Platinum — Path Tracing in Metal

[winged veldt]

[copper halo]

so it works with links

but not files

ofc

[winged veldt]

i get the preview when writing the message too

but then it doesn't embed

[copper halo]

lmao

[winged veldt]

common discord L

[copper halo]

oh the link embed isn't correct hdr either

it converts to webp lul

[winged veldt]

man fuck webp fr

anyway ima just use libavif fuck discord

[copper halo]

oh avif doesn't support metadata either

fack

but they do support gain maps

[winged veldt]

🤔

does that mean you can't control how hdr gets displayed?

[copper halo]

yea

[winged veldt]

right

[copper halo]

you can't say what it was tonemapped for

so the browser/OS/monitor each get to choose what they do

one after the other

unless one of them disables the tonemapping in the later stages

argh i stacked my monitors and my old monitor looks so oversaturated

i love 'wide' gamut monitors that just stretch the input

and the srgb clamp mode just tints everything red

[winged veldt]

right i'm just reading a bit about formats and i think i might as well give up on getting hdr images to look good on anything other than the exact display they were authored on/for

[copper halo]

or you can export a video with a single frame

[winged veldt]

i mean isn't that kinda what avif is

[copper halo]

idk if av1 supports hdr metadata then lol

hey you can use ultra hdr

jpeg tho

it's pretty nice tho

it's sdr + gain map

so it looks good on sdr, and on hdr

[winged veldt]

does jpeg even support primaries other than 709?

[copper halo]

dunno

[winged veldt]

ah, it can embed icc profiles so yes

[dense jacinth]

what about .raw

[winged veldt]

is that an actual format? or you mean like, just output pixel data

[dense jacinth]

i was kinda joking but yes

[winged veldt]

i mean i can just output exr lol

but i still need colorspace data

[winged veldt]

any reading on how this works?

i assume it's storing the gain map in one of the APPn markers

[copper halo]

https://developer.android.com/media/platform/hdr-image-format

[winged veldt]

seems to be supported by openimageio as well

so now i just need to figure out icc color profiles and jpeg should do what i need

it's so jank though, i hate it

can we please retire fucking jpeg already

[winged veldt]

oh nvm input only

fucking libraries

fine, i'll do it myself

[copper halo]

lol

[copper halo]

if i render in rec2020 and don't tonemap at all i get this

(idk what the screenshot is doing to color)

[winged veldt]

looks about right for an underexposed render, are you clipping to srgb?

(if you aren't, your display probably is)

[timid frigate]

https://jcgt.org/published/0014/01/04/

[copper halo]

why would it

I'm outputting to a rec2020 pq swapchain

[hazy zenith]

So the input colors and textures are in sRGB and you convert them to rec 2020 and render in that colorspace?

[copper halo]

yeah

and that's the same as rendering in sRGB and then converting to rec2020 just before I write the pixel

which makes sense

the multiplier factors out

I don't get a similar effect to this

with the reflection changing

[dense jacinth]

that effect is propably because the light color is the same or just (1.0,1.0,1.0) but the meaning of that has changed

otherwise what I’ve noticed is just slight changes in saturation or brightness, fairly similar to bluescreen’s sponza

if you have more complex light simulation I could very well imagine some effect shifting colors more drastically in BT.2020 in comparison with smaller colorspaces

[winged veldt]

ah

ight then

i assumed srgb display

[winged veldt]

it shouldn't, i might have a bug

also the dragon color is different in mine, i made them saturated to test gamut, i'd have to try the original scene

[dense jacinth]

i know but i just noticed that i didn’t consider light color in my renderer which would be the reason my renders look different in every color space

so might have the same issue here

[winged veldt]

i am converting my emission from srgb, so that shouldn't be happening

[copper halo]

wdym by this

[copper halo]

no I have ascended to HDR master race

[timid frigate]

@winged veldt what mean const 2000.0f here?
and why 2000.0f?
https://github.com/teofum/platinum/blob/975dd47a8884a4e5e4048f5f6c91fa43dcf051a5/src/renderer_pt/renderer_pt.cpp#L974-L976

[copper halo]

did you change your russian roulette luminance calc

[winged veldt]

i did not

shouldn't make a difference though

iirc i'm just using the highest component

[copper halo]

well, that's wrong, for one

[winged veldt]

1000 because it's converting mm (conventional unit for focal length) to meters (unit length for my renderer), 2 because it's converting a diameter to radius

[winged veldt]

is it

iirc what you use for rr doesn't matter

it's unbiased either way

so it won't change the results

[copper halo]

hmmmm

since you're biasing by the same value you roulette with, i guess?

[winged veldt]

ye

[copper halo]

try using real luminance tho

conv to xyz and yoink y

[winged veldt]

fuck no

[copper halo]

why

[winged veldt]

i can try lol

[copper halo]

you already have the code for that i assume

[winged veldt]

but yeah won't make a difference since it's unbiased

[copper halo]

unless you hardcoded rec 709 -> rec 2020 without going through xyz

[dense jacinth]

in my renderer my sun was always just emitting (1,1,1) light regardless of color space, so having a different internal color spaces will obviously make the output change

[copper halo]

oh

[winged veldt]

i go through xyz cpu side and feed the shader a 709->2020 matrix

[dense jacinth]

I have constexpr functions for this so that it automatically does this

(another reason metal > glsl/slang/...)

[copper halo]

bro it's a constant the shader compiler will optimize out

[winged veldt]

remember i have the option to change working space in the ui

it's not hardcoded 2020

[copper halo]

branch on bool lul

or just hardcode to 2020 because that's based

[dense jacinth]

yes

or use templated shaders

[winged veldt]

what's the problem with taking a matrix in the args buffer

[dense jacinth]

template<typename T>
kernel void bar(device T *x) { … }

template [[host_name("bar_int")]] kernel void bar(device int *);
template [[host_name("bar_float")]] kernel void bar(device float *);

do this but for color spaces

(this is just a sample from the metal spec)

[winged veldt]

but why

[dense jacinth]

why not

[copper halo]

oh i should make a col<T> type

[winged veldt]

cause it adds code complexity for no good reason?

[dense jacinth]

how is it more complex

your conditionals for switching color spaces just become if constexpr, and then the few extra lines of template instantiations

[timid frigate]

ok. thanks

[winged veldt]

i don't have any conditionals in the shader right now

[dense jacinth]

ah true

[winged veldt]

i also don't have any hardcoded conversion matrices

they're calculated from the primaries and whitepoint, which gives me the flexibility to define any colorspace i might want

[dense jacinth]

well ok

[winged veldt]

probably not strictly necessary, but it's cleaner imo

[dense jacinth]

I just slapped the whitepoint etc. into an online calculator and then got my matrices

and mightve copied quite a bit from jaker

[copper halo]

🤓

you should render in the monitor's colorspace directly

[winged veldt]

why? there's no real benefit to matching the working and display colorspace afaik

better to render in as wide a colorspace as possible

[copper halo]

how would you tonemap to the display then

[copper halo]

hmmmmmmm how do you do autoexposure for HDR output?

my SDR autoexposure maps the average luminance to middle gray (~0.18)

but i need an output in nits

[winged veldt]

no idea, i don't do ae

but you should be doing that in hdr anyway, no?

