Fwog and co.

[long robin]

u need both

[shell inlet]

btw fast method to get stable luminance is to write ln of luminance and generate mip chain to 1x1 then do exp on the pixel

this gives you geometric mean which is insensitive to small very bright spots

and vice versa

[long robin]

yeah I did something similar to that when I implemented auto exposure

[shell inlet]

I think though one could get fancier and compute both means and fine tune by lerping between them

[long robin]

except I also used buckets

to make a histogram

[shell inlet]

hows that working

[long robin]

https://github.com/JuanDiegoMontoya/Gengine/blob/master/data/game/Shaders/generate_histogram.cs.glsl

[shell inlet]

I was hoping for a verbal overview

[long robin]

I'm trying to rember

basically the idea is that you take the log of the luminance of each pixel, then adds it to a bucket (each bucket is an atomic int)

there are e.g., 128 buckets

the 0th bucket corresponds to the minimum allowed exposure or something, and the 127th bucket corresponds to the highest exposure

then you have another pass which sums all the buckets together and computes the exposure

you can use a parallel reduction in shared memory to sum the buckets, or you can use a 1x1x1 dispatch (which I do, since it's super cheap anyways)
https://github.com/JuanDiegoMontoya/Gengine/blob/master/data/game/Shaders/calc_exposure.cs.glsl

I think this guy uses a similar approach
https://bruop.github.io/exposure/

[shell inlet]

I am actually reading this rn cuz your explanation doesn't make sense

the 0th and 127th bucket part specifically

[long robin]

I'm trying to remember

ah

[shell inlet]

it's an optimization prolly

[long robin]

in my impl, the user can set a min and max allowed exposure, which corresponds to some luminance values

[shell inlet]

just a compute version of the mip reduction

except you also compute averages per some bucket assigned to a tile

and then reduce this set of averages you got from buckets and stored in a histogram

mip reduction would be more wasteful I guess since it reduces in a 2x2 tile at a time

GPUs can chew more pixels per workgroup

[long robin]

I guess one nice thing about the histogram approach is that you get more control over the influence of each range of luminance values

[shell inlet]

btw one thing I'm kinda struggling to understand with compute shaders is that there is a local size written in the source

[long robin]

I think I spoke with dot about my approach at length when I wrote this last year

[shell inlet]

but the size at which it would operate efficiently varies between hardware

you can even obtain those in vulkan

[long robin]

the subgroup size?

[shell inlet]

yeah

on most pascals it's a 4x8 tile iirc

but turings are bigger

[long robin]

the subgroup size is just 32 on all nv cards

[shell inlet]

is it?

[long robin]

the ideal dimensions to use for image processing is either 4x8 or 8x4, I can't remember

on older AMD cards, the wave size is 64

RDNA+ can switch between wave32 and wave64 depending on SC heuristics and whether a driver engineer has forced a mode for your pipeline

[shell inlet]

https://github.com/NVIDIA/Q2RTX/blob/e7d671d4702fa01f8fbabaed2b4645cf78d93eb3/src/refresh/vkpt/shader/asvgf_gradient_reproject.comp#L49

why would it be the most efficient with this size

[long robin]

it could be a million things

finding the ideal workgroup size is just trial and error

[shell inlet]

awesome, I love it when you can't calculate shit

[long robin]

the use of shared memory here has some implications w.r.t. scheduling

a large workgroup means you can maximally leverage shared memory, but it also means more sync

and some gens might not be able to retire waves that finish early, I can't remember precisely

tl;dr you gotta profile

[shell inlet]

well I guess my assumptions were wrong then

today I learned

[long robin]

there are some good heuristics you can use though

e.g., always use a multiple of 64 threads in a wg

[shell inlet]

it might have been a coincidence that 8x4 performed best for a tracing megakernel

[long robin]

it's possible that a workgroup shape like that can access memory slightly more efficiently

definitely more efficient than 32x1

maybe slightly more efficient than 4x8 due to how textures are laid out in memory

[shell inlet]

I don't think that texture locality has a say in here

if I was rasterizing then maybe

[long robin]

maybe it's better for ray tracing because the fov changes less over the x axis than the y axis, meaning the rays are more coherent

[shell inlet]

ray tracing simply destroys coherence beyond primary rays

[long robin]

well, the fov doesn't change

[shell inlet]

it was a megakernel so there were all the bounces in a loop

wavefront is less brutal

mainly on register pressure though

[long robin]

it rekts bandwidth though

[shell inlet]

ye but it's a beneficial trade it turns out

[long robin]

not always

it depends on what you're doing

[shell inlet]

if you have more than 1 material I anticipate gains

[long robin]

it depends beyond that

[shell inlet]

if you have lots of divergence I anticipate gains

because it orders every aspect of the ray tracing such that there is almost no wasted work

[long robin]

wavefront rt uses much more bandwidth than megakernel rt though

[shell inlet]

that's true

[long robin]

bandwidth is far from free

UE has a feature to enable a wavefront rt-like thing and it's not enabled by default because it's not automatic perf gains

I'm not sure if the feature is public, might be an AMD-only thing

can't find mentions of it by name on google, so I'll refrain from discussing it in further detail

[daring surge]

looks good on my 160p monitor!

[shell inlet]

btw >60hz is cheating

for me rsm is very ghostish

[long robin]

I need to fix the depth rejection heuristic still

[long robin]

it ghosts a lot less for me at 60hz when I lower phi depth to 0.1

🦐le scenes like this work at 16x16 upscaling

[golden schooner]

it looks good still, final output that is

[long robin]

it looks horrible in geometrically complex scenes though

[golden schooner]

ah

[long robin]

watch

[golden schooner]

i only noticed the blocks at the last few secs on the right

[long robin]

unfortunately good filtering cannot create new samples

[golden schooner]

i see

who needs trees

maybe

that means different complexities need different settings?

[long robin]

ye

scene-dependent parameters and whatnot

[golden schooner]

ye

or could there be a general solution for everything?

[long robin]

2x2 upscaling looks okay in this scene

[golden schooner]

besides true PT

[long robin]

a little flickery, but probably solvable with TAA

[long robin]

SVGF also struggles with high complexity scenes because the spatial filters aren't very effective

[golden schooner]

ah

[long robin]

I think I'm basically done implementing stuff for this

now the code just needs to be refactored to be readable

[golden schooner]

: )

then we can finally go to the clustered volume thing

https://tenor.com/view/chris-pratt-rubbing-hands-smile-happy-plan-mischief-gif-17115137

[shell inlet]

yes after tonemapping

[golden schooner]

https://tenor.com/view/dr-strange-spiderman-no-way-home-ill-allow-it-gif-22816215

[shell inlet]

it's not as hard to add as GI and clustered shading but is a good thing to have in a real time renderer

[long robin]

[golden schooner]

phunky

[shell inlet]

was this always a thing

[long robin]

no

[shell inlet]

something broke along the journey?

