Fwog and co. | Graphics Programming | Page 3

shell inlet Oct 27, 2022, 6:28 AM

#

wait a sec

#

no

#

ah yes

#

that's right

#

wait no that's not right

long robin Oct 27, 2022, 6:32 AM

#

like a ninja

shell inlet Oct 27, 2022, 6:43 AM

#

I'm not sure how to confirm it's right on paper

#

but this integrates to 1 and is a uniform PDF

#

for a circle with radius R

#

now 1/r does not integrate to 1

#

this does

#

1/r integrates to 2piR

#

but when put into code it's obviously wrong

#

because it blows up

#

wait no it's not I just forgot

#

to remove 2pi from circle mapping weight

#

ah nah it's losing radius invariance

#

that's a head scratcher

shell inlet Oct 27, 2022, 7:44 AM

#

need to ask criver probably

long robin Oct 27, 2022, 7:45 AM

#

he's just gonna tell us to use a better overall method 🐸

shell inlet Oct 27, 2022, 7:45 AM

#

was about to say that

#

I'll try to think about it tomorrow, I need to go sleep now

long robin Oct 27, 2022, 8:43 AM

#

I decided to look at renderdoc's timing numbers a little closely for a laugh

#

it says my frame takes over 10x longer than it does in reality

#

every time I click the button, any individual event's timing can vary by up to 5x

heavy cipher Oct 27, 2022, 8:45 AM

#

pile of dogs on ye

long robin Oct 27, 2022, 8:45 AM

#

https://tenor.com/view/inconceivable-princessbride-impossible-gif-5806265

Tenor

#

at least the relative timings between expensive events are somewhat accurate

shell inlet Oct 27, 2022, 11:03 PM

#

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

Desmos

test

#

I am losing my mind

#

according to this the pdf is 1/r2piR^2

#

I made 3 estimators and 1 integral

#

all to validate this

#

2 estimate in area measure and 1 in directional, and 1 integrates over hemisphere

long robin Oct 27, 2022, 11:07 PM

#

I accidentally checked "label" under the 10000 random numbers

#

my browser froze

shell inlet Oct 27, 2022, 11:07 PM

#

rookie mistake

golden schooner Oct 27, 2022, 11:07 PM

#

tree-trunk-section.svg

#

looks cool somehow

shell inlet Oct 27, 2022, 11:10 PM

#

so the thing to try now is float worldSpaceSampledArea = TWO_PI*rMaxWorld*rMaxWorld;

#

instead of PI

#

TWO_PI

#

this looks fine as far as I can eyeball

long robin Oct 27, 2022, 11:10 PM

#

aight

shell inlet Oct 27, 2022, 11:11 PM

#

although it's no longer sampled area

#

because area of a circle is piR^2

#

you really need cornell box to validate

#

actually nah that's not going to work because the light source is different lol

#

need to validate using same scene with blender

#

some simple one like you have in deferred but with less occlusion screwing it

long robin Oct 27, 2022, 11:15 PM

#

ye

#

when I'm off work I can try that

long robin Oct 27, 2022, 11:39 PM

#

man

#

whenever I watch one of these remote graphics conferences, I am reminded that these people specialize in graphics, not audio

shell inlet Oct 27, 2022, 11:42 PM

#

I specialize in audio

long robin Oct 27, 2022, 11:44 PM

#

watching HPG and these guys all seem to have $5 mic setups

#

coupled with the fact that I'm listening through monitor speakers KEKW

shell inlet Oct 27, 2022, 11:44 PM

#

what are you watching

long robin Oct 27, 2022, 11:45 PM

#

https://youtu.be/au4cPLuEpNM

shell inlet Oct 27, 2022, 11:46 PM

#

hey that's the guy cem yuksel

long robin Oct 27, 2022, 11:46 PM

#

ok actually a couple of them have decent mic setups

shell inlet Oct 27, 2022, 11:47 PM

#

the guy involved in making restir pt (GRIS) and stochastic lightcuts papers

#

lmao clapping

shell inlet Oct 28, 2022, 12:03 AM

#

bro listening to daqi lin is hard 💀

#

accent much

#

SVGF motion blur is nice though

shell inlet Oct 28, 2022, 12:33 AM

#

pi vs two pi

golden schooner Oct 28, 2022, 12:41 AM

#

2nd one looks acceptable

long robin Oct 28, 2022, 12:44 AM

#

shell inlet Oct 28, 2022, 12:47 AM

#

https://tenor.com/view/youre-welcome-gif-21319408

Tenor

#

I consider what I've put in desmos enough evidence and I'm done defending my implementation, I rest my case

long robin Oct 28, 2022, 12:49 AM

#

Your paycheck will be in the mail tomorrow

dire badge Oct 28, 2022, 5:09 AM

#

🥧 vs. 🥧🥧

long robin Oct 28, 2022, 5:40 AM

#

btw instead of modding the noise after you add them, you can just divide it by two
xi = (xi + noise.xy) / 2;

shell inlet Oct 28, 2022, 5:59 AM

#

what

#

this is not how it works

#

you effectively shrink your random by 2

long robin Oct 28, 2022, 6:04 AM

#

adding two uniformly distributed numbers in [0, 1] gives you one uniform random number in [0, 2]

shell inlet Oct 28, 2022, 6:05 AM

#

for each sample there is now constant blue noise value and [0; 1/2] lds

#

blue noise is meant to do toroidal shift on the lds set

#

so mod is required

#

repent

long robin Oct 28, 2022, 6:07 AM

#

ah

#

I didn't realize it had more meaning than just being an extra source of randomness

shell inlet Oct 28, 2022, 6:15 AM

#

what I want to try now is to distort the distribution even more towards the center

long robin Oct 28, 2022, 6:15 AM

#

I also didn't know what a toroidal shift was and the only useful info I found was the docs for some math function for some R package

Applies a random shift simultaneously to all the points of a point pattern, or to selected sub-patterns, with wraparound at the borders of the window.

shell inlet Oct 28, 2022, 6:16 AM

#

I have no idea what result it will give

shell inlet Oct 28, 2022, 6:16 AM

#

long robin I also didn't know what a toroidal shift was and the only useful info I found wa...

also known as Cranley-Patterson rotation

long robin Oct 28, 2022, 6:17 AM

#

https://psychopath.io/post/2014_06_28_low_discrepancy_sequences

#

very neat

#

I thought the mod was just a way to keep the two random numbers in [0, 1] after adding them 😄

shell inlet Oct 28, 2022, 6:49 AM

#

in case you want to try more distortion on the mapping

vec2 ExponentialCircleMapping(vec2 uv)
{
    //r^x, x=4
    float r = uv.x;
    r = r*r;//2
    r = r*r;//4
    float theta = uv.y * TWO_PI;
    return vec2(r * cos(theta), r * sin(theta));
}

float ExponentialCircleMappingWeight(float r)
{
    //xr^(2x-1), x=4
    float rpow = r*r;//2
    rpow = rpow*r;//4
    rpow = rpow*r*r*r;//7
    return 4.0*rpow;
}

#

gives more noise(boiling from moving the camera) in areas removed from the directly illuminated part

#

also tried uniform and it's just bad, can see boiling when moving the camera all around, even near directly illuminated areas

#

by the way the 1/2pir integrates to 1 over the expected domain (it doesn't integrate to 1 for R!=1 but we don't need that)

#

don't need that because we compensate for it by multiplying whole integral by the world space area of the circle

#

~~I guess technically divide?~~

heavy cipher Oct 28, 2022, 6:57 AM

#

thats 1/2pir * r

shell inlet Oct 28, 2022, 6:57 AM

#

no

heavy cipher Oct 28, 2022, 6:58 AM

#

i mean thats what you wrote in the integral

shell inlet Oct 28, 2022, 6:58 AM

#

r is the surface element

#

1/2pir is what we integrate

#

similarly we would use sin(theta) for integration over sphere

heavy cipher Oct 28, 2022, 7:00 AM

#

if it was intended, then all g

long robin Oct 28, 2022, 7:01 AM

#

actually it looks like void just wrote 1 the long way

shell inlet Oct 28, 2022, 7:04 AM

#

I remember now that we don't really multiply the thing by the area anymore

#

so there should be R^2 in the pdf

#

wait it still doesn't make sense

heavy cipher Oct 28, 2022, 7:07 AM

#

what is the integral calculating?

shell inlet Oct 28, 2022, 7:07 AM

#

validating more like

#

making sure it integrates to 1 because it's a condition for a valid pdf

long robin Oct 28, 2022, 7:27 AM

#

fyi martty is out of the loop

shell inlet Oct 28, 2022, 7:44 AM

#

GP::User martty;
for(auto&& thing : all_things)
{
...
}

long robin Oct 28, 2022, 9:41 AM

#

hmm
https://youtu.be/nnZUYg5iToY

YouTube

Holger Dammertz

Filtering Reflective Shadow Maps

A small experiment of filtering high resolution Reflective Shadow Maps (RSM) for real time global illumination. This reduces flickering when there is high variance in the geometry.