so ig it would still be the same

[copper halo]

you should, but that would map to 0.18

what is that in nits

for output

[winged veldt]

it should be half of the sdr range

hdr doesn't mean you should always use the full range, if the entire dynamic range of the image fits in sdr you display sdr

[copper halo]

what is SDR range

[winged veldt]

good question

[copper halo]

lol

[winged veldt]

ask your monitor vendor

[copper halo]

what a monitor does when given an SDR signal doesn't really matter

[winged veldt]

i mean you'd want consistency between sdr and hdr versions of the same image ideally

[copper halo]

hmm

can't really get that

it's why HDR handling is so fucked

SDR isn't consistent in the first place

most monitors just have their own primaries, and ofc brightness can't be controlled at all

I guess that's what the 'paper white' settings in games is

[winged veldt]

yea

i mean as far as brightness goes middle gray is just "half the max sdr luminance of the display, at the current brightness setting"

as for the primaries well just assume the monitor is properly calibrated, if it isn't that's a skill issue

[copper halo]

I guess I make that configurable

[winged veldt]

i mean you can calibrate output for the monitor's primaries ig if you know them

that's kinda what monitor calibration is

[copper halo]

technically EDID tells you

[winged veldt]

edid?

[copper halo]

display info

that the monitor sends

[winged veldt]

ah right

assuming it's correct

[copper halo]

(it never is)

[winged veldt]

[dense jacinth]

[200,200]

[copper halo]

ooooh i can ask windows for the SDR content brightness and set my white point according to that

[copper halo]

https://www.gdcvault.com/play/1023512/Advanced-Graphics-Techniques-Tutorial-Day

[copper halo]

do you have any idea what any of his words mean

because i don't

[winged veldt]

if you read his articles it makes some kind of sense

the last few are totally unhinged though

[copper halo]

hg2dc?

i have read them

i still don't understand shit

[winged veldt]

yeah took me a while and i still don't get some parts

[burnt peak]

he sounds like a weird guy and i want to imagine he was totally normal before getting into color math

I don't mean just that reply...

[copper halo]

color math more like just color

[burnt peak]

"not even once"

[winged veldt]

if you read the articles in order you can tell when he goes off the rails

[copper halo]

oh new rabbithole

ABL-aware and mini-LED-zone aware local tonemapping

[winged veldt]

no thank you

i've been gone for a while cause uni semester started and i've been getting settled in but anyway

getting started with my jpeg lib

i've decided to just write a library so i can use it for other projects as well (and also because there's a distinct lack of simple image i/o libs that don't suck)

[dense jacinth]

fastjpeg??

[winged veldt]

so here's a jpeg

it's basically libjpeg hello world but hey it's working

now i gotta make it work with different input data types (int16, float32 etc) and figure out how icc profiles work so i can do colorspace stuff

then, ultra hdr support with the gain map thingy

oh and also be better at cmake so i can have the thing fetch and build libjpeg-turbo instead of relying on it being installed in the system (which it probably is because it's ubiquitous, but i'd rather not anyway)

[dense jacinth]

and then find a viewer that supports said icc profiles

or is Preview good enough?

[winged veldt]

ye, preview works

so does any web browser, afaik

arc (chromium) does indeed display p3 jpegs, so does safari

[dense jacinth]

ye but what about hdr

i mean even stuff like VLC can't fucking display HDR content on Macs

[winged veldt]

it just works™

i can open ultra hdr "jpegs" just fine in preview

and iirc works in a browser as well

[dense jacinth]

what does "ultra hdr" even mean here?

[winged veldt]

macos is actually like the one os where hdr is sort of sane

[dense jacinth]

well apple was very early on the hdr train

[winged veldt]

it's android's jank format for doing hdr in a regular old jpeg

[dense jacinth]

ah yes ghetto solutions

love it

[winged veldt]

i've only skimmed the spec, but basically you have a container file with two jpegs

the first one is the actual sdr image

second one immediately follows and is a gain map

any software that doesn't support ultra hdr will just read the first, sdr jpeg and display it

however in that jpeg, the app1 marker has some metadata

that basically tells whoever is reading the thing "psst hey this looks like a regular jpeg but actually it's hdr, there's another jpeg with a gain map appended to the end of it"

that software is then supposed to read the gain map and compute the hdr values

[dense jacinth]

just use a normal format what the fuck

[winged veldt]

it's absolute jank but hey it works, degrades gracefully to a sdr jpeg, and is more widely supported than... uhh most hdr formats

[winged veldt]

if this thing i'm making (i'm calling it simple_jpeg) turns out well i might make simple_avif or whatever

[dense jacinth]

i disagree

[dense jacinth]

oh stbi 2.0

[winged veldt]

exr is not good as an output format though

[winged veldt]

but with less suck

afaik stbi does the decoding itself and it's slowww

i'm just wrapping libjpeg and adding colorspace/hdr stuff

[dense jacinth]

ah okay

so not the #bikeshed-😇 NIH mindset

fair

[winged veldt]

not yet

[dense jacinth]

oh what about that other library

wait

[winged veldt]

i mean i want to actually make this thing and have p3/hdr output for my renderer, not bikeshed over the jpeg spec forever

[dense jacinth]

https://github.com/google/wuffs/tree/main

supports png and jpeg under one api

and is quicccc

oh wait nevermind

doesnt include encoders

[winged veldt]

yea

it also has no docs

and the c api is generated from their nih language thing

but yea decoding only iirc

[dense jacinth]

really?

i remember it was quite easy to get started

heavy use of c++ in the api so it has to be good

[winged veldt]

if it has i could not find them

it's a whole ass language they came up with

[hazy zenith]

They actually managed to achieve memory safety^tm

[winged veldt]

fake, it's not written in rust so it's obviously unsafe ||/s||

[timid frigate]

@winged veldt maybe you can be interested in this https://psychopath.io/post/2022_07_24_owen_scrambling_based_dithered_blue_noise_sampling

Blog also contains info about blackmagic luts for color manipulation

https://psychopath.io/post/2022_04_23_blackmagic_design_color_spaces

I found implementation of some from psychopath's owen scrambled sobol
https://github.com/NVIDIA-RTX/RTXPT/blob/main/RTXPT/PathTracer/StatelessSampleGenerators.hlsli

[winged veldt]

got icc profile embeds working, the left image is display p3 and right is srgb (they contain the exact same data not adjusted for color space, so the left one should be more saturated and clip differently on srgb displays)

now working on a minimal icc profile generator so you can write images in any arbitrary colorspace

turns out a minimal icc profile is pretty shrimple, you can basically just define the primaries and transfer function and that's it

[copper halo]

i realized something amazing today

for all the talk about colorspaces i do, my gltf importer doesn't import base color as srgb

[old hollow]

hey mate

you know you're on mac, can you give me your compiled dylib for imguizmo?

