (i32 si = 0; si < num_surfaces; si++) {
      for (i32 y = surface_ctxs[si].y_min; y < surface_ctxs[si].y_max; y++) {
        num_fragments += surface_ctxs[si].x_max - surface_ctxs[si].x_min;
      }
    }
  }
  actx->crctx->num_fragments_generated = num_fragments;
  p_fragment_t *frags = p_malloc(sizeof(p_fragment_t) * num_fragments);
  
  p_fragment_data_t frag_data = {0};
  frag_data.t1 = p_malloc(sizeof(p_triangle) * num_fragments);
  frag_data.bc = p_malloc(sizeof(float4) * num_fragments);
  frag_data.color = p_malloc(sizeof(float4) * num_fragments);
  frag_data.screen_coords = p_malloc(sizeof(float4) * num_fragments);
  frag_data.pixel_coords = p_malloc(sizeof(int2) * num_fragments);
  frag_data.depth = p_malloc(sizeof(f32) * num_fragments);

  {
    // Generate pixel coordinates for frags
    i32 fi = 0;
    for (i32 si = 0; si < num_surfaces; si++) {
      for (i32 y = surface_ctxs[si].y_min; y < surface_ctxs[si].y_max; y++) {
        for (i32 x = surface_ctxs[si].x_min; x < surface_ctxs[si].x_max; x++) {
          frag_data.t1[fi] = surface_ctxs[si].t1;
          frag_data.pixel_coords[fi] = (int2){x, y};
          fi++;
        }
      }
    }
  }
  u8 *bm = raster_ctx.attachments.bitmap;
  i32 pitch = raster_ctx.ro_raster_ctx.pitch;
  u64 *pixel = (u64 *)bm;
  f32 width_f = raster_ctx.ro_raster_ctx.double_scaled_dims_f.x;
  f32 height_f = raster_ctx.ro_raster_ctx.double_scaled_dims_f.y;

  for (i32 i = 0; i < num_fragments; i++) {
    frags[i].screen_coords = (float4){((f32)frags[i].pixel_coords.x), 0.f, 0.f, 0.f};
    frag_data.screen_coords[i] = (float4){(frag_data.pixel_coords[i].x), 0.f, 0.f, 0.f};
    frag_data.color[i] = (float4){1.f, 0.f, 0.f, 0.f};
    frag_data.depth[i] = 0.1f;
    frag_data.bc[i] = (float4){1.f, 0.f, 0.f, 0.f};
    rgba_to_rgba16(frag_data.color[i], pixel);
  }

is the relevant test code

I have both lines uncommented here

    frags[i].screen_coords = (float4){((f32)frags[i].pixel_coords.x), 0.f, 0.f, 0.f};
    frag_data.screen_coords[i] = (float4){(frag_data.pixel_coords[i].x), 0.f, 0.f, 0.f};

but in my test I just had one or the other

rgba_to_rgba16 actually writes to the bitmap

you can see that none of the other fields are currently being used

all the slowness is in the for (i32 i = 0; i < num_fragments; i++) { loop

[bronze socket]

yeah it's pretty clear to see why that'd cause cache perf gain then

[cloud rivet]

why is that

[bronze socket]

you can keep the stuff you actually use in cache without it being padded out by the parts of the struct you don't

basically the concept behind SOA in general

[cloud rivet]

ah right

typedef struct p_fragment_t {
  p_triangle t1;
  float4 bc;
  float4 color;
  float4 screen_coords;
  int2 pixel_coords;
  f32 depth;
} p_fragment_t;

typedef struct p_fragment_data_t {
  p_triangle *t1;
  float4 *bc;
  float4 *color;
  float4 *screen_coords;
  int2 *pixel_coords;
  f32 *depth;
} p_fragment_data_t;

yeah I have padding issues I noticed with the vector types

I saw in the debugger that it would add padding in other instances

I didn't realize that would cause issues here

[bronze socket]

yeah padding as in perpetually-unused space would be even worse, but it's also in the context of the loop

if you're not using depth for example, it's as good as unused space for a tight loop

[cloud rivet]

oh I will use all of those

this was just testing

I need to start adding functionality back in

hrm

yeah I'll just test as I add pipeline stuff back in and see what happens

[cloud rivet]

although I wasn't using them in the same loop so it was dead space actually

[pseudo dock]