▶ Play video

#

Note also that basing irradiance in terms of the radiant flux allows us to forgo integrating the area of the light as part of our calculations.
that's very interesting. I wonder how the math backing that up looks /s

shell inlet Oct 28, 2022, 4:52 PM

#

this doesn't make any sense

#

shell inlet Oct 28, 2022, 8:49 PM

#

I'm puzzled by uniform estimator diverging

#

probably numerical issues

#

it kind of converges with less samples

#

also I forgot lambertian brdf in the test graph

#

so everything converges to pi

#

should converge to 1

golden schooner Oct 28, 2022, 8:51 PM

#

subtract 2.1415 from it

shell inlet Oct 28, 2022, 8:52 PM

#

almost but divide by 3.141592653589793238462643

golden schooner Oct 28, 2022, 8:52 PM

#

hehe

shell inlet Oct 28, 2022, 8:52 PM

#

that's the lambertian brdf

#

albedo/pi

shell inlet Oct 28, 2022, 10:10 PM

#

ok so I can't figure out the proof of why weights are the way they are

#

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

Desmos

test

#

here's the latest test

#

not much has changed

#

just made it cleaner and added 1/pi for lambertian brdf

#

meaning the 2pi change may be wrong

#

because lambertian brdf cancels it out

#

but best is to compare to reference

long robin Oct 28, 2022, 10:13 PM

#

I see the interactive 3D axes, but I'm not sure what else I'm supposed to see

shell inlet Oct 28, 2022, 10:14 PM

#

maths on the left

long robin Oct 28, 2022, 10:14 PM

#

alright

#

I see the visualized distributions too

shell inlet Oct 28, 2022, 10:15 PM

#

any visualization is borderline useless

long robin Oct 28, 2022, 10:17 PM

#

are the surfaces estimators supposed to integrate to a specific value?

#

ah nvm

#

they're numerical versions of the surface integrals

shell inlet Oct 28, 2022, 10:38 PM

#

integrals are "symbolic" versions (numerical under the hood anyways) and estimators are numerical monte carlo integrals

#

I only kinda trust the directional integral and use it as a reference to validate surface ones

long robin Oct 29, 2022, 4:25 AM

#

I wonder if the RSM paper is technically correct in their usage of flux and not having to integrate over the sampled area, but it just ends up looking bad in practice with the rMax parameter

#

the brainworm won't leave me

#

because the thing is that point lights are already not realistic

#

but idk, it feels like I need to study radiometry for five years to fully understand

long robin Oct 29, 2022, 7:33 AM

#

it would be nice to be able to reflect certain info from shaders

#

particularly compute shader work group sizes

#

kinda lame having these magic numbers scattered around the code
const int localSize = 8;

long robin Oct 29, 2022, 8:27 AM

#

I pushed my imgui stuff and void's RSM fixes to examples-refactor
https://github.com/JuanDiegoMontoya/Fwog/tree/examples-refactor

long robin Oct 29, 2022, 8:47 AM

#

also, I wonder what a good way to bind whole buffers would look like (since that is what you do most of the time)
writing someBuffer.Size() gets annoying, e.g.,
Fwog::Cmd::BindUniformBuffer(0, globalUniformsBuffer, 0, globalUniformsBuffer.Size());

#

guess I could use a default parameter which is a constant like constexpr uint64_t WHOLE_BUFFER = -1;

#

also I noticed that Buffer::Size() returns size_t, but all the functions that take a buffer size use uint64_t

#

that probably won't end well on 32-bit platforms

golden schooner Oct 29, 2022, 8:56 AM

#

is that rsm paper the only one?

#

or are there other independent ones?

long robin Oct 29, 2022, 8:57 AM

#

it's the only one that I'm aware of

golden schooner Oct 29, 2022, 9:00 AM

#

i see

long robin Oct 29, 2022, 9:01 AM

#

there are papers that build on it though

golden schooner Oct 29, 2022, 9:02 AM

#

and they all have the same outcome i guess

#

but none of them fight with the math thing you 2 do

long robin Oct 29, 2022, 9:02 AM

#

I'm guessing no one questioned the math or something, idk

golden schooner Oct 29, 2022, 9:03 AM

#

yeah

long robin Oct 29, 2022, 9:03 AM

#

I suspect the math is actually correct for what they claim it does, but just ends up with an ugly and undesirable result

#

but at the same time, I can't confirm it so it's just pure speculation KEKW

golden schooner Oct 29, 2022, 9:16 AM

#

: >

long robin Oct 29, 2022, 10:29 AM

#

@golden schooner

golden schooner Oct 29, 2022, 10:30 AM

#

ill take a look later

long robin Oct 29, 2022, 10:55 AM

#

hmm^2
https://people.mpi-inf.mpg.de/~ritschel/Papers/ClusteredVisibility.pdf

#

oh jeez it looks complex nvm

#

anyways, just gonna save this list for later
https://extremeistan.wordpress.com/2014/05/11/realtime-global-illumination-techniques-collection/

golden schooner Oct 29, 2022, 10:57 AM

#

peopl should get punished for making these thin kind of blogs, layout wise

long robin Oct 29, 2022, 10:58 AM

#

punished by being forced to read a blog with an even thinner layout

#

also here's another thing
https://people.mpi-inf.mpg.de/~ritschel/Papers/GISTAR.pdf

golden schooner Oct 29, 2022, 11:01 AM

#

Original RSM: https://web.archive.org/web/20160327214626/http://www.vis.uni-stuttgart.de:80/~dachsbcn/download/rsm.pdf
Splatting Indirect Illumination: https://web.archive.org/web/20110907135823/http://www.vis.uni-stuttgart.de/~dachsbcn/download/sii.pdf
MultiRes Splatting https://web.archive.org/web/20140811154504/http://homepage.cs.uiowa.edu/~cwyman/publications/files/multiResSplat4Indirect/multiResolutionSplatting.pdf
Sparse Voxelization https://web.archive.org/web/20220201212921/https://www.seas.upenn.edu/~pcozzi/OpenGLInsights/OpenGLInsights-SparseVoxelization.pdf

long robin Oct 29, 2022, 11:01 AM

#

ah you managed to get that link to work

#

ok I gotta sleep for real now before I get caught up looking at these 😄

golden schooner Oct 29, 2022, 11:04 AM

#

when brave cant resolve a link i get a button top right which says "check older versions" 🙂

long robin Oct 29, 2022, 11:05 AM

#

Oh that's cool

golden schooner Oct 29, 2022, 11:06 AM

#

let me link the other missing ones

shell inlet Oct 29, 2022, 12:30 PM

#

long robin <@194502965749350400>

gotta love how you pinged deccer and I asked you to ping ME

golden schooner Oct 29, 2022, 1:07 PM

#

@shell inlet there you go

#

😄

#

i also had a huge interest in the debug view facilities

shell inlet Oct 29, 2022, 2:10 PM

#

long robin I pushed my imgui stuff and void's RSM fixes to examples-refactor <https://githu...

oh no you did not refactor screenshots

#

on the other hand maybe it's too early for that because of how jank my proofs of correctness are, perhaps there will be more changes later

#

although I've ran out of ideas

#

I really insist on running tests with a production renderer

#

ideally with matching camera and a difference plot

golden schooner Oct 29, 2022, 2:18 PM

#

devsh probably has something in his pocket

shell inlet Oct 29, 2022, 2:19 PM

#

there is a suspicion that 2pi should be just pi because lambertian brdf cancels out one pi but if it turns out that production renderer matches 2pi then it would make me even more confused

golden schooner Oct 29, 2022, 2:19 PM

#

another possiblitlity is that the rsm paper is wrong, author put a typo or whatever

shell inlet Oct 29, 2022, 2:19 PM

#

hard to outright say it's wrong because it misses a lot of details

#

but there are quite a few ass pulls

golden schooner Oct 29, 2022, 2:20 PM

#

turn it into a new project 🙂

shell inlet Oct 29, 2022, 2:20 PM

#

what do you mean?