I can't figure out how to compile it and of COURSE there's no compiled binaries anywhere online

:C

@winged veldt :)))))

btw I did a cool thing I need to make a thread on community projects for my "engine" 😄

[old hollow]

also do you know why on mac, when going into fullscreen imgui looks sharp, but in a window it's slightly blurry, it's also breaking my fucking picking as the viewport pos I'm clicking doesn't align perfectly with the render texture faaack

like

in fullscreen it works perfectly

in window, it has a y offset so you can click top of an object it won't select but under the bottom where the object isn't it does

like by maybe 2-3 mm on my screen (13" screen)

😦

[harsh sonnet]

might be your metal layer's contentsScale isn't getting updated

[old hollow]

I’m using OpenGL 🤣

Sorry for confusion haha

[harsh sonnet]

https://cdn.discordapp.com/emojis/483582351260254209.webp?size=48&name=Thonk

[old hollow]

It’s OpenGL via openTK

[harsh sonnet]

should probably move the discussion over to #opengl then

[old hollow]

Yeah mainly was asking bluescreen anyway

I know it’s to do with retina dpi

Like I have my Mac scaled to 2560x1200 I think it is

Which makes dpi scale 1

Going to recommended res makes the dpi scale 2 and my entire imgui drawing ends up cut into the top left quarter

Mac is strange like that

[harsh sonnet]

yeah retina layer scaling is a bit jank

[old hollow]

It also sucks for scalability when it comes to their ui, if you run at the legitimate max res

It goes tiny

Like idk I don’t understand if running native res makes it sharper I don’t think it does

Just makes the ui go tiny for everything

Retina is very strange yeah

But yeah I think what I need to do is derive the dpi scale from the openTK window and multiply my gl viewport with it

Btw @winged veldt tell me to go away if you don’t want me to message here 🙂

My work in progress windowing 🙂

[harsh sonnet]

if you can get at the NSWindow that contains your application the relevant value is the backingScaleFactor ivar

which is always either 1.0 or 2.0 afaik

[old hollow]

Hmmmm I can calculate the dpi scaling anyway mate, using Size and Width/Height

Size is the larger of the two

Physical size ig

I just need to find the correct place to stick in the imgui windowing stuff

But I’m pretty sure it just goes in onresized

[winged veldt]

sorry, didn't integrate imguizmo yet—it's pretty high on the todo list though, so i'll probably get around to it after i'm done with the colorspace stuff

i don't know when that'll be, i've put my projects essentially on hold for a bit as life stuff happened

might come back to this next month, or whenever i have the time and regain motivation to work on it

[winged veldt]

probably your dpi scaling is fucked

how are you handling it?

[old hollow]

I’m not currently lol

[winged veldt]

well yeah, you're going to 1x scaling on a display with tiny pixels

[old hollow]

I wasn’t aware of Mac dpi

[winged veldt]

you're not supposed to do that

retina displays are designed for 2x scaling

[old hollow]

Yeah I run my Mac on higher res so like

It’s 1x dpi scale so when it’s that resolution it behaves

I don’t like how big everything is on the suggested res

[winged veldt]

right, you need to use the right scaling for imgui

you can check my code for it

[old hollow]

But I think i just need to add the scale factor for imgui when I do the resize event

[winged veldt]

though iirc i don't handle changing monitors so it might break if you have two displays with different dpi scaling settings

[old hollow]

Yeah I don’t plan to support Mac fully

Not until I pull the plug and use wgpu c#

[winged veldt]

it's pretty straightforward dpi scaling

[old hollow]

Yeah it’s just weird to me because it’s a Mac specific thing

Windows always runs native monitor res so

Unless you like blurry shit

[winged veldt]

no it doesn't

you can set scaling and it defaults to 150% iirc in a lot of laptop displays

[old hollow]

It’s also doubling my res for the fbos when dpi scale goes up

Oh yeah but it doesn’t do this shit lmao

[winged veldt]

windows having major scaling skill issue is a separate thing

[old hollow]

Like what I mean you just don’t need to worry about your window size being different to pixel res on windows usually

I don’t do any dpi shit atm and had no problems with windows but mac it broke as soon as i changed res to test

But I may consider going to recommended res on Mac as it fixes the blurry imgui

[winged veldt]

yeah idk if windows handles it behind the scenes but mac does expect you to handle dpi yourself

[old hollow]

Yeah haha

It’s fun anyway

Also guess what

Metal being the actual api on openTK on Mac has its perks

[winged veldt]

/*
 * Scale fonts for high DPI rendering
 * TODO: should rescale on monitor change
 */
int renderWidth, windowWidth;
SDL_GetRendererOutputSize(m_sdlRenderer, &renderWidth, nullptr);
SDL_GetWindowSize(m_sdlWindow, &windowWidth, nullptr);
m_dpiScaling =
  static_cast<float>(renderWidth) / static_cast<float>(windowWidth);

[old hollow]

I can for the life of me get 10bit output on windows from my engine lol

[winged veldt]

that's what i do with sdl

[old hollow]

Yeah that’s all I need to do and then pipe in to imguicontroller.resized

Just scale gl viewport with the dpi

Basically anyway

Imgui resize in that controller does more stuff

Like ui mouse hit detection has to work with scaling too so

Will see if it fixes it when I pipe in the dpi scale

[winged veldt]

yeah you need to use the scaling factor in multiple places, input is one of them

[old hollow]

Yeah

I’m pretty sure tho

The imguicontroller does it

Simply by calling resize

It should get the pixel width then

[winged veldt]

what's the imguicontroller

[old hollow]

Height etc

Something from the openTK examples

[winged veldt]

ah

[old hollow]

It can do any imgui stuff

Btw when you do imguizmo mac

[winged veldt]

you're on your own with that, i don't even know what opentk is

[old hollow]

Let me know if you need to build the Dylib

Either way I would legit buy you a game on steam if you can give me the Mac lib for gizmo haha

[winged veldt]

i mostly statically link my deps tbh, but i'll share my cmake to do so

[old hollow]

It’s not in any nuget packages

Ah ok heh well if you could that’d be ace

I’ll treat you 🤣

[winged veldt]

i'll let you know when i do

[old hollow]

Thanks mate

[winged veldt]

no need for that lol

[old hollow]

What you been up to lately, this thread disappeared and I had to search metal path tracing to find it because I forgot the name haha

It fell off my discord pinned threads haha

Btw I’m not versed in c++ I only use c# - dynamic linking is when there’s a DLL and static is when the lib is part of your app?