Always cool/interesting to see a real-life example of the benefits of cache friendliness in the wild

[bronze socket]

yeah thats what I mean

[cloud rivet]

over the last few days something I may have noticed is that conditional logic to avoid doing a small bit of extra work actually doesn't seem worth it, it's either not much of a win, or hurts perf

idk I haven't really confirmed it

[astral hinge]

it depends on several factors

e.g. likelihood of taking a particular side of the branch, where the compiler put the code for each side, and whether the compiler actually emitted a branch instruction

[pseudo dock]

I read a book that had some content about performance related to branches that I found interesting: https://www.amazon.com/gp/product/B09BZTGJM2/ref=ppx_yo_dt_b_d_asin_title_351_o07?ie=UTF8&psc=1

[cloud rivet]

the author looks like me lol

[broken fog]

x86 krill issue for not having conditional instructions

but yeah if the branch is consistently taken/not taken the branch predictor will get it right basically 100% of the time and the perf hit shouldn't be too bad

then again, if the work you're avoiding is like two adds and a multiply it's probably not worth it

always benchmark

[cloud rivet]

idk, that's just as good as it's going to be for a bit

have some frustum bugs

I kind of want to work on a gpu thing and take a break from the software rasterizer for a bit

I'm going to work on my spirv builder idea

see if I can replace my shader hello world triangle

[cloud rivet]

I need to get my ui rendering on gpu renders

I think I just need to set the clear color to 0 for the bitmap each frame

hrm

and then I sample it from a full screen triangle

[bronze tendon]

love to see this update

[cloud rivet]

Same

So my next goal is a hello triangle with my own spirv builder and then make that a full screen triangle I will sample my cpu raster bit from so I can have a UI with my gpu render

[bronze socket]

ooh you decided on writing a spir-v VM as well?

or rather just having your software rasterizer run spirv somehow

[cloud rivet]

my project is not a software rasterizer, it has a software rasterizer

I was intending to use spirv for the gpu stuff

TIL about https://github.com/heroseh/hcc

[tight torrent]

isnt that just vcc?

[cloud rivet]

no

it's a different project

with different goals

hcc looks like it is a shader language

[tight torrent]

i meant like, isnt that just trying to do the same thing as vcc

[cloud rivet]

no

vcc's goal is to be able to write programs on the GPU using the full capabilities of a programming language, like pointers everywhere

recursion

etc

vcc doesn't have these limitations https://github.com/heroseh/hcc?tab=readme-ov-file#limitations

hcc is more of a shading language in C

see https://shady-gang.github.io/vcc/

Vcc supports advanced C/C++ features usually left out of shading languages such as HLSL or GLSL, in particular raising the bar when it comes to pointer support and control-flow:

https://shady-gang.github.io/vcc/why/

Gob has a very strong opinion about the unnecessary limitations of shader languages and a vision for how it should be, it's a very different project.

I actually don't know what I want to do right now

I think I might fork palinode and just make a game

and I can add whatever libraries I want to the fork just to make the game

and then go back to palinode when I get bored with that

and add anything that might useful to palinode back from what I did in the game

[pseudo dock]

Actually making something (like a game) is a good way to figure out what an engine needs. Seems like not a bad idea if you want a break to do something different!

[cloud rivet]

ok, I got an idea

I think not too hard

[pseudo dock]

An MMO?

[cloud rivet]

lol no

[pseudo dock]

Sorry, I didn't mean to interrupt your idea with my silly comment... Tell us your idea!

[cloud rivet]

wasn't silly at all, was funny :P

it's gonna be a small racing game

maybe just a race against time to avoid having to do any game AI, at least at first I don't know, anyway, just a small thing

going to cut every corner and reduce scope to bare minimum

[pseudo dock]

Seems like it could work and have a pretty limited and feasible scope

[cloud rivet]

I would like to use splines and a mesh shader for the track

[cloud rivet]

switching between the graphics pipeline on the GPU and my software rasterizer (via key pressing so it's a little janky)

I didn't have any 3D set up via the vulkan graphics pipeline, I didn't have a depth image or depth testing, no mesh or index buffers, no transforms or camera or scene data. So I just set it all up the last couple of days