golden schooner Oct 29, 2022, 2:20 PM

#

creating a reference renderer of sorts, which can be used to test papers against

shell inlet Oct 29, 2022, 2:21 PM

#

I mean it's entirely possible to add ray tracing to fwog

#

I've made opengl ray tracer in the past

golden schooner Oct 29, 2022, 2:21 PM

#

thats not what i was saying but ok fair enough

shell inlet Oct 29, 2022, 2:22 PM

#

this is opengl with nbidia extensions

#

for bindless textures and buffers

#

running on gtx1050

#

I doubt amd opengl could achieve similar performance

#

I tried using buffer textures and shit but it caused more indirection in the shader and it was slow as hell

#

amd only has buffer device address on vulkan

#

because it's core

golden schooner Oct 29, 2022, 2:45 PM

#

bretty cool

shell inlet Oct 29, 2022, 3:01 PM

#

jaker your gltf viewer example has broken filtered version

#

I have no idea why it doesn't work

#

indirect is all black

golden schooner Oct 29, 2022, 3:03 PM

#

it worked for me before

#

without moving any light, in default settings

shell inlet Oct 29, 2022, 3:04 PM

#

also deferred example is missing albedo modulation step

#

and the shader is missing it as well

#

did you break albedo demodulation on purpose???

  return worldSpaceSampledArea * sumC * surfaceAlbedo / rsm.samples;

#

it's going to blur textures now awesome

#

you must have never tried it on a textured model

long robin Oct 29, 2022, 8:26 PM

#

shell inlet gotta love how you pinged deccer and I asked you to ping ME

Sorry, I have dementia

long robin Oct 29, 2022, 8:28 PM

#

shell inlet indirect is all black

Oh it's probably because I removed the final albedo pass

#

I tested with and without that pass and saw 0 difference, so I removed it from the shader

#

Oh frog I see what you mean

#

Yeah I can add that back

shell inlet Oct 29, 2022, 8:31 PM

#

lol what a way to see if code is useful or not

long robin Oct 29, 2022, 8:33 PM

#

I was just testing in the deferred cube scene 🐸

shell inlet Oct 29, 2022, 8:34 PM

#

even though it says "modulate albedo" on the host side

#

and the albedo term is commented out in illumination computation function

#

if it was meaningless I wouldn't bother adding it

long robin Oct 29, 2022, 8:36 PM

#

shell inlet even though it says "modulate albedo" on the host side

Well obviously I didn't understand what that meant 😔

shell inlet Oct 29, 2022, 8:36 PM

#

so are you going to push changes any time soon?

#

I wanted to either ask you to add "skip filter passes" checkbox or adding it myself

long robin Oct 29, 2022, 8:37 PM

#

I haven't looked at your recent changes

shell inlet Oct 29, 2022, 8:37 PM

#

recent changes?

#

wdym

#

what is this about

long robin Oct 29, 2022, 8:38 PM

#

Nvm I misread

#

I can add the checkbox

#

While fixing the albedo modulation

long robin Oct 29, 2022, 10:58 PM

#

@shell inlet is this what you meant with "skip filter passes"? (indirect lighting view)

shell inlet Oct 29, 2022, 10:58 PM

#

unfiltered result yes

long robin Oct 29, 2022, 10:58 PM

#

ok I will push

shell inlet Oct 29, 2022, 10:59 PM

#

I'm assuming you didn't skip albedo modulation pass

long robin Oct 29, 2022, 10:59 PM

#

correct

shell inlet Oct 29, 2022, 10:59 PM

#

nice all good

#

is 16x16 noise alright though?

long robin Oct 29, 2022, 10:59 PM

#

hmm actually I'll add a checkbox for that too

#

it seems okay. I can experiment with other noise

#

2x2 bayer matrix sucked bad when I tested though

#

4x4 bayer seemed worse than blue noise too

shell inlet Oct 29, 2022, 11:01 PM

#

btw "Filter RSM" sounds like it's added on top, I'd name it "Filtered RSM" because it's a technique switch

#

minor nitpick

#

you can also add filter iterations settings

#

it's a loop anyways easy addition

#

though then you can throw away the skip checkbox because you could set iterations to 0 for all filter passes

#

also why not use ImGui::InputInt instead of a slider?

#

and allow more than 20 samples for filtered

long robin Oct 29, 2022, 11:07 PM

#

I pushed btw

long robin Oct 29, 2022, 11:08 PM

#

shell inlet also why not use ImGui::InputInt instead of a slider?

slider is easier to use 🤰

#

btw you can use a slider as an InputInt by ctrl+clicking on it

shell inlet Oct 29, 2022, 11:08 PM

#

but it has a hard limit

golden schooner Oct 29, 2022, 11:08 PM

#

the "light" in the left corner looks odd compared to the other corners

shell inlet Oct 29, 2022, 11:08 PM

#

no input int has +/-

#

you can click those to increment/decrement or hold to speed it up

long robin Oct 29, 2022, 11:08 PM

#

hmm I'll try it

shell inlet Oct 29, 2022, 11:10 PM

#

#

found a screenshot

long robin Oct 29, 2022, 11:10 PM

#

ye I just tried it

#

crashed when I went to 0 samples 😄

shell inlet Oct 29, 2022, 11:11 PM

#

so I'd replace checkbox with filter iterations via inputint and also replace samples with inputint

#

just to allow more customization to see how it does

long robin Oct 29, 2022, 11:12 PM

#

I still kinda prefer the ergonomics of the slider

golden schooner Oct 29, 2022, 11:12 PM

#

add a checkbox to toggle between the two input fidelities 😛

long robin Oct 29, 2022, 11:12 PM

#

oh jeez

golden schooner Oct 29, 2022, 11:12 PM

#

jk

shell inlet Oct 29, 2022, 11:12 PM

#

no add tabs, one tab is everything checkbox, one sliders and one input int

#

also a radio list
1 sample
5 samples
10 samples
15 samples

#

or a dropdown list

golden schooner Oct 29, 2022, 11:13 PM

#

not 1, 2, 4, 8, 16?

long robin Oct 29, 2022, 11:13 PM

#

also a button to randomize samples

shell inlet Oct 29, 2022, 11:14 PM

#

wow we now truly a khronos commitee

golden schooner Oct 29, 2022, 11:14 PM

#

we do be khronossing in other channels too

shell inlet Oct 29, 2022, 11:15 PM

#

it's khronossin

#

time

golden schooner Oct 29, 2022, 11:15 PM

#

heh

shell inlet Oct 29, 2022, 11:16 PM

#

can at least filter iterations be inputint

long robin Oct 29, 2022, 11:16 PM

#

yes

#

do you also want a checkbox to skip the box blur?

shell inlet Oct 29, 2022, 11:17 PM

#

slider is fun and all but if there is a big range it's harder to precisely pick values without typing it in manually

golden schooner Oct 29, 2022, 11:17 PM

#

there is dragInt too

long robin Oct 29, 2022, 11:17 PM

#

I don't expect anyone to want more than 3-4 blur passes

golden schooner Oct 29, 2022, 11:17 PM

#

i believe you can dooble click and enter precise value if need be

long robin Oct 29, 2022, 11:18 PM

#

you can double click/ctrl+click most things to turn them into input ints

shell inlet Oct 29, 2022, 11:18 PM

#

no I think both box and subsampled iterations should have customizable values, let's go over the top with customization

#

let the example inspectors express themselves

#

what if I want to filter everything with a wide box filter and skip subsampled passes

#

just to hear that pleasant coil whine

shell inlet Oct 29, 2022, 11:20 PM

#

long robin you can double click/ctrl+click most things to turn them into input ints

me said:

without typing it in manually

long robin Oct 29, 2022, 11:20 PM

#

ree

#

I'm going to expose a slider for albedo modulation passes too

golden schooner Oct 29, 2022, 11:21 PM

#

re ree: no, void makes fair statements

long robin Oct 29, 2022, 11:21 PM

#

I wish there was a slider with the + and - things at the end

#

maybe I can make an InputInt with just those

shell inlet Oct 29, 2022, 11:22 PM

#

you can copy the slider widget code and try to marry it with input one

#

but that's going to beat the purpose of what imgui is for: less work

#

at least from your perspective

long robin Oct 29, 2022, 11:24 PM

#

I wonder if there are "repeat" buttons in imgui

shell inlet Oct 29, 2022, 11:24 PM

#

repeat?