[long robin]

I'm trying to fix the depth weight by comparing world space positions

I messed up badly

[shell inlet]

iirc they compensate depth with motion vector z

something like this?

d1 depth from previous frame, d2 also depth from previous frame but reprojected and compensated

d reprojected is reprojected depth from previous frame

[long robin]

objects have no motion in my thing so I think it's okay

[shell inlet]

motion vectors include motion of the camera

or rather objects that move relative to it

[long robin]

ye but I'm comparing world space pos in the weight

rather than view space pos

so it should be stable even if the camera moves

[shell inlet]

you're not comparing depth?

[long robin]

I'm trying something new

[shell inlet]

should have went with motion vectors right away tbh

even if there is no motion in the scene

[long robin]

adding motion vectors requires effort

I don't think it would have changed anything since I have the previous viewproj matrix

[shell inlet]

well yeah you normally keep previous transforms but you don't have to since there is no motion, otherwise it's next to none in terms of effort to add

it's a diff between mvp*p and mvp_prev*p

[long robin]

2hard4me

[shell inlet]

you could use procedural motion like vertex-animated foliage with motion vectors

[long robin]

yeah once stuff starts moving I'll be forced to add them

but rn I don't have to

ok I just realized that comparing positions for the weight is dumb

well, it sorta makes sense for reprojection

but for the filters it makes 0 sense

changing your distance to the surface will change the blur width

[long robin]

something unrelated to rsm for once

[heavy cipher]

tf

the frog

[long robin]

you feeling froggy, punk?

[heavy cipher]

i try

[golden schooner]

i like it 🙂

[heavy cipher]

ye p nice, except for the yellow light's thing, thats a bit jarring

[golden schooner]

the cubicle volume thing looks also quite cool

[long robin]

you mean the fog cube?

no one dares to diss the fog cube

[heavy cipher]

oh its wildly unrealistic on porpoise

nvm then

[long robin]

it really do be

  // Fog cube.
  d += 1.0 - smoothstep(0.0, .25, sdBox(wPos - vec3(3., 2., 0.), vec3(0.75)));

[heavy cipher]

or perhaps fog cubes are more common in that part of the US where u be

i have not encountered them yet

[long robin]

https://tenor.com/view/the-fog-is-coming-the-fog-june262022-the-fog-is-consuming-me-funny-gif-25904105

[heavy cipher]

papa bless

[long robin]

can't believe you didn't even question the realism of having a fog frog

a volumetric frog, if you will

[heavy cipher]

thats called faith

[long robin]

faith in frogity restored

[rugged notch]

https://tenor.com/view/swag-gif-25350620

[long robin]

The frog is consuming me

[shell inlet]

damn boi the giraffics 👌

[shell inlet]

you pushed upscaler?

oh yeah you did

deccer should bench it

there might be noticeable gain in perf

[golden schooner]

requires shader fixes 🙂

invalid shader operandisms

[shell inlet]

to division?

you can fix it yourself real quick

[golden schooner]

i dont even know which shader file it is

ah wait

when i hit debug, callback will tell me

[shell inlet]

^

[golden schooner]

got it

ye

10ms unfiltered

2ms filtered

[shell inlet]

is there a slider for upscaling feature?

I haven't tried it yet

[golden schooner]

nope, not in deferred at least

let me check gltf viewer

nope

[shell inlet]

is it a default

probably

[golden schooner]

there is no word about "upsample" in the code

[shell inlet]

there's a word about upscale

[golden schooner]

oh wait

i cant read

[shell inlet]

I already closed git page

would have linked

[golden schooner]

illuminationUpscaled

[long robin]

No

[golden schooner]

i walked upstairs after i typed in the thing, and fell asleep

[long robin]

Frogor

[golden schooner]

: )

[long robin]

I think I just had a revelation from this talk
https://youtu.be/2GYXuM10riw?t=266

maybe

what if I stored incoming radiance + its direction instead of irradiance

then I can sample and denoise radiance at low res and later shade at full res

man, it actually sounds plausible

ah, it's not radiance + direction that they store, but an 8x8 probe per pixel

that makes a lot of sense

[long robin]

using a radiance cache means you can get specular indirect for free™️ even

[long robin]

I wonder what data structure they're using to adaptively place probes like this

maybe it's just N textures of increasing resolution

the technique seems 🦐le until you get to this

[heavy cipher]

but now it is 🆑🇦🇲🇵🇱🇮🇨🅰️🧸

[long robin]

that right there is a modern art masterpiece

[long robin]

I also found this technique which might be interesante (pdf included)
https://www.nvidia.com/en-us/on-demand/session/gtcsj20-s22699/

I just realized that the direction for the radiance probes is only necessary if you want specular

I think I can split the pixel light calculation into two parts then

basically, the radiance part (the light bouncing off the first surface and the distance to the next surface) and the surface part (light bouncing off the second surface to the eye)

then I can denoise radiance, which is probably much easier than denoising illuminance

it should take the normal weight out of the denoising equation

@shell inlet do you have any thoughts on my recent ramblings?

[shell inlet]

I don't understand a thing

is what I'm thinking

denoising radiance sounds especially incomprehensive

maybe you mean irradiance

radiance is like 1 direction only

[long robin]

tl;dr: store incoming light and denoise that

[long robin]

my terminology with light units is bad

but yeah I guess it'd be irradiance since we care about all directions

[shell inlet]

A spectral radiance measurement expresses radiance as a function of wavelength or frequency. In contrast to radiance measurements, irradiance is the radiant flux received by a surface or the flux that is incident on the surface. The irradiance unit is determined as power per unit area.

shit this sounds even less intuitive

[long robin]

ok so radiance probes store radiance coming from several directions

but I only care about the light coming from all directions for diffuse

[shell inlet]

ok when you use cubemap for specular you can call the cubemap radiance map

cuz 1 direction tells you how much radiance is there for it

[long robin]

seems okay

[shell inlet]

you can integrate it to get irradiance

like what deccer did recently

that would be irradiance map

so which one you have

[long robin]

irradiance

I think radiance probes would work great for adding specular to RSM, but it seems extra complex since you need to store "all" the possible incoming directions of light (lumen just uses 8x8 oct encoded probes)