[winged veldt]

yea

[winged veldt]

just life happened, been going through some stuff and also super busy with uni and work so i haven't had time to work on any projects

i do want to get back to this soon, though, there's a lot i want to experiment with

[old hollow]

Yeah, someone posting in showcase has some nice caustics idk if you’d be able to do that real time tho idk

Oh but then you aren’t doing real time anyway yet so

[winged veldt]

"yet" lol

i'm not even planning on it

least, not for this project—i might mess around with realtime rt down the line, but it's not in scope for now

though i do want to try implementing better light transport for caustics etc, starting with bdpt

just needed to get the colorspace crap out of the way first—i actually already have working jpeg output for display p3, so i think rather than adding support for arbitrary color spaces, cleaning it up and making it a library as i originally intended i'm just gonna integrate it into platinum and be done with it

[old hollow]

Ahhhh good idea 🙂 and yeah deffo would be cool to see caustics 🙂

[winged veldt]

yeah bdpt and volumetrics are the two things i most want to try out

[copper halo]

ok but when hdr tonemapping

[burnt peak]

by my back of the napkin calculations, bdpt is at least twice as good as udpt

[winged veldt]

what about tdpt then

[winged veldt]

no

i mean some day

but probably not any time soon

[copper halo]

[winged veldt]

or i might just tackle it now, idk

that was the plan, anyway

i'm just not too motivated to work on hdr when i'm missing a bunch of more important stuff, tbh

[burnt peak]

I'll have to do the math

[burnt peak]

if you like you can just yoink my thing when it's done and then scale it for pq or whatever.

if that's what HDR is in this case... I think, assuming Rec. 2020 working space, it could just be something like

m1 = 0.1593017578125
m2 = 78.84375
c1 = 0.8359375
c2 = 18.8515625
c3 = 18.6875

def pq_inv_eotf(x):
    ym = pow(x, m1);
    return pow((c1 + c2 * ym) / (1.0 + c3 * ym), m2)

rec2020_lin_val = transmittance * some_projector_brightness
pq_val = [pq_inv_eotf(ootf(x) * (paper_white_nits / 10000.)) for x in rec2020_lin_val]```

in theory, projector brightness could just be a, uhh... HDR-ness control knob of sorts, with radiance extremes already reined in

[burnt peak]

or - wait, you're doing apple stuff, right? I guess they have their own display p3 thing?

[winged veldt]

yeah, haven't looked into hdr output

[burnt peak]

wikipedia says dci uses pq, but apple is .... apple, so I have no idea

[winged veldt]

it's probably pq

oh good, a new version of openimageio broke the project

i hate image libs so much

[burnt peak]

errr... I dunno

I've had nothing but trouble with oiio

I guess they use PQ for consumer stuff anyway.

[winged veldt]

yeah it's kinda dogshit

but what image i/o lib isn't

stb is simple but horrendously slow and only really handles jpeg and i think png

[burnt peak]

python's imageio is mostly decent, but I'm guessing that doesn't help you

[winged veldt]

oiio is pretty solid and supports a lot of formats but it's a major pain in the ass to build and/or link hence why i end up relying on it just being installed on the system

[burnt peak]

pain in the ass to use too

I'm not a fan of their docs or API, IIRC

[winged veldt]

but then apparently they make breaking changes on a patch bump and my code no longer compiles against it (3.0.4.0 vs 3.0.1.0)

[winged veldt]

api is kinda jank but fine, i don't think you can do much better while being as generic as oiio is

the docs are kinda ass but at least they fucking exist lmao

stb has some comments in a header file, wuffs has literally nothing

might just deal with libjpeg, libpng and libexr directly, i'm already switching to libjpeg for output to support color spaces

but also i wish writing images to a file would just get out of the fucking way, because it's very boring code to write and has nothing to do with what i actually want to do (rendering)

[burnt peak]

I don't know. I find it very sparse and lacking options where it should be detailed and explicit and very thorough everywhere that idgaf.

[winged veldt]

i just haven't found any better

for c/cpp at least

[burnt peak]

the state of image io in general strikes me as surprisingly raw and immature-looking for how ancient and fundamental it is to just about anything

not the library, but just all of it

[winged veldt]

yeah you'd expect there to be at least one decent lib for it by now

honestly, i might just extend my libjpeg wrapper to input, add support for png and exr, and forget about image io libs

i don't want to spend the next month writing image i/o code, but it might be easier long term

[burnt peak]

like, even in python, which is reputed to be largely plug-and-play for most things, I've concluded on several occasions that the most appropriate image format is a raw binary dump

because there's always something... tiff API is broken, EXR is corrupted on linux, etc, etc, etc

[winged veldt]

ppm is unironically the best image format

[burnt peak]

PIL/Pillow is beyond help

that's like... too far the other way.
API simplified to the point that you get a fisher price image library.

"save this for interwebs and I don't care how"

https://tenor.com/view/king-of-the-hill-jpeg-do-i-look-like-i-know-what-a-jpeg-is-distort-koth-gif-27539179

[winged veldt]

yeah i'm convincing myself writing a libX wrapper is the way to go

[burnt peak]

reading a little more about HDR standards now I would like to reaffirm my position that the whole thing is unhinged and improvisational

[hazy zenith]

Tiff is a horendous format really

[winged veldt]

why

the little i've seen of the spec seems alright

[timid frigate]

What formats you want to use also?

[hazy zenith]

I've briefly contributed to a tiff decoding library and while the spec is nice, it's missing a lot of details. TIFF is one of those "the spec is what the official implementation (libtiff) does".

And libtiff's source code is... one of the worst I've read. Like no wonder the thing gets CVE's almost every year.

It's the first codebase where I've seen the register keyword used. It's that ancient

I don't think the format does anything particularly well.

[winged veldt]

ah, yeah that sucks

if you're gonna have the code be the spec at least make the code good

[winged veldt]

fair enough, it's kind of an ancient format and dng/exr probably do a better job at what it's used for

tbh so are jpeg and png but we're still using them

[winged veldt]

need to read exr (for luts and hdris), png and jpeg (gltf textures), write jpeg (render output) and exr (raw output, lut gen)

[timid frigate]

tinyexr?

For save png I use lodepng. Works slow but good compress

[winged veldt]

i'm just gonna use libexr and libpng directly in my own abstraction

[hazy zenith]

Or you could convert all textures to QOI with a cmdline tool

[timid frigate]

Why not tinyexr?

[dense jacinth]

since tinygltf I am quite scared of “tiny” libs

[timid frigate]

because of it you start to use "fast"

[winged veldt]

because i've had enough of dealing with someone else's shitty library

i'm making my own shitty library instead

[timid frigate]

what is not ok with tinyexr? I used it for read and write and no problems

[dense jacinth]

fastexr???

[timid frigate]

boost

[winged veldt]

fast uhh something

i'm probably just going to write a wrapper around libexr/libpng/libjpeg though, unless they're bad enough i actually go fuck it, i'm doing it myself

but i really hope it doesn't come to that because i don't actually want to write image i/o code

[dense jacinth]

well to be fair I would rather write a wrapper around that thing google made

forgot the name of it

[timid frigate]

wuffs?

[dense jacinth]

yes wuffs

use wuffs as much as possible is what i'd say

and then use other libraries for the things it doesnt support

[timid frigate]

wuffs scared @winged veldt 🤣

[dense jacinth]

🐕

[winged veldt]

it only does decoding, and isn't documented, like at all

so that's a no from me

[dense jacinth]

you need decoding for importing no?

and it is documented I am fairly sure

or use my code as documentation

[winged veldt]

yeah, but i'd rather not have two separate libs for encoding and decoding

[dense jacinth]

what difference does it make

there's gonna be two separate code paths regardless

[winged veldt]

if whatever i use for encoding decodes as well might as well just use that, reduce the amount of dependencies

yeah, i know wuffs is fast, but i'm not decoding png in a performance critical context anyway

[winged veldt]

if it is, i could never find it

honestly the whole thing is just confusing as hell to me, every time i've looked at the repo i just have no idea what's going on

[dense jacinth]

yeah true

I just use it because I don't need exporting, and it is safe and fast

basically my entire graphics programming endeavours boil down to having the fastest possible asset loading times

never got to the graphics part

[winged veldt]

yea, i do need exporting and don't care about safety or speed (within reason, stb is just too slow)

though honestly i'd just use stb despite the slowness if it had the features i need

[winged veldt]

actually idk what you've worked on besides fastgltf but it's great

i wish more libs were like that

[burnt peak]

use it for importing. for exporting, have a popup tell user to photograph screen with phone

[dense jacinth]

I made a DDS library

uh for graphics that's about it

I had a raytracer back in the day which didn't converge correctly and had broken diffuse shading, I had a Vulkan mesh shader renderer that just put triangles on the screen, I had a Metal mesh shader based renderer where I actually managed to get cascaded shadow mapping in, and I had a Metal based mesh shader based renderer with a basic VSM impl that borked the drivers and Xcode

I think that's about all I've managed to achieve 👍

oh and I've done MVK work

it's all just random garbage basically

besides my libs

[hazy zenith]

tinyexr is not great tbh

I use it in my renderer and would like to replace it eventually.