long robin Oct 29, 2022, 11:25 PM

#

buttons that you can hold down

#

like the + and - buttons for inputint

shell inlet Oct 29, 2022, 11:25 PM

#

there is a state query

#

I guess you can query smth like ImGui::IsItemActive9) or smth

#

need to look at the imgui demo for reference

long robin Oct 29, 2022, 11:25 PM

#

ImGuiItemFlags_ButtonRepeat = 1 << 1, // false // Button() will return true multiple times based on io.KeyRepeatDelay and io.KeyRepeatRate settings.

shell inlet Oct 29, 2022, 11:26 PM

#

for the proper event check

#

ah so they already thought of that

long robin Oct 29, 2022, 11:28 PM

#

lel

shell inlet Oct 29, 2022, 11:30 PM

#

that damn label

long robin Oct 29, 2022, 11:30 PM

#

yeah

#

there is a way to change where it is, but I forgor

#

I guess I can just use an ImGui::Text

golden schooner Oct 29, 2022, 11:31 PM

#

i think the way to change it is to make the label invisible via ## and use a separate TextUnformatted?

long robin Oct 29, 2022, 11:31 PM

#

golden schooner Oct 29, 2022, 11:31 PM

#

button is to slim, at least 2px

long robin Oct 29, 2022, 11:31 PM

#

yeah it's because of that damn font you made me use

#

now it's thicc

golden schooner Oct 29, 2022, 11:33 PM

#

listenyoupieceofshit

#

that default proggyclean font is ugly af, but ill take it

long robin Oct 29, 2022, 11:34 PM

#

or I can figure out how to use PushItemWidth

golden schooner Oct 29, 2022, 11:34 PM

#

nah

long robin Oct 29, 2022, 11:40 PM

#

god it's so smooth

#

also now the GUI is a little weird since some elements have text on the left 😄

golden schooner Oct 29, 2022, 11:41 PM

#

heh

shell inlet Oct 29, 2022, 11:41 PM

#

push push push

golden schooner Oct 29, 2022, 11:41 PM

#

how come the transition is so odd on the left and right

#

but its fine at the back

long robin Oct 29, 2022, 11:42 PM

#

edge-avoiding filter does not blur when the normals differ too much

#

the back wall and the floor are the same color, so it looks like there is a smooth transition

golden schooner Oct 29, 2022, 11:44 PM

#

i see

shell inlet Oct 29, 2022, 11:44 PM

#

by the way, about the gui...

#

you could have used a dropdown list for which version to use

#

and displayed the options for the selected one

#

just you know

#

a minor nitpick

long robin Oct 29, 2022, 11:45 PM

#

a

shell inlet Oct 29, 2022, 11:53 PM

#

jaker are you working on something?

#

I thought you pushed the changes

long robin Oct 29, 2022, 11:54 PM

#

I'm working on the imgui layout since you mentioned the dropdown thing

shell inlet Oct 29, 2022, 11:55 PM

#

I was only half serious but it must have sounded like a good idea probably

long robin Oct 29, 2022, 11:55 PM

#

fr

long robin Oct 29, 2022, 11:57 PM

#

long robin lel

I also decided that the optimal effort-ugliness ratio can be achieved by having the sample count in the middle like this

#

ok I pushed

#

might be easier to use a checkbox instead of the dropdown though lel

#

oh well

#

now I want to put this gui and full screen pass stuff into a function so I can use it in gltf_viewer without big pasta

shell inlet Oct 30, 2022, 12:06 AM

#

should have done that tbh

#

I'm going to try the latest version of the branch now

golden schooner Oct 30, 2022, 12:08 AM

#

hmm can you make it into a Fwog::DebugRsm call of sorts

long robin Oct 30, 2022, 12:09 AM

#

wdym

#

to display this imgui stuff?

golden schooner Oct 30, 2022, 12:09 AM

#

that function does all the Imgui nonsense

#

ye

long robin Oct 30, 2022, 12:10 AM

#

yeah that's what I was thinking

#

it would just have a load of inout parameters brainworm

shell inlet Oct 30, 2022, 12:12 AM

#

I had no idea that even 1 subsampled pass is so effective 👀

#

perhaps varying the kernel offset size between passes is going to make it even more OP

#

long robin Oct 30, 2022, 12:13 AM

#

yeah I noticed the difference between 2 and 3 passes was almost nothing

#

earlier

shell inlet Oct 30, 2022, 12:14 AM

#

it's basically pulling samples in from same offsets

long robin Oct 30, 2022, 12:15 AM

#

yeah

shell inlet Oct 30, 2022, 12:15 AM

#

but what if you make second one 1/2 of the first?

#

it would in theory bring more new info although from closer pixels

long robin Oct 30, 2022, 12:16 AM

#

I can try

shell inlet Oct 30, 2022, 12:17 AM

#

I don't understand those lines though

#

maybe doing cranley patterson rotation is inferior to shifting hammersley seed?

long robin Oct 30, 2022, 12:18 AM

#

could also be due to the repeat address mode

shell inlet Oct 30, 2022, 12:18 AM

#

it works good with white noise but perhaps low discrepancy sequence is just too deterministic

long robin Oct 30, 2022, 12:18 AM

#

when rmax is huge, doesn't it start sampling outside the RSM texture?

shell inlet Oct 30, 2022, 12:19 AM

#

#

this is unfiltered

#

same patterns

shell inlet Oct 30, 2022, 12:19 AM

#

long robin when rmax is huge, doesn't it start sampling outside the RSM texture?

rmax is in uv space, no?

#

so 1 is at best 1 repetition

long robin Oct 30, 2022, 12:20 AM

#

I think it's gonna look like ass with a big rMax no matter what

shell inlet Oct 30, 2022, 12:20 AM

#

yeah no it's not related to rmax

#

see for yourself that there is no difference in pattern when you change rmax

#

it only appears to expand

#

it's the hammersley LDS most likely

long robin Oct 30, 2022, 12:21 AM

#

yeah

#

maybe the solution is bigger noise

#

but eh

#

I think what's happening is that the lines are larger than the blue noise size

#

if you go closer, the blue noise fails to hide it even more

shell inlet Oct 30, 2022, 12:29 AM

#

I tried to offset the seed with blue noise but realized it beats the purpose and it didn't work anyway

#

if we change the underlying rng with dither mask we break the spatial blue noise properties

long robin Oct 30, 2022, 12:30 AM

#

I'm gonna try a massive blue noise texture real quick

shell inlet Oct 30, 2022, 12:30 AM

#

it will give better results I'm sure

#

but it's then harder to filter

#

need a wider filter size

long robin Oct 30, 2022, 12:31 AM

#

yeah it be like that

#

it be splotchy now

shell inlet Oct 30, 2022, 12:32 AM

#

tbh expected with that high amount of variance

long robin Oct 30, 2022, 12:33 AM

#

I have another idea

shell inlet Oct 30, 2022, 12:33 AM

#

use ray tracing?

#

and path guiding?

long robin Oct 30, 2022, 12:33 AM

#

nah I just wanted to add temporal filtering

#

but not fancy temporal filtering

shell inlet Oct 30, 2022, 12:33 AM

#

yeah that is another option

#

wait temporal filtering?