But irradiance seems like a trivial extension of what we currently have

It's like demodulated albedo 2.0 in how it simplifies denoising

[shell inlet]

I still don't get the full picture of what you're describing

radiance probes is like, render scene to a cubemap around some position

with direct illumination or whatever

[long robin]

So the original ComputePixelLight function has two dot products, but only one depends on the normal of the surface we're looking at

[shell inlet]

irradiance probe would mean you already integrate that map also with some brdf

[long robin]

Forget I said the word probe

It's in screen space still

It's just storing illuminance at each pixel

I mean irradiance

light going into the pixel (which represents the surface element) from all directions

The difference is that you put the brdf, which introduces noise, at the end of the pipeline

[shell inlet]

surface element 🤔

[long robin]

ignore that term too

[shell inlet]

okay

[long robin]

I stole these ideas from this talk that I linked earlier
https://youtu.be/2GYXuM10riw?t=266

[shell inlet]

are they only computing irradiance at sparse points in screen space?

and then interpolate?

[long robin]

They're computing radiance and putting them in screen space probes

But radiance is harder because the incoming direction matters for the brdf, so each probe is 64 values instead of 1 value

[shell inlet]

yeah that is basically a mini cubemap, well half of it

and with different mapping also probably

[long robin]

exactly

[shell inlet]

idk

very low quality radiance in there, so no sharp specular

[long robin]

Yeah, I suppose so

[shell inlet]

but also they are using ray tracing and generate rays on a hemisphere

while rsm generates samples on a surface

how are you going to cache the radiance, especially since you don't have visibility, so many points may overlap each other along same direction

[long robin]

well I'm not actually planning on using radiance cuz it seemed too hard

[shell inlet]

like what you saw in the screenshot of my brute force rsm specular

you could see the wall right through the box

because some samples landed on a wall and some on the box

[long robin]

irradiance is a single value on the other hand, and works fine for diffuse as long as you have a shrimple brdf

[shell inlet]

this may be a fitting thing to try for when we make ray tracing in fwog

[long robin]

yeah

when

[shell inlet]

is this a question

[long robin]

I forgot to put the > before it

I like the ambiguity, however

I'm gonna try my irradiance idea tomorrow since it seems really easy to add in my head

I'm less warm on the specular idea after you reminded me that leaking exists

[shell inlet]

I'm just curious how you're gonna cache the radiance

or are you talking about something different?

[long robin]

I'm not going to cache radiance

I'm just gonna move part of the brdf to the end of the pipeline and then comment out the normal weight in the denoiser

[shell inlet]

uhh but the brdf is already constant and is cancelled out

by normalization factor

[long robin]

Ok then I'm talking about something slightly different

The part where you calculate the pixel light

[shell inlet]

it was 2piR^2 but 1/pi brdf made it 2R^2

[long robin]

I'm going to move one of the two dot products

[shell inlet]

you were mentioning the geometry term cos_src*cos_dest/dist^2

[long robin]

ye

apologies for poor use of terminology

[shell inlet]

move where? both cosines depend on ray direction

not on the outgoing direction but incident direction

which is unique per sample

[long robin]

wait shit

yeah I'm tripping hard yo

Everything I said in this conversation

[shell inlet]

lol

yeah happens

[long robin]

We're literally already computing the irradiance

The function is even called that
ComputeIndirectIrradiance

So then the only other thing is radiance probes for specular, but that'll suck because 1) sampling specular RSM sucks, and 2) light leakage

[shell inlet]

sampling surfaces sucks in general

rsm uses surface estimator

I tried to make it directional but it doesn't really work without visibility term

and we can't even properly remap between a direction and a rsm texture to begin with

[long robin]

I guess the only thing that might improve RSM now is the hierarchical adaptive sampling idea in that talk, but I don't even fully understand it

[shell inlet]

my idea was to generate random point on a hemisphere with some distribution, project that to rsm uv and then compute the metric tensor of the transformation to integrate irradiance still using surface form but it's too hard to derive

[long robin]

I remember that

iirc you and criver had a long talk about it with a sad conclusion

[shell inlet]

yep

[long robin]

The sampling is probably perfect then

Ignoring the possibility of creating a structure from the RSM

like in LPVs

[shell inlet]

I think it's possible to use resampling for RSM

but I'm not gonna think about that now

resampling is essentially a numerical way to refine samples to better match integrand

[long robin]

I see

It sounds like you need to take extra samples though

Or do you refine it temporally?

[shell inlet]

well it depends

RIS afaik needs more than one sample but ReSTIR can exchange them spatially

wait maybe RIS can also go with 1

I need to re-read the paper to answer that

it's not a hard thing, it's explained in one paragraph in ReSTIR paper

[long robin]

I'd say don't worry too much about it

We've been polishing this turd for long enough

[golden schooner]

its super smooth now 🙂

[shell inlet]

1 sample works also for RIS

[long robin]

ye, it looks better than I'd ever expect after seeing the og version

[shell inlet]

here have a totally out of context equation

[long robin]

What's f(y)/p(y)

[shell inlet]

integrand over target distribution

[long robin]

And what's w(x_j)?

[shell inlet]

I forgor

I guess the whole thing in parentheses is expected value of w

https://cdn.pharr.org/ReSTIR.pdf

start with section 2 and in 2.1 it's explained

there is one error in their terminology in original restir paper however

[long robin]

I gotta sleep but I'll try to remember to read it tomorrow

[shell inlet]

they're calling some function a pdf while it clearly doesn't satisfy the conditions of being called a pdf

it's more like a weighting function if anything

cuz for resampling you don't need a function that integrates to 1

only some function to correlate your samples against

[long robin]

I'm the kind of person to call some weighting function a pdf

[long robin]

I was tempted to ask you if w here is the pdf

[shell inlet]

set of RIS weights apparently

[shell inlet]

idk where to write this but apparently min() function is also a form of mean

https://www.desmos.com/calculator/2vjmv0snac

here I was trying to see how would each mean behave (we talked about means and autoexposure recently)

kinda seeing the trend of each mean going lower, and turns out there is a generalized mean and minimum is some special case of it

¯_(ツ)_/¯

maths

maximum is also a mean

[golden schooner]

mean as in evil?