[timid frigate]

What wrong with it?

[hazy zenith]

I had a problem with outputting images in XYZ iirc.

[winged veldt]

WE'RE SO BACK

i got the thing building again, without relying on clion now just plain cmake+make

yeeted oiio, added tinyexr for lut loading

just need to bring in a png lib to bring back texture loading and render export and i can start working on it again

[bronze sedge]

what's that ui?

looks aaaaamazing

[winged veldt]

imgui

a nice font and a bit of custom theming goes a long way

it supports a light theme as well

[bronze sedge]

damn

doesn't look like imgui at all

btw I meant the whole render and everything looks amazing

[winged veldt]

thanks

it doesn't have a huge number of features, but the basics should be solid

[bronze sedge]

do you like argentinan bbq

[winged veldt]

once i get everything working again i want to get volumes in

[winged veldt]

hell yeah

asado is the best

[bronze sedge]

are you in buenes aires?

[winged veldt]

yup

like a third of the country's population

[bronze sedge]

I was thinking of visiting clarolab team

the flight is like 14 hours or something :S

it's crazy though 14 hours and very little jet lag

because the direction is north to south

my brain has a hard time with that

[winged veldt]

yeh, it's a long flight

i'm used to them lol

flight to europe is about the same time

you should visit, it's a cool city

if you have the time travel around the country while you're at it

[bronze sedge]

that sounds nice

[winged veldt]

hdri loading is back in, now supports exr (with tinyexr) and hdr as before with good ol' stb_image

uh oh my glass may have a little nan problem

[bronze sedge]

oh

interesting though

[winged veldt]

so the only thing that changed is the lut loading

and the other materials work fine, so it's gotta be some issue with the 3d luts the glass material uses

[quick zealot]

It's honey for my eyes 🍯

[hazy pelican]

You lowkey really have amazing looking ui

[winged veldt]

i think tinyexr might be ass

it can't cope with single channel exr files, just spits out rgba data so now i have to process it instead of doing a memcpy to a buffer

got it working, we're so back

[winged veldt]

gltf import is back, i now just fall back to stb_image for texture loading because it handles different formats seamlessly (unless it's exr, then it uses tinyexr)

it's slow, but whatever, i'll just multithread asset import later

[winged veldt]

p3 output works 🎉

if your display is srgb those two images should look exactly the same, but if you have a p3 display (and an image viewer that's properly color managed) the walls shold be way more saturated in the left image

[quick zealot]

Suurrelly Discord is an image viewer that's properly color managed

[winged veldt]

actually, yes

[quick zealot]

no waydge

[winged veldt]

it's chromium which handles color profiles fine

[quick zealot]

nice

[winged veldt]

well on macos at least

[quick zealot]

So what's next now that the image IO stuff is sorted out?

[winged veldt]

uhh i still need to fix the lut export i think then i'm done

i mean

not that i'm using it anymore, but i do want it to work

i think ima add support for volumes next

simple beer's law first, but i want to do it in such a way that it supports proper volumetrics as well and not just the ugly hack yart had

which does mean messing around with the renderer architecture a bit, because now i'll need to actually track what medium rays are in

[quick zealot]

Yeah the RayVolumeState is quite a bit of a feature creep in my renderer, you always have to bring this one along in many places because there are many places where you may want to know if you're inside a volume or not, it's a bit annoying

[winged veldt]

yepp

i'm probably just gonna like

make it a global

[quick zealot]

May not even be a bad idea

[winged veldt]

keep it as shrimple as possible

[quick zealot]

I suspect my RayVolumeState is spilled to gmem anyways because of how long its lifetime is

So having it in gmem just simplifies the variable passing a bit and may not even be slower than having it in (spilled) registers

[winged veldt]

hm wdym

i'm not too familiar with memory management on the gpu

it would still be thread local ofc

just a global in the sense that i'm not passing it around as a parameter all over the place

[quick zealot]

yeah so in a global VRAM buffer right?

thread local because each thread wants its own state but by "global" you mean in an SSBO?

[winged veldt]

the fuck is an ssbo

i don't speak opengl

[quick zealot]

shader storage buffer object

just a buffer

like the one you store your materials in

[winged veldt]

oh

[quick zealot]

and geometry indices and whatnot

[winged veldt]

no need for that though

i don't need that data cpu side

i'm talking about literally just a global variable in the shader

thread RayState *rayState;

done

[quick zealot]

Ah so it behaves like a local variable just that it's accessible everywhere?

[winged veldt]

ye

i mean i'm pretty sure msl does let me do that

[quick zealot]

I see

[winged veldt]

if it doesn't i can just pass it around whatever

[quick zealot]

then:

Basically the ray volume state is created at the very beginning of your kernel (where you ray is created)
And you have to keep it around until the very end of the kernel when your ray exits the scene / has bounced the maximum number of times

So the ray volume state exists from beginning to end. That's why it's lifetime is long.

The few cosine terms and contribution terms when evaluating NEE for example don't have a long lifetime because there only there for a few lines of code and when you exit your NEE(), these variables are "freed" and registers allocated for them are also freed

The ray volume state is never freed, from the beginning of the kernel to the end. And big-ish variables / structures that have long lifetimes like that are often spilled by the compiler ime to save on vector registers usage

[winged veldt]

right

yeah in that case it's probably getting yeeted to vram

i honestly don't think about that too much

[quick zealot]

probably yeah

[winged veldt]

how many registers does a gpu have

[quick zealot]

I didn't either until I realize that the sole existence of my full BSDF code halves the performance of my entire rendering if not more

[winged veldt]

wait how

[quick zealot]

depends totally on the GPU. I literally how no idea how that works on Apple

In general it's 65k per SM for NVIDIA/AMD

[winged veldt]

oh that's a completely different ballpark from like cpu registers

[quick zealot]

oh ye

[winged veldt]

are we even talking about the same type of thing

[quick zealot]

yes

[winged veldt]

huh

[quick zealot]

it's just GPUs have a massive ton to allow for ultra fast context switching basically

[winged veldt]

how does that not make the chips chungus

[winged veldt]

also, what was causing this?