#

temporal accumulation?

long robin Oct 30, 2022, 12:34 AM

#

yeah, that's the term

#

not filtering

#

then we can have a per-frame random seed

#

or jitter or whatever

shell inlet Oct 30, 2022, 12:34 AM

#

temporal filtering would be, I would imagine, when you filter then store and then next frame filter stored on top of new

#

maybe

long robin Oct 30, 2022, 12:35 AM

#

idk I thought it would be like SVGF or something

shell inlet Oct 30, 2022, 12:35 AM

#

we kinda do half of that with iterative filter invocations

#

oh svgf is this but variance estimation and temporal accumulation

#

I lowkey pulled atrous part from it

#

castrated quite a bit

long robin Oct 30, 2022, 12:36 AM

#

heh

#

I didn't recognize it and was about to replace it with my own™️ atrous filter from another paper

#

I noticed some differences though

#

yours does a different pattern

shell inlet Oct 30, 2022, 12:37 AM

#

inspired by™️ another source code?

long robin Oct 30, 2022, 12:37 AM

#

and uses hard cutoffs instead of calculating soft edge weights

shell inlet Oct 30, 2022, 12:37 AM

#

ye it is castrated in that way

#

this is what I meant

long robin Oct 30, 2022, 12:37 AM

#

basically I was gonna use the code at the end of this but without the color weighting
https://jo.dreggn.org/home/2010_atrous.pdf

shell inlet Oct 30, 2022, 12:38 AM

#

oh hey look I totally mentioned that

#

varying subsampled offset

long robin Oct 30, 2022, 12:39 AM

#

that's the "a-trous" part

shell inlet Oct 30, 2022, 12:39 AM

#

ye calling it a-hole would be a bit...

#

by the way, since you evaluate more than 1 sample per pixel, it's possible to calculate variance from that too

#

SVGF evaluates it temporally but you can have that for free

#

but I'm not sure if it would give any quality increase for such a low frequency signal as indirect illumination

#

I'm certain that it would make sense for denoising shadows

long robin Oct 30, 2022, 12:48 AM

#

is the variance used for choosing the temporal weight?

shell inlet Oct 30, 2022, 12:48 AM

#

variance is used for edge stopping

#

to not overblur parts that require more sharpness

#

e.g. shadows (again)

#

if you were to ignore variance your shadows will become a blurry mess regardless of how far away the shadow caster is

#

regardless of the actual penumbra size

long robin Oct 30, 2022, 12:49 AM

#

I see

#

it seems like it could be useful to have in specular denoising as well

shell inlet Oct 30, 2022, 12:50 AM

#

you need to figure out specular reprojection first

long robin Oct 30, 2022, 12:51 AM

#

ye that seems like a tough problem

shell inlet Oct 30, 2022, 12:51 AM

#

but if you have more than 1spp (not modern ray tracing) then yes

long robin Oct 30, 2022, 12:52 AM

#

so if you have more variance, you want a lower filter width for that pixel?

shell inlet Oct 30, 2022, 12:52 AM

#

vice versa

#

less variance means it's converged

#

so you don't want to blur it too much

long robin Oct 30, 2022, 12:52 AM

#

ah

#

I don't see how that helps with shadows exactly. Is it because shadowed areas have less variance?

shell inlet Oct 30, 2022, 12:57 AM

#

#

wait that's only tangentially related

#

basically you have high variance only in the penumbra, so you don't want to blur parts fully in shadow and fully illuminated, because those don't require any filtering as they are converged, otherwise you will lose detail if you mix sharp shadow with illuminated part next to it

#

I don't really have anything to add to it

long robin Oct 30, 2022, 1:04 AM

#

oh that makes sense

shell inlet Oct 30, 2022, 1:05 AM

#

1spp visibility is really a binary thing not a gradient

#

either visible or not

long robin Oct 30, 2022, 1:06 AM

#

yea so you'd get big variance in penumbras

shell inlet Oct 30, 2022, 1:06 AM

#

so you accumulate some samples to get the idea of how much there is variance, if it is either 1 or 0 all the time it's 0 variance, otherwise based on how much there is variance you want to blur more to hide lack of samples

#

ok you get it all good

long robin Oct 30, 2022, 1:07 AM

#

ye thx for explaining

#

now I'm gonna try this temporal accumulation

#

stage 1 will be accumulating when the camera is still

shell inlet Oct 30, 2022, 1:08 AM

#

maybe port the menu to gltf viewer?

long robin Oct 30, 2022, 1:09 AM

#

oh right let me do that stuff first

long robin Oct 30, 2022, 10:24 AM

#

maybe it's time to disable this extension

coral axle Oct 30, 2022, 10:38 AM

#

c# version when? 😛

golden schooner Oct 30, 2022, 10:38 AM

#

👉 EngineKit 😉 pretty much

long robin Oct 30, 2022, 10:39 AM

#

#1019740157798273024 has my official recommendation

coral axle Oct 30, 2022, 10:39 AM

#

long robin maybe it's time to disable this extension

hah, mine does that

heavy cipher Oct 30, 2022, 12:04 PM

#

long robin maybe it's time to disable this extension

it is just showing that that function compiles to the x64 instruction vgrur

#

just the guarantees of using a low level language

shell inlet Oct 30, 2022, 8:04 PM

#

long robin it be splotchy now

is this blue noise?

#

somehow under closer look it looks no different than white

long robin Oct 30, 2022, 8:30 PM

#

It is blue noise, supposedly

shell inlet Oct 30, 2022, 8:34 PM

#

yeah but it doesn't look like it

#

this is how blue noise is supposed to look

#

but then there are stupid lines apparent due to not enough randomness being there

#

this is what happens if you rotate both xi components with one noise value

long robin Oct 30, 2022, 8:37 PM

#

that deffo looks bluer

shell inlet Oct 30, 2022, 8:38 PM

#

ye all random numbers are correlated with blue noise now

#

same blue noise

#

which is how it was supposed to be

#

how to get rid of the lines tho

long robin Oct 30, 2022, 8:40 PM

#

lower rmax, bigger noise, or temporal accumulation

#

or more samples

shell inlet Oct 30, 2022, 8:48 PM

#

this is fine

golden schooner Oct 30, 2022, 8:49 PM

#

this looks beautiful

shell inlet Oct 30, 2022, 8:50 PM

#

me feet got cold from the gpu making wind

golden schooner Oct 30, 2022, 8:50 PM

#

hehe

#

let me run it on my potato

#

its on examples-refactor neh?

shell inlet Oct 30, 2022, 8:51 PM

#

yep

golden schooner Oct 30, 2022, 9:00 PM

#

oh boy

#

8 fps in default

long robin Oct 30, 2022, 9:01 PM

#

turn down the samples

golden schooner Oct 30, 2022, 9:01 PM

#

mouse/keeb movement doesnt work

#

ref picture

long robin Oct 30, 2022, 9:03 PM

#

woof 117 ms

golden schooner Oct 30, 2022, 9:03 PM

#

: D

#

filtered doesnt really do much on this hardware

#

fps wiggle from 6 to 23 constantly tho

#

there should be an RSM off option and just shadows on

shell inlet Oct 30, 2022, 9:06 PM

#

is this ubuntu

golden schooner Oct 30, 2022, 9:06 PM

#

manjaro

shell inlet Oct 30, 2022, 9:07 PM

#

anyways I think the pattern arises from uv sampling

golden schooner Oct 30, 2022, 9:07 PM

#

you mean what you can see in the first pic?

shell inlet Oct 30, 2022, 9:07 PM

#

never seen anything like that appear in ray tracing

shell inlet Oct 30, 2022, 9:07 PM

#

shell inlet this is how blue noise is supposed to look

I'm talking about this

golden schooner Oct 30, 2022, 9:07 PM

#

ah

long robin Oct 30, 2022, 9:08 PM

#

Ye it's just sampling a projected circle

#

Some geometry or sun angles will give worse line artifacts

shell inlet Oct 30, 2022, 9:09 PM

#

so it's preferable that there is less correlation in random numbers

#

I actually think if you were to use white noise xy instead of blue noise xy it will be distributed worse

#

with blue noise there is at least some correlation even if xy are not correlated

#

never mind there is no visual difference

#

just tried white noise texture

#

turns out blue noise properties are completely destroyed unless you correlate every random number in a sequence with the same blue noise value

long robin Oct 30, 2022, 9:18 PM

#

That explains some things

shell inlet Oct 31, 2022, 12:20 AM

#

I think we have one pi too many

#

I said that before but one pi is canceled out by lambertian brdf albedo/pi

#

so in the end it's just r and the sampled area terms

long robin Oct 31, 2022, 12:34 AM

#

one pi, I don't know why

shell inlet Oct 31, 2022, 12:35 AM

#

it doesn't even matter how hard you pi

#

so I think it's true because of how directional integral surrounded by emissive surfaces integrates to 1

#

without lambertian brdf it ofc integrates to pi

#

and I use directional integral as a reference for surface one, where it turns out it behaves same way

#

#

then estimators converge to the same result too

#

now look at the estimator with quadratic distribution on the circle

#

there is 2piR^2 but we also divide the integrated function by pi due to lambertian brdf being there

#

which means pdf is 1/(rPi), area normalization term is 1/PiR^2 and then lambertian brdf term is 1/Pi, all adds up to rPiR^2 weight if you cancel out one pi from pdf

#

PiR^2 turns out to be a constant that you can take out of the integrand and multiply whole integral by it (an optimization we do in code)

#

note that there is some error in what estimators converge to due stochastic nature and then precision issues when there is a lot of samples

shell inlet Oct 31, 2022, 1:08 AM

#

ok I am literally an idiot

#

it cancels out pi

#

and there is 2pi not pi^2

#

lmao

#

so 2 is what's left as a result

#

so this is apparently the right intensity...?