[shell inlet]

poe's law doesn't let me decide whether you're legit or joking

so haha and also as in average

[golden schooner]

: )

i wasnt sure about your min is mean thing earlier hence the question, but now i know

[shell inlet]

so anyways maybe harmonic mean is even better than geometric mean for autoexposure

it only involves divisions as opposed to natural log and exp

maybe even faster? Is log and exp slow on GPU?

harmonic is literally reciprocal of a sum of reciprocals times amount of reciprocals

and seems to be affected even less by tiny bright lights

needs to be tested on some real images I guess

[daring surge]

trying to say this outloud makes me suffer

[shell inlet]

clamplicatedd

[long robin]

throwback to last night when I forgot that diffuse illumination is not view-dependent

[golden schooner]

https://tenor.com/view/lion-king-toss-bye-gif-4963915

[shell inlet]

haha wrong

but I am nitpicking

[long robin]

yeah yeah I know a real diffuse brdf has view-dependent parameters

microfacets and shiz

[shell inlet]

no I mean lambert is just constant that's why it's independent to anything

any other brdf you'd find is not a constant

e.g. even oren-nayar

it's not necessarily a microfacet brdf, but it was fitted to real world data to add roughness

I think lambertian may be the only one like that

[long robin]

I will make another

[shell inlet]

yes double lambertian brdf 2/2pi

[long robin]

in terms of fwog, I think copying vulkan for the formats was a bad choice
e.g., stuff like this R8G8B8A8_UNORM

I'd rather write RGBA8_UNORM

memory barriers are also slightly annoying, but I don't see a way to fix that outside of a full-blown dependency graph

maybe I could do something simple

actually nah it would immediately become not simple

[golden schooner]

R8G8B8A8 is fine

d3d does something similar

imaging libraries do RgbaX isms

[long robin]

it's explicit which isn't bad

just kinda difficult to read sometimes lmao

R8G8B8A8 is really bad because every single character has similarities

[heavy cipher]

just make an rpsl backend

[long robin]

idk what that is

[golden schooner]

RedEightGreenEightBlueEightAlphaEight_YouNormalizedIt

[golden schooner]

AMDs new Mantle 2.0 shading langauge

oh meant to reply to jaker 🙂

[heavy cipher]

eh not really

[golden schooner]

i tried to be funny here, emphasis on tried

[shell inlet]

you tried so hard 🎵

and got so far 🎵

[long robin]

https://tenor.com/view/itve1-itve-hl2-half-life-in-the-end-gif-26248969

[shell inlet]

remember someone told that this meme will die shortly after someone suggested the emoji

well guess what, not on our watch

[shell inlet]

what's next for fwog?

[brisk narwhal]

c l o u d s

[shell inlet]

it has volumetric fog already

not very different from c l o u d s

[rugged notch]

non-euclidean rendering

non-photorealistic rendering

non-virtual rendering (going outside)

[shell inlet]

no rendering (miserable life)

[golden schooner]

clustered volume rendering is next

[brisk narwhal]

What would you like to cluster

[golden schooner]

lights 🙂

its on the frog's list

[brisk narwhal]

but that's boring

no fancy VFX

[golden schooner]

maybe jaker adds some SFX next

[shell inlet]

apparently fwog thread has the most messages out of all

[brisk narwhal]

Since I'm at a good spot to get some rest

I'll spend it wisely

By cloning Fwog and adding stuff

Yes this is a good way to spend my time

[shell inlet]

what are you gonna add

[brisk narwhal]

Dunno yet

I'll probably find some interesting paper and experiment

[shell inlet]

atmospheric scattering example?

[brisk narwhal]

Oh that sounds nice

any reference papers you know of

[shell inlet]

nope, never implemented any

[brisk narwhal]

aight, I'll do some research then

[shell inlet]

but I've heard UE one is the best out there

maybe outdated info

https://github.com/sebh/UnrealEngineSkyAtmosphere

this maybe

#graphics-resources message
the only one I found in #graphics-resources

[golden schooner]

that cant be true 🙂

[shell inlet]

wow I missed your thread

my apologies

[golden schooner]

no reason to apologise

[shell inlet]

my not apologies

[golden schooner]

: )

[shell inlet]

checked it out, this is not for real time...

[brisk narwhal]

I'll start with good ol' GPUGems

https://developer.nvidia.com/gpugems/gpugems2/part-ii-shading-lighting-and-shadows/chapter-16-accurate-atmospheric-scattering

[long robin]

I hate how they don't fix the code samples on this site

[shell inlet]

what's more some of them have invisible unicode characters

I remember long long time ago I had fun time due to me copypasting shaders, nvidia does a little trolling

[long robin]

Amazing

[long robin]

ah copying stickers is annoying

[long robin]

was watching a talk when it mentioned this enhanced-performance bilateral filter algo
https://people.csail.mit.edu/sparis/publi/2007/siggraph/Chen_07_Bilateral_Grid.pdf

[shell inlet]

jesus local tonemapping looks cursed

[long robin]

I read a thing about it once, but forgot everything
https://bartwronski.com/2022/02/28/exposure-fusion-local-tonemapping-for-real-time-rendering/

[shell inlet]

btw today I had an idea in my dream

remember that extended reinhard has white point value that tells up to which point to compress the highs

[rugged notch]

literally revealed in a dream, nice

[shell inlet]

what if we used some another average to get that white point

how cursed would it be

[long robin]

I dunno

[shell inlet]

sometimes both averages can be really close though so I guess you'd need to clamp the value

or maybe if the difference between averages is so tiny then the image's contents is not really detailed

the idea surely came from my research on generalized mean aka power mean

[long robin]

I gtg to bed, but I'll be back in 7-8 hours to read the rest

I'll have to study the extended reinhard thing a little too because I just copied your function without reading it carefully

[shell inlet]

I posted graph and my thoughts here^

[shell inlet]

which one, luminance scaling only version?

[long robin]

I think so

[shell inlet]

it's ez, same reinhard except multiply each channel by same factor

[long robin]

ok, so you're talking about something else now

[shell inlet]

1 float value of luminance instead of RGB

[long robin]

ye, shrimple enough

aight gotta dip

[shell inlet]

and the luminance you get by dotting with a vector that basically contains how sensitive human eyes are to color components in terms of luminance

[shell inlet]

https://www.desmos.com/calculator/lwxfio2zsk

here updated the thing and included power mean

can use a really high power to get upper white point that is not absolute maximum to still have some overblown parts, plus tiny bright lights will definitely ruin stability with maximum

[long robin]

this means a lot

[long robin]

is this for finding the mean luminance for auto exposure?