[quick zealot]

because the rest of the GPU core are quite small compared to a CPU core, that's why GPU cores are way slower on their own than a CPU core

[quick zealot]

yep hold on, writing two things at the same time, the explanation is cooking in the clipboard

[winged veldt]

multithreaded convo

[quick zealot]

Because the BSDF code is so massive that it requires a ton of registers. So many that tons of registers are yeeted to VRAM.
And I suspect (haven't checked the assembly though) that the registers spilled to VRAM are the ones used in the eval/sample function of the BSDF

This makes the eval/sample functions super slow because they're pretty much operating on VRAM instead of registers lol. That's a bit of a simplified way of saying that but that's the idea

And so the difference between

• a Lambertian BRDF
• my Uber BSDF with all materials using parameters such that only the diffuse lobe is evaluated

is 2x in perf

[winged veldt]

huh

how big is your bsdf

mine is kinda chonk ngl

[quick zealot]

Quite big, it's almost a full OpenPBR impl basically. It's around 2.5k lines I'd say, 8 lobes

[winged veldt]

i've thought about using metal's fancy compile time branching thing to build specialized shaders for each material before rendering

might do that at some point, see what it does for perf

would also eliminate a lot of branching in the shader

[quick zealot]

Yeah I think the best way to go for me at this point is wavefront path tracing + material sorting + specialized shaders such that it only compiles the shader code that is exactly needed for the material being path traced: do not compile the chonky glass lobe / thin-film interferences / double-metal / ... if the material is just specular + diffuse

I think doing this could be really good for registers (and thus perf)

[quick zealot]

Would you have to do sorting too then?

[winged veldt]

yep

i'd have to figure that out lol

[quick zealot]

the thing that worries me a bit with wavefront path tracing is that it doesn't like path guiding 🙁

or path guiding doesn't like wavefront

[normal solar]

I see it

in chrome browser

works as is

why do my photos I send from phone never have hdr previews though

if you could get p3 to show up

[winged veldt]

no idea, discord krill issue i think

hdr images only show in hdr when i zoom in for me

well shit

just tried to render intel sponza and it froze the entire os

this worked before

[normal solar]

at least it's not setup_loader_term_phys_devs: Failed to detect any valid GPUs in the current config that I have

though I do wonder why entire os has to die

[winged veldt]

likely something the gpu driver didn't like

i had full os crashes before when i had like an infinite loop in the shader or something

[copper halo]

have you tried getting a better driver/os

[winged veldt]

shut

[copper halo]

[winged veldt]

ok yeah i'm definitely doing something to lock up the gpu

the question is wtf is it

[copper halo]

https://tenor.com/vWNbnW2wcRl.gif

[winged veldt]

intel sponza rendered fine before and not much changed

well uhh

the old build also crashes

and if i import the gltf instead of loadin gmy old scene it renders fine

what the fuck

[copper halo]

gippy yous

[winged veldt]

gippy moment

sponza now with curtains, ~5 min at 4k samples

i should actually do stuff lol

anyway nice to have this functional again, and exporting in p3 too

[copper halo]

rec2020 smh

[winged veldt]

i technically support 2020 export

do you have a 2020 display

[copper halo]

uh

sorta

[winged veldt]

exactly

[copper halo]

[winged veldt]

[bronze sedge]

[spice frost]

👀

i was planning on moving onto volumes right after doing mis

we can be volume buddies

[keen bear]

https://github.com/teofum/platinum/blob/975dd47a8884a4e5e4048f5f6c91fa43dcf051a5/src/renderer_pt/shaders/kernel.metal#L554 @winged veldt why do you consider the bxdf sample when sampling direct lighting,,

https://github.com/mmp/pbrt-v4/blob/f140d7cba5dc7b941f9346d6b7d1476a05c28c37/src/pbrt/cpu/integrators.cpp#L724

[copper halo]

no point doing NEE if the brdf is perfectly specular?

[quick zealot]

Not doing NEE on emissive materials? What is this biased felony

[winged veldt]

am i not supposed to do that? the specular flag indicates the bsdf is a dirac delta so there's no point doing nee (bsdf pdf will be zero)

[winged veldt]

actually i'm not sure why emissive is in there

i think it used to be a separate lobe with a nonsense pdf, so i skipped nee if i sampled the emissive lobe

but then i merged it into diffuse so it should be fine

[keen bear]

The flag indicates that for a particular sample of bsdf though

Not the entire bxdf

Which is my point

Your bxdf might have both specular and diffuse components

But if you picked specular component when sampling you won't do nee

Even though it would still be beneficial??

Because there's diffuse component

[winged veldt]

yeah that's a good point

[keen bear]

Pbrt snippet I linked shows this btw

They have a flags method on bxdfs

That returns flags of all components basically

[winged veldt]

ye just saw that

interesting

[keen bear]

That contribute to bxdf eval

[winged veldt]

i'll try fixing that later today, cause i think i do have some missing energy from just that type of material

[keen bear]

Read,,

[winged veldt]

thanks for pointing that out btw

[keen bear]

Ye np

[winged veldt]

i probably have countless issues in my mis and bsdf code

[keen bear]

I'm dumb at light transport, I'm soaking things up from pbrt but I'm looking at your code for how do you organize things

Btw you should add programmable materials

It makes everything at least 10x more fun

Sampling is as easy as before tbc but you get more plumbing to piece together

[winged veldt]

ngl i'm also dumb at light transport

my bsdf is probably a mess so feel free to keep pointing out anything wrong/shitty code it helps

[winged veldt]

how's that work

[keen bear]

Alr

[winged veldt]

like blender's node based thing?

[keen bear]

Yeah

[winged veldt]

yeaa i'm skipping that for now

[winged veldt]

@keen bear removed the condition on light sampling, doesn't seem to make any visual difference in the render 🤔

but maybe it'll help convergence on a more complex scene

it's certainly more correct, though

i've changed it to just the bsdf itself not being fully specular (either roughness >0, or metallic+transmission < 1 so it has a diffuse component)

also i clearly have a major firefly problem

i suspect something in my bsdf is bad

[keen bear]

I'd figure it should yeah

[spice frost]

if the light is double sided, i would edit the model to bring the light even closer to the ceiling just to be sure it's not the top of the light that's hard to sample and generating those fireflies

i'm getting fireflies too with your scene despite only having diffuse materials

[winged veldt]

hmm, doesn't change anything, so i don't think it's that

the light is double sided yea, but it's just two tris

so if it gets sampled you can't sample one specific side, if that makes any sense

if a ray happens to hit the top—or sample the light from the ceiling then yeah it'll hit the backside, but that should be no problem

though that does make me think if a sample at grazing angles might be causing fireflies with some insane pdf due to numerical instability

[spice frost]

but it's much more unlikely to happen, no? specially after a few bounces

[winged veldt]

well any ray that hits the ceiling will sample that light

hm

[spice frost]

before vs after moving the light

and yes i still haven't fixed it

[winged veldt]

hm, there may be multiple sources of fireflies

you have way fewer than i do

how many samples is that btw

[spice frost]

the left was 15k i think, but honestly it was getting worse over time

[winged veldt]

changing the light to a sphere should minimize grazing angle nee hits, but does little to reduce the fireflies

[winged veldt]

oh

[spice frost]

you're right it isn't many fireflies

[winged veldt]

lmao i'm doing like 250 samples

15k takes forever

lemme just do a lower res render

uhhh yeah 2.5k and counting, i definitely have waaay more fireflies

i don't have a pure diffuse brdf to test with though

i would hope i don't get fireflies with that cause it's super shrimple

[spice frost]

could also be caustics maybe?