[cloud rivet]

next thing is to get my ui to render on the graphics pipeline

[cloud rivet]

this AI review bot has saved me from so many problems and headaches over the last year

GPUMeshVertex exists because the clang language extension vectors can't be sent to the GPU as they are

the bot understands that, which is cool

typedef struct MeshVertex {
  float4 position;
  float4 color;
  float4 normal;
  float2 uv;
} MeshVertex;

typedef struct GPUMeshVertex {
  f32 position[4];
  f32 color[4];
  f32 normal[4];
  f32 uv[2];
} GPUMeshVertex;

[astral hinge]

why not? do they not have the same representation in memory as sequential floats?

[cloud rivet]

  vk_buffer_t source = actx->gctx->staging_buffer;
  size_t total_data_size = 0;
  for (i32 i = 0; i < num_meshes; i++) {
    size_t data_size = mesh_data[i]->num_vertices * sizeof(MeshVertex);
    memcpy((u8 *)source.data + total_data_size, mesh_data[i]->vertices, data_size);
    total_data_size += data_size;
  }
  VK_CHECK(copy_buffer(actx, source.vk_buffer, dest.vk_buffer, total_data_size));

  vk_buffer_t source = actx->gctx->staging_buffer;
  size_t total_data_size = 0;
  for (i32 i = 0; i < num_meshes; i++) {
    GPUMeshVertex *gpu_vertices = p_malloc(mesh_data[i]->num_vertices * sizeof(GPUMeshVertex));
    for (i32 vi = 0; vi < mesh_data[i]->num_vertices; vi++) {
      MeshVertex mv = mesh_data[i]->vertices[vi];
      gpu_vertices[vi] = (GPUMeshVertex){
          .position = {mv.position.x, mv.position.y, mv.position.z, 1.f},
          .color = {mv.color.x, mv.color.y, mv.color.z, 1.f},
          .normal = {mv.normal.x, mv.normal.y, mv.normal.z, 1.f},
          .uv = {mv.uv.x, mv.uv.y},
      };
    }
    size_t data_size = mesh_data[i]->num_vertices * sizeof(GPUMeshVertex);
    memcpy((u8 *)source.data + total_data_size, gpu_vertices, data_size);
    total_data_size += data_size;
    p_free(gpu_vertices);
  }
  VK_CHECK(copy_buffer(actx, source.vk_buffer, dest.vk_buffer, total_data_size));

the docs don't have anything about their layout in memory

what's interesting is in renderdoc the vector types look correct

one thing I know about them is they are 16 bytes aligned

because if I don't align the memory I allocate that way the application crashes

Note: The implementation of vector builtins is work-in-progress and incomplete.

[astral hinge]

oh ok

[cloud rivet]

what's interesting is that I'm using the OpenCL vector type

which if you use OpenCL works on the GPU? I don't know anything about OpenCL

[cloud rivet]

I think I will need a per frame image for my UI

I'm going to have to do some math for these texture coordinates

oh ezpz

[cloud rivet]

I remember someone saying that there's a severe limit on many memory allocations you can do in vk and I checked my vulkan info

 maxMemoryAllocationCount                        = 4294967295

it's max32 int

so

        maxMemoryAllocationCount                        = 4294967295
        maxComputeSharedMemorySize                      = 49152
        minMemoryMapAlignment                           = 64
VkPhysicalDeviceCopyMemoryIndirectPropertiesKHR:
VkPhysicalDeviceExternalMemoryHostPropertiesEXT:
        maxTaskSharedMemorySize               = 32768
        maxTaskPayloadAndSharedMemorySize     = 32768
        maxMeshSharedMemorySize               = 28672
        maxMeshPayloadAndSharedMemorySize     = 28672
        maxMeshOutputMemorySize               = 32768
        maxMeshPayloadAndOutputMemorySize     = 48128
        maxMemoryAllocationSize           = 0xffe00000

I'm going to just do one off memory allocations until something breaks or gets slow

I'm not cargo cult building a GPU memory allocator unless there's a reason for it that I run into

maybe it's something those poor souls who want to support android have to deal with