[shell inlet]

one could be fore the luminance, another for the white point

so that both are adaptive, not just exposure but also tonemapper

like you'd use geometric or harmonic to find average luminance to use for autoexposure

and cubic or quadratic (or arbitrary power) for white point

since those are above arithmetic mean but still below maximum

so you'll have some blown out parts that will make bloom shine

if you take maximum as your white then reinhard will compress hard

and won't have any noticeable bloom

this is again just an idea to try out btw it's not from somewhere

I'd like to do it myself but I am in college jail

[long robin]

I see

[shell inlet]

are you gonna make tonemapper in fwog?

if you add basic one with autoexposure I might fork and experiment with it

working on something from scratch would be too time consuming for me

[long robin]

I might add it as a new example or maybe cram that into the gltf viewer as well

But don't hold your breath

I saw that they use tricubic filtering for the fog volume in ghost of Tsushima

I might copy that

[shell inlet]

tricubic filtering 🤔

[long robin]

bicubic but 3d

[shell inlet]

is it like... ok

this shouldn't be hard to add I think

but you have SDFs tho

[long robin]

yeah, just one function

just gotta apply the tricubic in the final pass where you composite the fog into the scene

[shell inlet]

you're raymarching via sphere tracing I guess?

[long robin]

it's marching through every cell in a 3D texture

[shell inlet]

wait hows that working

I mean I know how marching thru cell working

[shell inlet]

talking about this

[long robin]

the in scattering and transmittance are calculated and stored in a 3D texture which is trivial to apply to geometry on the screen

you just convert world space to 3D texture space and sample

[shell inlet]

oh sure yeah

3d blur kinda?

[long robin]

kinda

[shell inlet]

well filtering

blur would be different

[long robin]

Trilinear filtering has obvious artifacts, but tricubic does not

[shell inlet]

it still had visible aliasing with shadows

but yeah better than linear

[long robin]

37 minutes in
https://youtu.be/GOee6lcEbWg

[shell inlet]

pretty cool

from what they shown it is indeed magical solution

[long robin]

The whole talk is ridiculously packed with info

At the end they talk about tone mapping and exposure

[shell inlet]

I am now gonna watch tonemapping section

[golden schooner]

should go into #graphics-resources at some point 🙂

[long robin]

I think it has too much info lol

A lot of it went over my head

It is interesting though. It covers a lot of topics

I guess what I'm trying to say is that its information is not super accessible

[shell inlet]

https://advances.realtimerendering.com/s2021/jpatry_advances2021/index.html

here's the presentation

with links and all

funny they also use desmos

oh hey they use white balance

pretty cool

didn't know it's actually used in games

their white balance doesn't seem to be balancing the whites tho

it's "artistic"

[golden schooner]

Cube Map Shadow Tracing 🙂

[shell inlet]

maybe you should start some project on top of fwog where we could put all advanced stuff in

@long robin

[golden schooner]

like an air plane dog fight game, where people fly around in sponza and moana island

[shell inlet]

nah just a renderer would do imo

a game will take time away from graphics

jaker been saying fwog examples are getting big and suggesting me to create a project but I already have a project

[golden schooner]

ye

[shell inlet]

a compromise would be a separate renderer

[golden schooner]

a renderer for different situations

one for your indoor only game, one for your mmo openworld one, one for your side scrolling pure 2d sprites

your naive deferred renderer, and something more complicated with all sorts of effects, for "true" physical goodness

to show off techniques and how to use fwog

or something like serious-sam tech demo world, which would show off different effects in various rooms

[shell inlet]

that'd be quite a task

sprite renderer shouldn't be hard though

[golden schooner]

i mean its something else for a change

and it doesnt have to be implemented in 2hrs, all those things 🙂

[long robin]

Making a separate project with all these advanced renderisms seems like a decent idea

[shell inlet]

call it something based on toad

to get frog and toad

[long robin]

wtf they actually embedded a desmos graph into the presentation

[shell inlet]

yea lmao

can go to desmos page too if you click logo in the left right corner and play with it there

[long robin]

it would be cool to load HL2 maps (or Source maps in general) and view them in a newer renderer

idk how complicated .bsp files are though

might be feasible to load .vmf files instead, idk

ugh writing yet another loader would suck though

maybe there's a .bsp -> model format converter out there (then I can just convert the model to gltf in blender)

https://github.com/luckydye/vsource-decompiler

this seems clamplicated
https://developer.valvesoftware.com/wiki/Source_BSP_File_Format

seems like most point lights get completely baked into the lightmap and have no other representation in the bsp

I'd rather just load gltfs and forget about these oldisms

[golden schooner]

but half-life architecture is super super boring

ugly textures in low res

harsh transitions between faces

[long robin]

use AI upscaler

I mean what can you expect from a 2004 game that's trying to look realistic

[golden schooner]

elderscrolls stuff is probably cooler

and the game is practically still 1998 😄 from the looks of it

[long robin]

the thing with hl2 is that it's obvious how it can look better since it's attempting realism

[golden schooner]

: )

[long robin]

so you "just" have to add modern lighting

and textures ofc :))

[golden schooner]

i shtronkly disagree

[long robin]

the models would probably have to be remade too

or you can just abuse tessellation

[golden schooner]

but would be cool nonetheless, HL3 renderer made in fwog

[long robin]

I think HL2 with the latest PBRisms and GIisms would look cool

[golden schooner]

you might be able to f ind some hl inspired models on sketchfab, or even exports... hmm i wonder 😄

https://sketchfab.com/search?q=half+life&type=models 😄

https://sketchfab.com/3d-models/half-life-frenzy-multiplayer-map-377670023195446aa8eeb9c1e4a56f70

would be cool to run around in these as well not just fly

[long robin]

could add some physics library to the new renderer

jolt seems shrimple to add compared to physx

physx makes me want to die whenever I integrate it

could never figure out the cmake incantation required to summon it, so I ended up building it and copy-pasting the dlls/libs/headers myself

[golden schooner]

: )

maybe i could invest some braincells into the cmake-template-thing of mine, and see how physx could be integrated

[long robin]

someone posted a cmake snippet once, but I forgor to save it

[golden schooner]

: )

you wanted to try jolt anyway

[long robin]

ideed

ineded

indede

indubitably 👍

[golden schooner]

indubiousedly

[rugged notch]

Undeviously

[brisk narwhal]

unequivocally

[golden schooner]

unschleppingly

also time to schlepp my ass over into my cave to work on the things i wanted to work on