#

2R^2?

#

man idk anymore

#

this is unfiltered for comparison (2piR^2)

#

really need blender to see which one is closer

#

since jaker won't do tests I'm going to have to do that

shell inlet Oct 31, 2022, 2:00 AM

#

#

blender ref with 1 bounce idirect

#

2R^2 seems to be the right one

#

so neither pi nor 2pi in the end

#

📎 ref.blend

#

you can rename area variable to normalization factor or something

#

also there is a hidden canceled out pi in it but whatever

long robin Oct 31, 2022, 2:16 AM

#

I'm back

shell inlet Oct 31, 2022, 2:24 AM

#

I still have no idea why we don't normalize by the area of the circle but whatever

#

it converges to the right result in my graph that should be enough for me

#

if anyone asks say it's "empirically proved"

long robin Oct 31, 2022, 2:26 AM

#

I'll look at fixing the math with your updated numbers after I finish putting RSM into its own file for ez reuse

shell inlet Oct 31, 2022, 2:26 AM

#

it's just one change in the shader

#

change TWO_PI to 2.0

long robin Oct 31, 2022, 2:26 AM

#

just this line?
float worldSpaceSampledArea = TWO_PI * rMaxWorld * rMaxWorld;

#

and I'll change the variable name

shell inlet Oct 31, 2022, 2:27 AM

#

ye

long robin Oct 31, 2022, 2:28 AM

#

hmm what about the noise

#

should I make it only use one component of blue noise

shell inlet Oct 31, 2022, 2:29 AM

#

if you want lines but narrower filter requirements go ahead

#

blue noise is supposed to allow you to shrink the filter because of how it's distributed

#

I personally think it looks uglier with lines visible

#

they won't be lines all the time, basically depends on the scene as you said yourself

long robin Oct 31, 2022, 2:32 AM

#

yeah I'll mess with that later

shell inlet Oct 31, 2022, 2:32 AM

#

but an artifact is an artifact

long robin Oct 31, 2022, 5:33 AM

#

https://i.imgur.com/6q5s9OR.jpeg

Imgur

long robin Oct 31, 2022, 8:24 AM

#

with mipmaps the annoying aliasing is mostly gone (also increased the sun's brightness a bunch in this one)

long robin Oct 31, 2022, 8:43 AM

#

https://gfycat.com/AlienatedFatalBlackfish

Gfycat

ShareX

▶ Play video

#

this goes hard

golden schooner Oct 31, 2022, 8:51 AM

#

https://tenor.com/view/trailer-park-boys-the-way-she-goes-oh-well-gif-14173918

Tenor

long robin Oct 31, 2022, 9:05 AM

#

temporal blend vs without (though I am basically cheating)

#

now I "just" need to reproject I guess

#

https://gfycat.com/ExcellentAlertCarpenterant

Gfycat

ShareX

▶ Play video

#

without reprojection or sample rejection you get this

golden schooner Oct 31, 2022, 9:08 AM

#

Goto settings -> Graphics -> [ ] Motion Blur -> Done

long robin Oct 31, 2022, 9:13 AM

#

gonna fumble around with reprojection tomorrow I guess

#

hmm, I think that means I need the old depth buffer too (and possibly some other old buffers to aid in rejection heuristics)?

#

not gonna have moving geometry so I can ignore motion vectors

#

just need to account for the camera's movement

#

which means I can calculate the motion of a pixel using just the current and previous frame's viewProj matrix nice

shell inlet Oct 31, 2022, 10:33 AM

#

long robin with mipmaps the annoying aliasing is mostly gone (also increased the sun's brig...

small rmax small pp

shell inlet Oct 31, 2022, 10:35 AM

#

long robin temporal blend vs without (though I am basically cheating)

very smooth, if only there wasn't that artifact in the distance from my bad filter

#

btw temporal accumulation is usually done on unfiltered result to avoid excessive bias

#

here it doesn't matter I guess since the technique is one huge bias due to lack of visibility for indirect

#

as well as you don't need the unfiltered history for other stuff

#

I am also going to say that temporal reprojection is going to give you pain if you haven't implemented it before, it's a thing that appears deceptively simple but in practice you will face quite a few stepping stones, so if you want a good reference implementation I suggest you to look at SVGF code https://github.com/NVIDIAGameWorks/Falcor/tree/master/Source/RenderPasses/SVGFPass

dire badge Oct 31, 2022, 6:54 PM

#

I think reprojection will always have some issues so I did not even try

#

Meanwhile MSAA will work just fine for me 🔍🐸

long robin Oct 31, 2022, 7:45 PM

#

@dire badge I'm trying to use it to improve the quality of indirect illumination

#

I'm not gonna try it for antialiasing because I'm aware that's super hard and way more sensitive to small issues 😄

long robin Oct 31, 2022, 8:22 PM

#

I also realized I could use variable rate shading for this (instead of just 2x2 shading)

heavy cipher Oct 31, 2022, 8:23 PM

#

does ogl have VRS?

long robin Oct 31, 2022, 8:24 PM

#

yes (nv extension) but I'm not referring to the hardware kind

#

I just need to refresh myself on common ways to create the shading rate image

heavy cipher Oct 31, 2022, 8:27 PM

#

🤔

#

i am lost

long robin Oct 31, 2022, 8:28 PM

#

for calculating indirect illumination

#

it is an expensive and low-frequency effect, so it would be nice to compute it at a lower res wherever possible

#

so I am thinking of something akin to software VRS to achieve that

#

I was reminded of its existence by looking at the most recent post in #showcase
https://github.com/BoyBaykiller/IDKEngine

long robin Nov 1, 2022, 5:29 AM

#

shell inlet btw temporal accumulation is usually done on unfiltered result to avoid excessiv...

so is it recommended to blend with the previous frame's unfiltered result, or should I blend the current frame's unfiltered with the previous frame's filtered?

shell inlet Nov 1, 2022, 5:30 AM

#

it is recommended to keep accumulating samples within the estimator effectively extending it across time

#

accumulating filtered results distorts the estimation

long robin Nov 1, 2022, 5:31 AM

#

I see. The filtered result is thrown away at the end of the frame

shell inlet Nov 1, 2022, 5:32 AM

#

this is besides the point but it also only really works for view angle invariant brdfs such as lambertian

#

for view dependent brdfs it is trickier because you need to re-evaluate the brdf for each sample, in restir pt paper this is referred to as shift mapping

long robin Nov 1, 2022, 5:33 AM

#

interesting

#

does that change what you store in each sample?