[cloud rivet]

it wouldn't be too much work to add a basic gpu allocator though

I just want to kind of see something break or behave badly to see what happens

it's a bit too much work still to add another graphics pipeline with new shaders and draw commands I need to make that more ergonomic

[cloud rivet]

spent some time today on improving my neovim experience, it is going well

I improved the status line and have information about what struct or function I am in, turned on line numbers and cursor line, which I think really helps

fixed a bunch of lsp problems

[cloud rivet]

have the UI on the gpu now

it's really expensive

it's like 2-3 ms clearing the buffer, rendering it and copying it to the gpu

idc

I'm going to work on the game now

I think I can move the UI to its own thread maybe

I'm not going to worry about it right now I have a lot of frame time to spare for now

I used this full triangle solution https://wallisc.github.io/rendering/2021/04/18/Fullscreen-Pass.html

    float2 uv = float2((id << 1) & 2, id & 2);
    return float4(uv * float2(2, -2) + float2(-1, 1), 0, 1)

so yesterday I built a 3D graphics pipeline for a scene, and today I got my UI to render on the GPU, those each took a full day's worth of work, which I wish I had made some progress beyond it, but it is what it is. I need to reduce the amount of code it takes to make these kinds of changes

I think what I might do is start a ui raster job before the submit and wait for it after present, but again, just not going worry about it right now

[cloud rivet]

I love these graphs the ai review bot creates

[cloud rivet]

placeholder vehicle

going to add a super tiny placeholder track

once I have a little track I'll need shadows, I'm going to use the RT pipeline I set up for shadows, that may take a bit

I have an RT pipeline set up but it just does a triangle

I think I will also use the RT pipeline for the background scenery that isn't anything on the track

the track will be via mesh shader

once I have the track and shadows I'll start on the gameplay

with just this little test thing

I will need physics to accelerate the thing

also if you go off the track you fall into obliviion

I think the initial gameplay will just be beat your previous time for now

[bronze tendon]

My friends and I were going to try and make a racing game for a game jam and I spent most of the time bikeshedding a tool for generating tracks

Marching squares can be handy for turning MS paint drawings into meshes

[cloud rivet]

I want to just do simple splines

it's going to be a 3D race track

there's going to be elevation

I don't make two dimensional games

[pseudo dock]

Do you not like them or are you just not interested in making one?

[cloud rivet]

not interested in making one

I love a lot of 2D games

I like doing 3D stuff is all

in terms of making something 2D is super boring for me

I want to be able to look around in a virtual world

I want to feel like I'm in the world I made

[pseudo dock]

That makes sense, and I think I'm the same way 🙂

Well, I don't know if I would say "boring", but I guess 3D is more interesting to me to actually work on

[cloud rivet]

I have dramatically reduced the amount of work it takes to create a new graphics pipeline, finally

internal VkResult init_road_pipeline(AppContext *actx, Arena *arena) {
  if (!actx)
    fatal("no actx");
  if (!arena)
    fatal("no arena");

  FileData road_shader = p_read_file(actx->arena, road_shader_path);
  FileData source_spirv[] = {
      road_shader,
      road_shader,
      road_shader,
  };
  const char *function_names[] = {"road_task", "road_mesh", "road_frag"};
  VkShaderStageFlagBits stages[] = {
      VK_SHADER_STAGE_TASK_BIT_EXT,
      VK_SHADER_STAGE_MESH_BIT_EXT,
      VK_SHADER_STAGE_FRAGMENT_BIT,
  };
  constexpr i32 num_shaders = 3;
  VkPipelineShaderStageCreateInfo create_infos[num_shaders];
  p_shader_cfg_t shader_cfg = {
      .source_spirv = source_spirv,
      .function_names = function_names,
      .shader_stages = stages,
      .create_infos = create_infos,
      .num_shaders = num_shaders,
  };
  p_create_shaders(arena, &shader_cfg);

  p_pipeline_layout_cfg_t pipline_layout_cfg = {0};
  pipline_layout_cfg.push_constant_size = sizeof(road_pc_t);
  pipline_layout_cfg.debug_name = "rosy v3 road_pipeline_layout";
  VK_CHECK(p_create_pipeline_layout(actx, arena, &pipline_layout_cfg));
  actx->gamectx->vk_road_pipeline_layout = pipline_layout_cfg.pipeline_layout;

  p_pipeline_cfg_t pipeline_cfg = {0};
  pipeline_cfg.shader_cfg = &shader_cfg;
  pipeline_cfg.debug_name = "rosy v3 road_pipeline";
  pipeline_cfg.pipline_layout = actx->gamectx->vk_road_pipeline_layout;
  pipeline_cfg.pipeline = &actx->gamectx->vk_road_pipeline;
  pipeline_cfg.enable_depth = true;
  VK_CHECK(create_graphics_pipeline(actx, arena, pipeline_cfg));
  return VK_SUCCESS;
}