[shell inlet]

just wait until rtx remix is out

people already mod the shit out of portal to get hl2 and stuff running on top of it

[daring surge]

ask sultim 🤠

[golden schooner]

indeed

[shell inlet]

if someone can manage to find hl2 sources where you can rewrite graphics code then maybe but I talked with sultim and neither him nor me could manage to do so

it probably requires premission to access a private repo or something

hl1 has an open source port of goldsrc called xash3d which is what he's using to get rt into hl1

[golden schooner]

https://www.youtube.com/watch?v=Uea9Wq1XdA4 could you implement something like this with plain opengl fwog

[shell inlet]

ah yes gibs I mentioned it before

not with rsm

needs ray tracing

[golden schooner]

ah then it was here where i read "surfels" in the corner of my eyes recently

i also like the rocky scene in the background of the thumbnail

[shell inlet]

on second thought you can

but radiance caching part will be very suboptimal

it's always about optimal sampling when it comes to irradiance

surface integral that rsm uses is the biggest problem

[golden schooner]

muh better than

[shell inlet]

I don't think so, not without visibility term

[golden schooner]

you can throw RTX on as much as you want at HL2 it will never look better than this

compared to the rocky thing in the background of the thumbnail

[shell inlet]

it's already ray traced

except it's baked

and only 1-2 indirect bounces

not sure how many exactly but not much

probably just 1 cuz there's no saturation

more than 1 bounces give saturation and you actually can see it in source 2 lightmaps

https://cdn.discordapp.com/attachments/707034102343139370/855428933993889792/unknown.png
https://cdn.discordapp.com/attachments/707034102343139370/855428972823969793/unknown.png
https://cdn.discordapp.com/attachments/707034102343139370/855429004172984351/unknown.png
https://cdn.discordapp.com/attachments/707034102343139370/855429028348428339/unknown.png
https://cdn.discordapp.com/attachments/707034102343139370/855429051170029568/unknown.png

https://cdn.discordapp.com/attachments/707034102343139370/855429076537180220/unknown.png
https://cdn.discordapp.com/attachments/707034102343139370/855429102344732672/unknown.png

0 to 5 bounces, last one is 128

you can see saturation increase with each new bounce

[golden schooner]

i see

[shell inlet]

also the reason why multiple scattering in brdfs give saturation

[golden schooner]

but i was also just refering to the uglyness of source worlds 🙂

the last frame is quite "red"

compared to the previous ones

[shell inlet]

it's red and beautiful

unfortunately not suitable for real time

sometimes I think that irl is surprisingly shadeless and flat

and old games with primitive shading looked way more realistic on old crts than games from 200X

but maybe it's increase in resolution that makes stuff look more obviously artificial

[shell inlet]

funny thing I thought about

what if during rsm indirect estimation you check if sample is on screen and if it is you add its contribution from previous frame

feedback loop from screen space for multiple bounces

[shell inlet]

wait that's not gonna work if we sample in rsm space, what if we sample in screen space though, disregarding visibility (again)

but that's like extra cursed probably

so like 1 rsm sample and 1 sample in screen space from previous frame

[long robin]

Hmm

Sounds orthogonal to the RSM sampling, so more like ssgi

[long robin]

Maybe I misunderstood this line though

[shell inlet]

ye ssgi

it is orthogonal

except ssgi doesn't necessarily use feedback loop

[long robin]

I see

[long robin]

fun fact: teardown does not have contact-hardening shadows, but you can trivially add them with the glsl editor in renderdoc

[heavy cipher]

#extemsiom GL_CONTACT_SHADOWS : enable

[long robin]

I'll pay you a dollar if you write an extension spec for that

a zimbabwean dollar (rare and exclusive currency)

[shell inlet]

how?

by distributing samples?

[long robin]

Yeah

all the code for generating rays and traversal is right there

[shell inlet]

so someone mentioned pcss maybe worth trying out first

before you do the other one

https://developer.download.nvidia.com/shaderlibrary/docs/shadow_PCSS.pdf

[long robin]

also https://www.gamedev.net/tutorials/programming/graphics/contact-hardening-soft-shadows-made-fast-r4906/

the author of that gamedev.net article does not know about bilateral filters

nvm, they do, but for some reason only mention it in the checkerboard upscaling part

[long robin]

hmm, is there a 🦐le way to add an email to all my previous commits on github?

it probably falls under changing the author of those commits

gh kindly provides this script that I'm afraid to run
https://web.archive.org/web/20200823163529/https://docs.github.com/en/github/using-git/changing-author-info#changing-the-git-history-of-your-repository-using-a-script

[heavy cipher]

why do you want to do this? pretty drastic

[long robin]

my contributions chart be looking kinda sparse doe

literally the only reason btw

[heavy cipher]

here

[long robin]

bruh reddit silver's value has depreciated, now I need like ten of those to feel sufficiently validated

[heavy cipher]

inflation

[golden schooner]

@long robin clone your repo into a new folder, run the thing there, check if it werked, do the same on the real one, push force

[golden schooner]

did it work?

[long robin]

gonna try it tomorrow 😄

12/22/2022 4:04

(sry for burger format)

[golden schooner]

heh

isnt steam sale starting today again

[long robin]

https://tenor.com/view/gabe-collector-steam-wallet-discounts-gabe-newell-gif-17620770

the gabening has arrived

[long robin]

I was wondering why my program was immediately closing

  auto appInfo = Application::CreateInfo{.name = "Volumetric Fog"};
  auto app = VolumetricApplication(appInfo, filename, scale, binary);

  return 0;

I frogor app.Run();

anyways, the volumetric ex🦐le has been refactored now

[heavy cipher]

exshrimple

[long robin]

shrimple as dat

when the goblin gas kicks in

[golden schooner]

i like it

[golden schooner]

fog color needs color

[long robin]

it could have color

AKA a different albedo

[golden schooner]

a different odebla

a tnereffid odebla

[rugged notch]

Spooky froge

Did you see that I made an sdf for btw

[long robin]

I'm going to bed rn, but if you pull you should be able to change the shader that controls it. Here's the line
https://github.com/JuanDiegoMontoya/Fwog/blob/examples-refactor/example/shaders/volumetric/CellLightingAndDensity.comp.glsl#L152

Link plox

[golden schooner]

we talked about it back when but jaker ignored it for some reason

[long robin]

I must've been on mobile if I didn't click it

[golden schooner]

good excuse

[long robin]