[winged veldt]

don't think so, it's consistent across almost every scene even without shiny materials

[spice frost]

made a new material by mixing metallic with diffuse just for testing and yeah, it's noisier but still not comparable to yours

[winged veldt]

yeah ima try to figure out what's causing all the fireflies

ugh

i wish i could write my shader code in a real language

fuck shading languages

not having any lsp support or even syntax highlight in my editor sucks absolute ass

[spice frost]

true...

time to write shader debug printfs all over the code

[winged veldt]

that's not even the issue, it's the complete lack of syntax highlight

tbf there are lsp implementations for glsl and slang

cause people like actually use those

msl is a pain in the ass to work with because there's zero support outside of xcode

and i'm not using xcode

tbh when kosmickrisp enables vk rt on mac i'm probably rewriting this thing with a vk backend

funny enough clangd_format actually works just fine with msl because it's just c++

but the clangd lang server shits itself with all the custom keywords and doesn't find the include dirs

[spice frost]

so you're basically coding in plain text? that does not sound fun

[winged veldt]

yeahh

i did manage to enable nvim's basic syntax highlight so i can at least tell keywords, comments, and numbers, but that's about it

[keen bear]

honestly your codebase seems small enough that a shrimple text editor, maybe with syntax highlighting + git grep is fine no?

I do struggle slightly without go to definition on larger codebases but these tools also often fall apart on the said larger codebases

like e.g. vscode's go to def in linux (the kernel (the sorse)) often lands me at incorrect definitions so I have resort to git grep again

it works slightly better with mesa but our driver does some cursed macro magic (there's other drivers that do the same thing) that vscode doesn't cope with and so I'm often back to using git grep again

I guess it's just a cee moment (similarly applies to c++ ofc)

[winged veldt]

it's manageable, and at least the autoformatter does work

but without lsp support the syntax is kinda ass

[keen bear]

lol I wish clang-format was actually good and we had ci require that things are formatted

[winged veldt]

what's wrong with clang-format

other than being kinda tedious to configure

[keen bear]

doesn't format designated init well too well imo

but it's a minor complaint

[winged veldt]

this is what nvim basic syntax highlighting looks like

better than plaintext but barely

and no autocomplete

i mean clangd should work on msl with a little trickery

if i could just get it to see the include paths

[keen bear]

I don't get autocomplete at work

well

it works sometimes

the other times it's borked

[winged veldt]

what the fuck

[keen bear]

because of the cursed macro magic

[winged veldt]

oh

[keen bear]

leme show lmao

[winged veldt]

oh no

fwiw xcode's autocomplete is also broken

so the situation is not better using apple's own stuff

[keen bear]

when I worked as an ue monkey vs proper (i.e. not vscode) would also be broken a lot

so I had to use,, you guessed it, git grep

for ue

yes

and ue code is NOT straightforward to navigate

[winged veldt]

[keen bear]

god I love when my sbc randomly reboots for no reason

alright I give up I'll boot into stable kernel

probs something is borked in kernel with my changes

so uhh to give you the first impression

vsc doesn't believe this is a function (but it is)

[winged veldt]

[copper halo]

i'm working on UE rn and clangd works almost perfectly tbh

[keen bear]

autocomplete? yeah sure buddy

[copper halo]

all the UHT macro spam resolves fine too

[winged veldt]

btw do you work on gippy drivers?

[keen bear]

yes

works elsewhere though

[winged veldt]

cool

[keen bear]

vX files are compiled once per gpu arch

with panvk_per_arch and bunch of other things substituted with names identifying the arch

[winged veldt]

only similar issue i ever had at work was certain libs murdering the ts language server

[copper halo]

just need to wait 20 min for reindex when i realize i didn't include a target when genning compile_commands

[keen bear]

I'm not a fan of this approach tbh but the other extreme is doing things like radv does and I feel like that's worse

in drivers using genx approach (like the above), the hw stuff is represented with structs

in drivers like radv which don't have this cursed macro magic things are arguably more tedious, you put things into bags of uint32_ts or w/e by yourself, shift and mask bits around, etc

anyway back into the kernel mines I go,,

[winged veldt]

so either way you don't get typechecking lol

[keen bear]

you do get typechecking with genx approach

it's just that fancy tools fall apart

[winged veldt]

well yea at compile time right

[keen bear]

hopefully eventually things get rustified at least a lil

(and hopefully it won't be taking morbillion years to compile, as many rust codebases seem to tend in that direction)

[copper halo]

it only uses 50 gigs of ram while doing that

and 70% of a 64 core threadripper

[keen bear]

my work laptop has 16G soldered I'm suffering

I got 48G for gf so she can have morbillion layers in krita

but I myself continue suffering

[copper halo]

i had a work laptop and it was too slow so they sent me a workstation

[keen bear]

I compile stuff on the sbc

easier than cross compiling

and the sbc is fast enough for compiling linux and mesa (though not vk validation layers, or vk cts, those are too chungus)

[copper halo]

i need the threadripper and UBA to compile in a sane amount of time

(still takes 20 min when i get a new build)

[keen bear]

the only downside is that it has an nvme and in summer nvme would get so hot in some workloads it would trip the thermal fuse

so like yeah

it would throttle the soc down into oblivion but it would still not help because even fully throttled cpu is fast enough to feed nvme with enough work to blow past whatever is the thermal threshold

[copper halo]

still, your iteration times seem a lot nicer than mine

[keen bear]

in certain workloads

yes

I hate C++

I did some work on ANGLE in the past

it was pain

[copper halo]

my iteration times suck less because of c++ and more because of UE

[keen bear]

well in case of ANGLE it's also down to how ANGLE uses C++ but ime a lot of C++ codebases just tend to have horrid compile times

so unless there's additional knowledge/assumptions, I prefer to avoid working with C++

but yeah UE also has this uhh

the last time I touched UE was in 4.26 times or something

I do recall there was a weird step before actual compile taking like 10-20 seconds and barely utilizing the hw

[copper halo]

UHT?