#

since you care about the direction the light came from

shell inlet Nov 1, 2022, 5:38 AM

#

I would like to say but I need to give restir pt an actual in-depth read first

long robin Nov 1, 2022, 5:38 AM

#

yeah I tried reading it, but I got hung up on the "reservoir sampling" part, so I had to take a detour to learn what that is first

#

pretty sure I know what reservoir sampling is now 😄

shell inlet Nov 1, 2022, 5:40 AM

#

for all I know there is a unique set of inputs that varies between different points on a surface spatially(if you try to merge neighboring samples together) and temporally(if it moved and you need to reproject), so you need to remap the results to correct for the change

long robin Nov 1, 2022, 5:40 AM

#

is this only addressed in restir pt? it seems like a problem for any temporal filter

#

or did they previously just not temporally filter specular reflection

shell inlet Nov 1, 2022, 5:41 AM

#

beats me

#

I haven't read all that many papers

long robin Nov 1, 2022, 5:42 AM

#

now that I think of it, temporally reprojecting+filtering specular reflection is hard af and there are a lot of hacky ways to address it

shell inlet Nov 1, 2022, 5:42 AM

#

svgf seems to ignore bias from reprojection because filtering is inherently adding bias

#

for ris based techniques starting from restir it's different because it's not only reusing them temporally but also spatially

#

which is what they say is a gradient domain rendering problem, whatever that means

#

totally goes over my head

long robin Nov 1, 2022, 5:43 AM

#

there was a presentation by EA about their PICA renderer which addressed what they did for reprojecting specular

#

this is probably it
https://youtu.be/MyTOGHqyquU

shell inlet Nov 1, 2022, 5:45 AM

#

if the technique doesn't claim to be unbiased then there should be no concern about any correction

long robin Nov 1, 2022, 5:45 AM

#

yeah idk about this presentation

shell inlet Nov 1, 2022, 5:48 AM

#

well if it's realtime oriented then you should expect corners to be cut imo

#

so my guess is that it's prioritized that it looks okay over whether it's mathematically correct or not

long robin Nov 1, 2022, 5:49 AM

#

that's what it's all about 👌 👌

shell inlet Nov 1, 2022, 5:50 AM

#

well that and that it's fast enough

long robin Nov 1, 2022, 8:11 AM

#

💀

    GLM_FUNC_QUALIFIER float uintBitsToFloat(uint const& v)
    {
        union
        {
            uint in;
            float out;
        } u;

        u.in = v;

        return u.out;
    }

#

    template<length_t L, qualifier Q>
    GLM_FUNC_QUALIFIER vec<L, float, Q> uintBitsToFloat(vec<L, uint, Q> const& v)
    {
        return reinterpret_cast<vec<L, float, Q>&>(const_cast<vec<L, uint, Q>&>(v));
    }

shell inlet Nov 1, 2022, 8:23 AM

#

UB? I don't speak C++

#

well first one is defo UB

#

those two should have been both just memcpy

long robin Nov 1, 2022, 8:25 AM

#

ye, those are stinky UB

#

sadly the kind of UB that "works"

shell inlet Nov 1, 2022, 8:30 AM

#

enable strict aliasing for "fun" time

long robin Nov 1, 2022, 9:34 AM

#

for some reason I am having trouble using glCopyImageSubData with what I'm 99% sure are valid parameters ( opengl moment?)

OpenGL Debug message (1280): GL_INVALID_ENUM error generated. <srcTarget> or <dstTarget> is not a valid texture target.
Source: API
Type: Error
Severity: high

#

oh well, I don't need to copy anything

long robin Nov 1, 2022, 10:04 AM

#

found another bug in Fwog

#

with the framebuffer cache again 😄

#

it caches pointers to textures and compares the pointers to see if the textures are the same

#

but if you swap two textures, the pointers are the same while the texture IDs are different

#

shit, it's still flickering

#

ah, the hash was messed up too

long robin Nov 1, 2022, 10:51 AM

#

pointers were a mistake

long robin Nov 1, 2022, 11:20 AM

#

sweet jesus that was a painful debugging session

#

framebuffer cache is now two vectors (key and value) so I don't have to implement stupid hashes and also so it's easier to debug

#

and it's probably faster 99% of the time too

dire badge Nov 1, 2022, 2:19 PM

#

Been there, done that, do not use pointers in caching

#

using explicit framebuffer objects bypasses the need of a framebuffer cache

#

I don't quite see why the framebuffer object should be hidden

long robin Nov 1, 2022, 7:31 PM

#

because it's unnecessary friction imo

#

it's a lot nicer to just say "draw with this list of render targets" and not have to worry about managing an FBO and binding stuff to it (especially if things change each frame or the window is resized, etc.)

#

tl;dr: I find framebuffers to be an annoying extra thing to handle, so I copied vulkan dynamic rendering

golden schooner Nov 1, 2022, 7:45 PM

#

siding with jaker

brisk narwhal Nov 1, 2022, 7:58 PM

#

why not do a sharedptr

#

'cept it's not a shared ptr

#

like

#

class ReferenceCountable {
    uint64 grab();
    uint64 drop();
    atomic_int ...
};

class Framebuffer : public ReferenceCountable {};

long robin Nov 1, 2022, 8:00 PM

#

because the user doesn't even know about the existence of framebuffers

#

the framebuffer cache exists as an optimization so Fwog doesn't create a new FBO every time you start rendering

golden schooner Nov 1, 2022, 8:00 PM

#

the only thing exposed to the user is BeginRender/EndRender

brisk narwhal Nov 1, 2022, 8:00 PM

#

hmm I understand, I never used VK_KHR_dynamic_rendering

golden schooner Nov 1, 2022, 8:01 PM

#

dynamic state just means setting various things willy nilly, but within the bounds of fwog which keeps track of it

#

(scissor, viewport, ...)

brisk narwhal Nov 1, 2022, 8:02 PM

#

with vulkan dynamic rendering you have to manage the images yourself though

typedef struct VkRenderingAttachmentInfo {
    VkStructureType          sType;
    const void*              pNext;
    VkImageView              imageView;
    VkImageLayout            imageLayout;
    VkResolveModeFlagBits    resolveMode;
    VkImageView              resolveImageView;
    VkImageLayout            resolveImageLayout;
    VkAttachmentLoadOp       loadOp;
    VkAttachmentStoreOp      storeOp;
    VkClearValue             clearValue;
} VkRenderingAttachmentInfo;```

#

is it not the case with Fwog?

golden schooner Nov 1, 2022, 8:02 PM

#

ok scissor and viewport are the worst examples here

long robin Nov 1, 2022, 8:02 PM

#

here's an example of it in action (read until Fwog::BeginRendering(...))
https://github.com/JuanDiegoMontoya/Fwog/blob/examples-refactor/example/02_deferred.cpp#L388

brisk narwhal Nov 1, 2022, 8:03 PM

#

Ah so you don't even let the user manage images

#

all hidden, very OpenGL-esque

golden schooner Nov 1, 2022, 8:03 PM

#

that was the idea

#

so that the user wont get the idea of using some obscure legacyism

long robin Nov 1, 2022, 8:03 PM

#

brisk narwhal Ah so you don't even let the user manage images

what do you mean by "manage images"?

#

textures are exposed to the user for them to manipulate as they please

#

that's what e.g., &frame.gAlbedo.value() is (a pointer to a texture)

brisk narwhal Nov 1, 2022, 8:04 PM

#

I don't know how opengl calls them

#

render buffers? (I think)

long robin Nov 1, 2022, 8:05 PM

#

yeah opengl has those as an optimization

#

you can use regular textures for rendering though

brisk narwhal Nov 1, 2022, 8:05 PM

#

Ah you don't have to use them

long robin Nov 1, 2022, 8:05 PM

#

and in fact I don't even expose render buffers (or use them internally)

brisk narwhal Nov 1, 2022, 8:05 PM

#

fair enough

long robin Nov 1, 2022, 8:06 PM

#

maybe I can add a createInfo flag for textures to make them write-only so I can secretly turn them into a renderbuffer

golden schooner Nov 1, 2022, 8:06 PM

#

hmm

long robin Nov 1, 2022, 8:06 PM

#

it would probably add some buggy edge cases though 😄

golden schooner Nov 1, 2022, 8:06 PM

#

that should be part of the framebuffer struct

long robin Nov 1, 2022, 8:07 PM

#

what framebuffer struct listenyoupieceofshit

golden schooner Nov 1, 2022, 8:07 PM

#

the thing you pass into BeginRender

brisk narwhal Nov 1, 2022, 8:07 PM

#

framebuffers don't exist it was all just a dream

golden schooner Nov 1, 2022, 8:07 PM

#

renderinfo

long robin Nov 1, 2022, 8:07 PM

#

golden schooner the thing you pass into BeginRender

I can't transmogrify arbitrary textures to suddenly be a renderbuffer though

golden schooner Nov 1, 2022, 8:08 PM

#

it shouldnt be a thing for texture creation

long robin Nov 1, 2022, 8:08 PM

#

you have to know how the texture is used up-front to know if it can be a renderbuffer