Ah btw the frog in the scene uses a different fog color

The fogfrog

[golden schooner]

allez dodo, we bikeshed this later

[shell inlet]

what the... is this a rainbow

[golden schooner]

hehe thats cool

[shell inlet]

also somehow fog doesn't seem to absorb (and/or out-scatter?) light

I'd expect some darkening both due to out-scattering and absorption

but maybe there is some of it and the in-scattering seems to be the greatest contributor to the fog's appearance

maybe I'm wrong, the maths needs to be checked to make sure, but I'm not going to do that

[long robin]

Yeah, the sun light doesn't get absorbed at all

But the only way to add that would be to make it an approximation

Like assuming it traveled through a constant amount of fog or something

[long robin]

It's a fogbow

[golden schooner]

https://tenor.com/view/frog-umbrella-stand-pardon-me-oh-im-sorry-gif-17124395

thats a frog bow, my frog

[long robin]

I can try experimenting with hardcoded optical depth for the sun light or somehow estimating it

I could make a separate scattering volume for the sun, but that means doing everything twice basically

[shell inlet]

honestly not sure how sun is different than any other light source

must be something implementation related, because in rt it makes no difference what your light source is and where it is, it's all just radiance and visibility

[long robin]

the problem is that it's harder to approximate scattering for the sun compared to local lights

for local lights, I just presume that the fog density between the light and the point in the volume is constant

[shell inlet]

but you said...

you can do variable density

[long robin]

yes

by "constant", I actually mean based on the density of the point

the math is probably completely sound
https://github.com/JuanDiegoMontoya/Fwog/blob/examples-refactor/example/shaders/volumetric/CellLightingAndDensity.comp.glsl#L78

anyways, that approximation is okay for local lights because the participating media density probably won't change all that much between the point and the light

[shell inlet]

well but this is obviously a cut corner

[long robin]

remember that the density is only stored in a screen-aligned froxel grid

tracing a ray from a point to the sun or light won't work if the light source if off-screen

[shell inlet]

this is view space?

[long robin]

what is?

[shell inlet]

the technique, if it's aligned to the screen

as far as I understand you mean there is no info outside of the view about fog density

[long robin]

yes

[shell inlet]

but you have world space sdfs

why don't you also use world space density

why does it have to be in a grid

[long robin]

I dunno what world space density is

[shell inlet]

well you have a sdf which is a function right

[long robin]

also this isn't only using SDFs for density

[shell inlet]

function of world space position

[long robin]

it can use noise or any other kind of injected density

ye

but not always

[shell inlet]

so why don't you have a density function of world space position

[long robin]

hmm

I could sample some of the density functions in the direction of the light

that would help

[shell inlet]

but not always
in which case it's not?

[long robin]

you could inject artist-made density shapes into the volume, hypothetically speaking

[shell inlet]

you can compose them with the function

by overriding or lerping and whatnot

or adding

[long robin]

you can compose anything with functions

[shell inlet]

ye so I'm still not understanding the idea

missing some pieces maybe

[long robin]

Like if you author an arbitrary 3D shape in a modeling program and want your fog to have that shape

[shell inlet]

are we on the same page?

I mean, this arbitrary 3d shape is still a map

be it a 3d texture or also a continuous function

you can have it in world space

[long robin]

true

[shell inlet]

and access it regardless of what you're looking at

[long robin]

I was thinking you might have an arbitrary number of them. But then again, I'm only using sdfs and noise for the fog so lol

[shell inlet]

so what I'm talking about is

float density_map(vec3 world_pos)
{
  //compose all shapes by adding densities
  return texture(your3dsampler, (vec4(world_pos, 1)*inv_your_shape_matrix).xyz) + density_continuous_function(world_pos);
}

you can then march through this, and you can also do same with sdfs, but it's more finnicky ig

[long robin]

Yeah, I get you

[shell inlet]

alright

[long robin]

I can probably fix it 2nite

[shell inlet]

maybe it's more expensive than the screen space grid approach?

I was also thinking you'd need to check borders of the volumes stored in textures, but maybe you can just set clamp to border and write zeros to borders

[long robin]

It won't change the grid, it'll just be taking extra density samples between the grid point and the light source in the first pass (the shader I linked)

[shell inlet]

(I didn't read the shader)

[long robin]

np

[shell inlet]

(but I looked at it out of consideration)

(and my brain did not process a thing)

[long robin]

It's just calculating density and lighting for each cell in the volume

The stuff we're talking about will affect the lighting

[shell inlet]

I need to do homework, unrelated to the discussion but I was working with SQL and yet again, as usual, I ended up googling "SQL sucks"

instead of googling actual things about it

[long robin]

lmao

I took an SQL class and barely understood a thing

the query syntax made me want to die

[shell inlet]

what makes me die is that it has a million dialects

and different software that uses it, that you have to also learn

[long robin]

I only had to learn how to use MySQL stuff (I think) when I took it

I guess part of the reason my SQL (pun not intended) class sucked is because the prof was just reading off slides in broken English in all the lectures

[shell inlet]

no surprises, education is shrimply broken worldwide

but my situation is unique though, although I can learn at my own pace, I just hate web related stuff

[long robin]

I had web classes for my last 1.5 years (due to beer virus) and those sucked

but at least I could sleep between classes every day

well tbf some of the profs made a decent effort to make their virtual classes interactive/tolerable

I was in a class with my friends and we would troll each other by writing obnoxious text-to-speech messages in discord while they were speaking in class

[shell inlet]

lol

[long robin]

fun fact: there is no way to turn down the volume of TTS in discord

[shell inlet]

damn tts is disabled here, I was gonna try it out

[golden schooner]

i have tts disabled too

[golden schooner]

that id stuff is some neat stuff

and potti still hasnt noticed

[shell inlet]

any updates?

[long robin]

unfortunately I started playing Valheim again yesterday and can't stop

my to-do list:

• fix light scattering
• run that one git script

[long robin]

something

[golden schooner]

: O

looks like a cotton ball of sorts 🙂

[long robin]

I added extinction for sunlight

By marching from the sun to the point and accumulating density

I tried doing it for local lights and it's way too expensive since I have to do a march for every light

[golden schooner]

unless the cloud volume was just a bunch of minecraft sized voxels

[long robin]

I still need to think about the math for in-scattering of the lights, but it shouldn't be too bad

[golden schooner]

i like it

those kind of things i also need later somehow

for local nebulae

or debris fields of sorts

[long robin]

Hehe I wonder if mieplot can do space debris

So you can generate an accurate phase function lut

[golden schooner]

maybe after i added some "physicsisms" we could check 🙂

[long robin]

Driving through fog right now (as a passenger) 😍

[golden schooner]

: )

you have some crazy weather going on there right now

[long robin]

Today the ground was snowy and foggy while it rained

I guess snow+rain=fog

sublimation, clearly

[golden schooner]

😄

TYL so to speak

[long robin]

tyl?