I don't have materials yet, obviously

it's all very basic still

[cloud rivet]

stuff is coming together, deleting the old stubbed out demo graphics and mesh shader pipelines, things are starting to feel focused

and purposeful

[cloud rivet]

I'll start on the RT shadows tomorrow, that will probably be a bunch of work for a few days

you can't tell, but the vehicle is floating, it's not going to be a car game

shadows will help

I should add a scrollbar to my ui at some point

[cloud rivet]

doing some research, people often use a g-buffer to generate ray queries to use for ray traced shadows

and a g-buffer is deferred rendering, which is not what I have right now

I really want shadows

I'm just going to add a simple single shadow map for now

I'll briefly look at doing a ray query

I don't know anything about those

oh I think that works

just load the meshes into the AS, and then see if anything at all obstructs the light direction, and if so the light is occluded

don't need to do set up the RT pipeline for it or have an SBT

and it will only do it if the fragment gets any light already

and I only need to do it for the road right now, since there's nothing that could block light from hitting the vehicle

ezpz maybe

I already have BLAS and TLAS setup code

https://github.com/KhronosGroup/Vulkan-Samples/blob/main/samples/extensions/ray_queries/ray_queries.cpp
https://github.com/SaschaWillems/Vulkan/tree/master/examples/rayquery

[tight torrent]

Yep, and then if it's not occluded, inverse square law has you covered for the lights intensity

I have it set up in my dx12 renderer where I output an intensity from my shadow pass from 0.0 to whatever, 0 if in shadow or calculate intensity and output that

Then I just put that into pbr calculations or whatever

[cloud rivet]

oh does dx12 allow ray queries from pixel shaders?

that's cool thank you

I imagine you don't use light intensity for sunlight though?

that's cool I can use additional ray queries for other light sources hrm nice

[tight torrent]

Not sure

I have a traditional rt pipeline setup

[cloud rivet]

are you using just RT or also a graphics pipeline?

[tight torrent]

Just RT

[cloud rivet]

nice

[tight torrent]

Doing a rewrite in vulkan with a raster gbuffer and rt shadows tho

[cloud rivet]

I'm going to use the RT pipeline for everything not the road and vehicles and things that are not part of the race

[tight torrent]

It's the same code path, but returns either a flat 1.0/sunIntensity or 0.0

[cloud rivet]

ah yeah maybe I end up doing this too eventually, but just to get a game going I will try and start with ray queries in the fragment shader, I don't have a lot of stuff so should be cheap I hope

I also won't have any skinned meshes in this game

the other thing that's nice about ray queries is I could use MSAA, whereas I'm not sure I could with deferred rendering

I'm not worried about AA right now

[tight torrent]

just looked at the spec, it does actually

inline RT is a thing, i havent touched it tho

[tight torrent]

raytracing is never cheap lol

[broken fog]

is this still cpu rasterized or are you doing sw raster on gpu?

[cloud rivet]

this is software rasterized on the GPU now

the UI is software rasterized still

[tight torrent]

why not just do normal gpu raster

[cloud rivet]

because I already wrote it for the CPU I guess

I don't want to redo it, it's fine, it's super fast

except for the upload the image thing

[broken fog]

where's the fun in that

[cloud rivet]

I'm going to eventually upload the UI image in a separate transfer queue via a thread or something

it's just a debug ui

it's not the game UI

[cloud rivet]

I'm sorry I mispoke, this just normal graphics pipeline stuff, it is not software rasterised on the GPU it's just the normal GPU pipeline

I don't know why I said that

[broken fog]

oh

will you be using the cpu raster stuff for anything?