[keen bear]

I guess so idk

but yeah I did not find UE pleasant to work with

[copper halo]

UHT takes a bit for me on full rebuilds yes

but incremental is fine

[keen bear]

also my workstation at the time was i7 6700 (non-k) iirc

though when I moved to wfh I got 5900x

[copper halo]

ok that explains it

local compiles are just unfeasible no matter what you've got

[keen bear]

5900x was p good tbh

though

I do feel like windows was making an impact on compile times

[copper halo]

ok actually i have to compile the engine 4 times, so uh

might be better if you don't

[keen bear]

I remember git grep being noticably slower on my 5900x with nvme ssd running windows than on my laptop with i5 6300U with sata ssd running fedora

I think I had to tinker with windows guard or w/e it's called to make things not garbage

but ofc it became the fastest when I stopped being an ue monkey, wiped windows and my 5900x became busy running linux instead of windows

[hazy pelican]

Can't you force treesitter to highlight it as a cpp file ?
It's cope but atleast it will be better than this lmao

[keen bear]

@winged veldt have you considered getting rid of LightSample::pos in favor of dist or smth

so pos would be equivalent to hit.pos + LightSample::wi * LightSample::dist

infinite lights like env light could specify dist = inf

so then the code that handles light sampling would do min(ls.dist, max ray dist or w/e)

in my toy program I also decided to factor out Li into its own function because Li will become non-trivial to compute once materials are wired into the lights so it'd be preferrable to only compute Li if the point has been hit

[winged veldt]

hm, what's the advantage here?

over what i have now

[keen bear]

your infinite light code assumes max ray length

of 100 rn

[winged veldt]

oh ye lol

i forgot about that

[keen bear]

that way you could push it onto the whatever is doing light sampling

[winged veldt]

yea i should just use dist=inf

and no limit for max ray length

[keen bear]

do you plan to make emission value be defined by a texture

[keen bear]

if so you'd probs benefit from doing this too

[winged veldt]

yep

hm yart had that did i not port it over to platinum yet

i frogot it's been a while

but yeah i'll make li a function if it gets any more complex

[keen bear]

afaiu you can't have maxt=inf

but you can know the scene bounds

which'd be effectively inf ig

and arguably lights should not be the ones handling that

[winged veldt]

nothing stopping you from just not setting it afaik

at least in metal

[keen bear]

inderesding

[winged veldt]

yeah lights shouldn't have to know the scene bounds

[winged veldt]

i mean ig you'd set it to float_max or something

which is effectively infinite

[keen bear]

I guess yeah

[winged veldt]

yeah i'll keep that in mind it'll be cleaner

rn i'm trying to find what's causing all the fireflies

i want to improve overall render/code quality

[copper halo]

select random parts of the brdf and only render them

[winged veldt]

like disable lobes?

[copper halo]

no straight up disable chunks of the code

just output D, G, blah blah

[winged veldt]

oh

last time i did that i didn't find anything weird

but it's been a while

[quick zealot]

Still have them with just brute force path tracing? If so, this may be the BSDF, otherwise probably something with light sampling?

[quick zealot]

Also @winged veldt , what if for GMoN we could look up the neighbor pixels to get more sets for not more VRAM?

Each pixel does 3 sets

And for computing the median of means, we could look at say the pixel to the right, left, above, below to add 12 more sets

That deviates a bit since neighbor statistics aren't the same as center pixel statistics but with pixels this close, this could work very well?

[keen bear]

@winged veldt when switching to spherical trongle sampling,,

[winged veldt]

uniform solid angle sampling?

[keen bear]

ye

for area lights

[winged veldt]

i don't understand it tbh, brain too smooth

[keen bear]

mood

[winged veldt]

i read the pbrt stuff and it was so much code just to sample a tri

i tried to implement it in yart (my earlier cpu path tracer) and gave up

if you know of a simpler implementation i can look at that would be super helpful tbh

i've since learned a lot of pbrt code is overcomplicated and hard to parse

[keen bear]

that's true

I don't like pbrt's light sampling interface for example

there's too much data in the light sample and it also has different interfaces for infinite and finite lights instead of unifying the two

[winged veldt]

yeah i don't like pbrt much in general, the book is great for theory but the code is way bloated if you're trying to figure out the concepts

[keen bear]

yea

[winged veldt]

hm, wouldn't this blur the image? unless you're considering the neighbors for the confidence function only and doing the mon on the pixel buckets

also that seems closer to that other method for yeeting fireflies that's mentioned in the gmon paper

[bronze sedge]

What is a spherical triangle. Three non-collinear points form a plane

[keen bear]

trongle projected onto a sphere

[bronze sedge]

Oh wikipededia explains

Cool stuff

[winged veldt]

neat, treesitter will happily highlight msl as if it was c++

so even if i can't get clangd to work, i at least get syntax highlighting and formatting

[bronze sedge]

that's sweet

[quick zealot]

I mean using the samples of the neighbors to decide on your MoN at the center pixel
It wouldn't be perfect but probably better quality/VRAM ratio

[winged veldt]

hm

possibly, but vram usage isn't my main concern with gmon

it's the effect it has on caustics and darker areas

[quick zealot]

Ah I see

How many sets do you use? I tend to use only like 5/7

I'd rather have a tiny bit of fireflies (but even with only 5 sets they are extremely mild) left than massive darkening because of using 21 sets or something

I think RWMC darkens a bit less than GMoN but still a little depending on the variance if that's a solution 😅

Or just get a better GI / caustics integrator

[copper halo]

"shrimply git gud" is always the best idea

[winged veldt]

yeah that's the plan atm

debug the bsdf and integrator to see what's causing the fireflies

as for caustics i do eventually want to add bdpt, path guiding and such

[quick zealot]

https://cdn.discordapp.com/attachments/1296777118037966862/1372522611292110940/646434005440397370.webp?ex=6852964a&is=685144ca&hm=c650b8d0cc82a7e8284af1f959387bf12f99ecb78b0652aad771ad55274b62c1&

[quick zealot]

Have you tried that yet?

[winged veldt]

nope, haven't had time for it

[winged veldt]

naive integrator also has firefly issues so yeah most likely the bsdf

i mean good news is, probably not a mis bug

bad news is, probably a bsdf bug

the fireflies look worse in the mis+nee renders but that's mostly because the noise floor for naive is much higher so they're less obvious, but i think they're just as bad

[winged veldt]

here's a set of "control" renders i'll be using to test the firefly stuff

cornell box renders are 10k spp naive and 1k spp mis+nee, veach lights thingy (turned out to be a great scene to expose fireflies) is 20k spp naive and 1k spp mis

[winged veldt]

alright test 1: replace the entire bsdf with a simple lambertian diffuse

absolutely no firefly issues in the naive renders as one would expect (the veach scene is super noisy ofc, it's literally designed to be a bad case for a naive path tracer)

mis renders are showing some fireflies though, especially visible in the veach scene

but also some in the dragons and even the basic cornell box has some brighter pixels in the small box closer to the camera if you look closely

so there are mis issues, i'll go ahead and fix those first

oh now i remember why i had it set to skip nee if the last hit was emissive @keen bear

i suspect what's happening here is ray hits a light -> nee ray samples the same light -> distance is near zero, numerical instability causes the sample to blow up

which i fixed by simply skipping nee if the hit was on an emissive material, idk if there's a better way there probably is

[quick zealot]

But the cosine term here should be 0, literally, very close to at least so it shouldn't firefly?

If anything, enabling MIS usually fixes some fireflies for me with everything Lambertian, probably because the BSDF sampling helps in some places where NEE is unstable

[winged veldt]

yeah, i suspect some numerical instability with near zero / near zero

i gotta eep now, but i'll keep looking into it tomorrow

[quick zealot]

I'll try see if MIS adds fireflies for me or not

that was on the dragon scene?

[winged veldt]

the veach test one

it's not mis vs no mis, btw

i'm always testing mis+nee

i don't have a nee only integrator

[quick zealot]

so what's on the right?

[quick zealot]

here

ah NEE vs no NEE

[winged veldt]

no

right is mis+nee without skipping nee on emissive hits