[golden schooner]

Today You Learned

Tyl is the surname of one of my eve online characters

[long robin]

Your subliminal thought surfaced

Subconscious I mean

[golden schooner]

ye

i read it as that

subcouscousse

[long robin]

Subliminally conscious

[golden schooner]

superluminal conscious

[long robin]

I haven't even watched die hard this year

[golden schooner]

xD

[long robin]

my dad made a tree ornament of it. I gotta show you a pic

[golden schooner]

bruce willis' german voice fits him much better than his original

[long robin]

I like how the characters just walk away at the end of that movie

[golden schooner]

dont they do that in all of them

but tis a good idea to watch all of them again, minus 5th one, that was crap

[long robin]

Idk I only saw the og one

[golden schooner]

WAT

[long robin]

The first one

[golden schooner]

i heard you

but just the first??

[long robin]

what did I do wrong

[golden schooner]

heh i thought you did a marathon, watching all 4

[long robin]

nein

[golden schooner]

ungeheuerlich

[long robin]

Are 2, 3, and 4 good?

[golden schooner]

3 is my fav

2 is like 1, just at the airport

4 is a little bit modern

but same principle, takes off shirt, vest is white at the beginning, movie ends, vest is pitch black 😄

[long robin]

Continuity fail

[golden schooner]

😄

[golden schooner]

since you know your way around in die hard

ill call you Argyle from now on

man the audio is really fucked

when they sit in the limo, it sounds like they talk in an echo-y room and not driving in a car

[golden schooner]

another funny thing with these old 60s/70s/80s/90s movies or series

you ALWAYS hear footsteps, as if they dubbed that for some reason

foley or whatever that is

[shell inlet]

nice, how's it looking on the nightmare now?

[long robin]

once I'm back on my PC I'll show

The effect isn't super obvious when you're in the fog on that scene

[shell inlet]

there should be visible extinction from everything

like if you look at a wall the radiance coming from it should also be attenuated with distance

and there also should be radiance coming from the sky

but I guess I am asking for too much from an approximation

btw nobody stops you from sampling lights stochastically, tho it'll be noisy (tho x2 you can sample more than one per pixel)

[golden schooner]

[long robin]

I'm sampling the lights per frustum cell btw

[long robin]

That should be accounted for already

Since I store accumulated density from the cell to the camera in the grid

I'm kinda accounting for sky radiance with a hacky approximation (basically constant ambient term)

[shell inlet]

leaking my beloved

but ye that's how it be without rt

tbh even with rt there is a lot of leaking if you use biased restir

in realtime apps

I didn't notice any in portal rtx they must be using unbiased restir

or maybe they don't 🤔

can't tell tbh because sometimes in disoccluded areas shadows disappear

like on ultra perf dlss

maybe that's just what disocclusion filter in relax does

they do use relax in rtx remix

[long robin]

Wait how does light leak in RT

[shell inlet]

restir spatial reuse + no visibility check

[long robin]

Hmm I thought it was just reusing a sampling direction or something (I actually know almost nothing about restir though)

[shell inlet]

when you take samples from neighbors and ignore if they are occluded or not for the pixel where you gather them at

[long robin]

That sounds like a spatial blur

[shell inlet]

restir gi reuses path vertices

[long robin]

I see how that could cause a leak

[shell inlet]

ye same path may not be valid for other pixels

so you have to trace a shadow ray to check

but it's obviously expensive

[long robin]

now the frog actually occludes the light somewhat

you can see it under the frog here as well

[shell inlet]

hey that's pretty good already

I wonder what causes fog to look like it's glowing

there is visibility from primary source

and also absorption now

you might not even need in-scattering 😂

wait maybe it's the in-scattering from the sun? 🤔

[long robin]

there is no in-scattering from the sun atm

er

yes there is

[golden schooner]

the volume needs to be a bit more caustic, its too artivisual looking

[long robin]

There are local lights that make it look like it's literally glowing if that's what you mean

[long robin]

You mean the volume is too uniformly dense? Otherwise I'm not sure

[golden schooner]

yep that

[long robin]

It is using simplex noise for density, but it's hard to tell when you're inside it

[golden schooner]

ah

[long robin]

I can try increasing the scale and amplitude to make it obvious

[golden schooner]

schrimplex being too schlepplex

tbh

i didnt even notice any simplex patterns

it looked like a unit cube filled with uniform smog of sorts

[long robin]

There is a unit cube and simplex "ground fog" that fades away as you go up

[golden schooner]

excuse my inaccurate use of terms (t)here

[long robin]

shrimplex

[long robin]

@golden schooner I hope the noise in the fog is more obvious now

[shell inlet]

cool looks like a wind

I like it

[shell inlet]

you should totally leave it animated, or at least optionally so

[long robin]

it was always animated, just hard to tell because the fog density was more uniform

[long robin]

wtf is this magic
https://www.shadertoy.com/view/MtjBWz

1-tap tricubic filtering

[long robin]

also TIL about iChannelResolution in shadertoy. I always used textureSize 😄

[shell inlet]

doesn't look like cubic filtering to me

looks like smoothstep hack

I showed you some time ago

[long robin]

rip

yeah, doesn't seem to actually fix aliasing like I was hoping

[shell inlet]

ye iq also says that in comments lol

[long robin]

that ghost of tsushima talk (I think) mentioned an optimized tricubic function, only requiring 16 taps or so

[shell inlet]

not 4?

[long robin]

maybe it was 4

[shell inlet]

wait no 8

[long robin]

either way, it was fewer than 64

[shell inlet]

like it's 4 in 2d, but add dimension and maybe it's possible to use just 8

https://gist.github.com/Fewes/59d2c831672040452aa77da6eaab2234

8 taps

[long robin]

btw here's the part in the video
https://youtu.be/GOee6lcEbWg?t=2223

he do say 8 taps

tricubic do be looking a lot blurrier

[long robin]

https://www.shadertoy.com/view/cts3Rj

[shell inlet]

informal fork

why not a real fork

[long robin]

I guess the extension I have turns the fork button into fake fork

btw harsh edges seem a lot smoother now