[cloud rivet]

if it makes sense to, maybe for UI editor widgets or something

maybe all my debug lines are just software rasterized tbh

I don't know

it's just a tool in the toolbox I have

[cloud rivet]

so Sascha's ray query example is exactly what I want to do

I'm just gonna try it

[tight torrent]

good luck!

RT shadows are wayyyyy easier to implement lol

[cloud rivet]

I cleaned up my BLAS and TLAS creation code so I can use my mesh types to create arbitrary tlas/blas now, with support for instances, and it also sets up the descriptors too and adds them to the big bindless set. Should just be mostly shader work now to do the ray query and add a shadow. I also added the ray query extension and features enabling.

one of the issues with using RT shadows is my track is created in the mesh shader and has no vertices to stick in an AS

I may bake its self shadowing

I think so

hrm

I'm not going to worry about it right now

[astral hinge]

that would be a neat thing to have

AO baker

[cloud rivet]

yeah

well the track is going to get procedurally generated from a long spline

it'll have a spline parallel to the track and then an array of curves perpendicular to the track as well

I am going for this https://www.youtube.com/live/MwnmN9yrCPI?si=dO2oJhXhb6GlnKNq&t=90

[astral hinge]

that would contribute more to realizing the game than AO tbh

[cloud rivet]

yeah

I'm not going worry about that now

I like this too https://x.com/Jakob_Wahlberg/status/1974118026263101932

Jakob Wahlberg (@Jakob_Wahlberg)

Magnetic wheels as they should be! Just hold a button!

X · Oct 3, 2025, 2:23 PM

I'll have a track

[astral hinge]

can't tell if this character is a baby or a bald guy

[cloud rivet]

a baby

idk either

[astral hinge]

I wonder how the suction works

[cloud rivet]

I'm not sure, I'm going more in the killer loop single track thing since that'll be easier

[astral hinge]

maybe he's using sdfs to find the nearest surface

[cloud rivet]

I just saw that video today

[astral hinge]

or just shooting rays in a sphere

[cloud rivet]

I'm basically making a killer loop clone

idk feels doable

I should be able to get something interesting with very little graphics work I hope, and then I can work on the graphics and do all the graphics learning I want

and then I'm doing that on a game instead of sponza

[cloud rivet]

ezpz

now it's just physics and game play for a bit

[cloud rivet]

So I am going to read the physics in a weekend document and see if it will work for me. If not I may just use Jolt but I would prefer to have my own physics code

I don’t need complex physics for this game

I would like it to feel fast and to handle well and for the drops to feel like falling

I am going to buy an xbox controller

I want rumbling support

I am also going to add audio

[tight torrent]

https://learn.microsoft.com/en-us/windows/win32/api/xinput/nf-xinput-xinputsetstate

[cloud rivet]

you know I'm fully just using win32, I should just be using DX12 instead of Vulkan tbh

it's the only cross platform code I have, the vulkan stuff

I'm happy with vulkan though, although I wish I could use PIX

[cloud rivet]

looked through the physics in a weekend books, seems ok, I'm going to refresh my quaternion math and write some orientation code and start adding basic physics

[solid grove]

i have custom behavior but I still use physx for handling the actual collisions

[cloud rivet]

yeah I am not opposed to just hooking up jolt, want to see how far I can get on my own

my vehicle has to hover, so I think I will start with that

[solid grove]

the sequel to F-Zero GX we never got

[cloud rivet]

F-Zero, that's the game I have been trying to remember

thank you

https://www.youtube.com/watch?v=76LGjuOmCGs

I should get a nintendo switch or whatever

[cloud rivet]

I used to have one of the elite ones but it kinda sucked, as the buttons would fall out :/

[solid grove]

they're still charging $50 for that? i think that's what I paid for it 10 years ago

[cloud rivet]

yes the eggs to xbox controller price ratio is collapsing

soon a dozen eggs will cost one xbox controller

so my goal is to get my vehicle to hover, move forward, turn and fall off the edge of the track

that's it

I have no idea how long this will take me

[wraith urchin]

My xbox 360 controllers are still going strong 💪

[cloud rivet]

yeah they are great, I should just have gotten a plain xbox controller

[broken fog]

why not a ds5 tho

it's just better

[cloud rivet]

I don't know what a ds5 is

but I bought this thing already