#

because you can't sample them

golden schooner Nov 1, 2022, 8:08 PM

#

you usually do know that

#

in 100.0% of the cases 🙂

long robin Nov 1, 2022, 8:09 PM

#

yeah, but specifying that in the renderinfo means you can make mistakes

golden schooner Nov 1, 2022, 8:09 PM

#

nah

#

semantically it has no place in the texturecreateinfo struct

#

</discussion>

long robin Nov 1, 2022, 8:09 PM

#

hmm

golden schooner Nov 1, 2022, 8:10 PM

#

that would confuse the consumer for no reason

long robin Nov 1, 2022, 8:10 PM

#

in vulkan it's part of the image create info (the usage flags)

golden schooner Nov 1, 2022, 8:10 PM

#

shit

#

well, i think id still say the same if we were doing Vwog

long robin Nov 1, 2022, 8:11 PM

#

also, specifying it in the renderinfo would mean you have to lazily create textures and renderbuffers, because you don't know which one it is until you try to sample or render to it

golden schooner Nov 1, 2022, 8:11 PM

#

but do you not know that upfront as well?

long robin Nov 1, 2022, 8:11 PM

#

textures and renderbuffers are two separate types in opengl btw

long robin Nov 1, 2022, 8:11 PM

#

golden schooner but do you not know that upfront as well?

you do! which is why you specify that in the texture create info!!!

golden schooner Nov 1, 2022, 8:11 PM

#

yo dont decide in the middle of whenever whether a texture is never sampled from but written to etc?

golden schooner Nov 1, 2022, 8:12 PM

#

long robin textures and renderbuffers are two separate types in opengl btw

yeah, thats why it should be part of the renderinfo in GL

#

or

#

you have a renderbuffer type itself (a derivate of texture)

long robin Nov 1, 2022, 8:12 PM

#

golden schooner yeah, thats why it should be part of the renderinfo in GL

I think you have a confusion

#

renderinfo can be created on-demand

#

it's a simple wrapper around a real texture

#

it's not an RAII object and doesn't own anything itself

#

  struct RenderAttachment
  {
    const Texture* texture = nullptr;
    ClearValue clearValue;
    bool clearOnLoad = false;
  };

  struct RenderInfo
  {
    std::string_view name;
    const Viewport* viewport = nullptr;
    std::span<const RenderAttachment> colorAttachments;
    const RenderAttachment* depthAttachment = nullptr;
    const RenderAttachment* stencilAttachment = nullptr;
  };

golden schooner Nov 1, 2022, 8:14 PM

#

i wasnt thinking about raii or ownerhip

#

ye, it should be part of RenderAttachment, whether Texture is a renderbuffer or not

#

is what i mean(t)

long robin Nov 1, 2022, 8:15 PM

#

why should it not be part of the texture?

#

is that not just another place where you can put the wrong thing?

golden schooner Nov 1, 2022, 8:15 PM

#

when do you use a texture as a renderbuffe outside of fboisms?

long robin Nov 1, 2022, 8:16 PM

#

well, never, but when I create a texture, how am I supposed to know it's an OpenGL texture or renderbuffer?

golden schooner Nov 1, 2022, 8:16 PM

#

you dont need to know

long robin Nov 1, 2022, 8:16 PM

#

or do you mean that texture should be a variant?

golden schooner Nov 1, 2022, 8:16 PM

#

you only need to know when you construct a renderinfo, because thats when you decide to slap textureA, B and C in there

golden schooner Nov 1, 2022, 8:17 PM

#

long robin or do you mean that `texture` should be a variant?

it could be an alternative

long robin Nov 1, 2022, 8:17 PM

#

okay, here's the least sucky solution (in my mind):

create a new type called RenderBuffer or something which is internally backed by an actual renderbuffer
change Texture* texture in RenderAttachment to std::variant<Texture*, RenderBuffer*>

golden schooner Nov 1, 2022, 8:18 PM

#

if you stick in a RendeBuffer* then renderInfo knows to construct a RENDERBUFFER, if not, then not

long robin Nov 1, 2022, 8:18 PM

#

construct

golden schooner Nov 1, 2022, 8:18 PM

#

let me put it in quotes

long robin Nov 1, 2022, 8:18 PM

#

but yes, it knows how to attach it to the FBO then

golden schooner Nov 1, 2022, 8:18 PM

#

"construct"

#

ye, i could live with that, it saves a possible bool isRenderBuffer of sorts in the RenderAttachment thing

heavy cipher Nov 1, 2022, 8:18 PM

#

renderbuffers 💀

golden schooner Nov 1, 2022, 8:19 PM

#

what a tangent 😄

long robin Nov 1, 2022, 8:19 PM

#

my original idea was to overload Texture by adding a createInfo flag to make it rendertarget-only, which secretly turns it into a renderbuffer

#

and makes every other operation with it invalid

heavy cipher Nov 1, 2022, 8:19 PM

#

have you verified renderbuffers are not just a useless vestige of ogl

golden schooner Nov 1, 2022, 8:19 PM

#

if you meant deriving from Texture, then ye 🙂

long robin Nov 1, 2022, 8:20 PM

#

heavy cipher have you verified renderbuffers are not just a useless vestige of ogl

well, graphite said one time that they do something on NV drivers KEKW

#

truly a man of source

heavy cipher Nov 1, 2022, 8:20 PM

#

~~have you verified graphite is not just a useless vestige of ogl~~

long robin Nov 1, 2022, 8:20 PM

#

I agree though, this BS is why I didn't support them in the beginning~~, plus the fact I completely forgot about them~~

golden schooner Nov 1, 2022, 8:21 PM

#

blame lvd

long robin Nov 1, 2022, 8:21 PM

#

it was a good point to bring up

#

final decision: Fwog will not be having render buffers because obviously they complicate shit for (probably) little benefit

heavy cipher Nov 1, 2022, 8:23 PM

#

tired: supporting some random legacy crap for NV driver
wired: not supporting some random legacy crap and going big, forcing NV to fix driver

brisk narwhal Nov 1, 2022, 8:24 PM

#

I didn't even know what renderbuffers were, I just googled "opengl framebuffer images" KEKW

long robin Nov 1, 2022, 8:25 PM

#

the troll was so subtle that you didn't even know you were doing it smart

long robin Nov 1, 2022, 8:29 PM

#

brisk narwhal with vulkan dynamic rendering you have to manage the images yourself though ```c...

FYI (since I didn't answer it earlier), this is indeed the case with Fwog

dire badge Nov 1, 2022, 8:49 PM

#

renderbuffer may have been slightly more useful on tiled architecture. Driver could get away without allocating memory for them.

heavy cipher Nov 1, 2022, 8:52 PM

#

why could it not for textures?

long robin Nov 1, 2022, 8:55 PM

#

after all, it's just one more in the pile of heuristics the driver is already doing okey

golden schooner Nov 2, 2022, 12:46 AM

#

when do you merge !46 ?

long robin Nov 2, 2022, 12:52 AM

#

Uh

#

I was planning to do that after I merge my thing

golden schooner Nov 2, 2022, 12:52 AM

#

https://tenor.com/view/community-ken-jeong-ben-chang-ill-allow-it-allowed-gif-4468552

Tenor

i'll allow it

▶ Play video

long robin Nov 2, 2022, 12:52 AM

#

But since it's such an easy thing to merge I'll just do it

#

Ok done

golden schooner Nov 2, 2022, 12:57 AM

#

was hard work

dire badge Nov 2, 2022, 6:39 AM

#

heavy cipher why could it not for textures?

You can do lazy allocation for textures, but I think the idea for textures is that you do sample from them. Meanwhile, depth renderbuffer likely exists only for the depth test, which can work with the on-chip tile memory without having to spill it to memory.

long robin Nov 2, 2022, 9:23 AM

#

https://gfycat.com/PoshDismalDuckbillplatypus

Gfycat

ShareX

▶ Play video

#

first attempt at reprojection

#

now time to discard samples

long robin Nov 2, 2022, 9:46 AM

#

https://gfycat.com/IdolizedLoneEagle

Gfycat

ShareX

▶ Play video

#

mostly works, except for the weird artifacts in the middle when I move straight forward (at the end)

#

something's wrong with my depth-based rejection

#

wait, no

#

something else is wrong

#

https://gfycat.com/TimelyTenseHawaiianmonkseal

Gfycat

ShareX

▶ Play video

#

disabled depth rejection

#

maybe that's just a weird inherent sampling artifact rgbemoji

#

I added a debug color to make rejection obvious, and none of the stuff in the middle is getting rejected

long robin Nov 2, 2022, 9:59 AM

#

long robin maybe that's just a weird inherent sampling artifact <:rgbemoji:8478302111015895...

I guess the problem is that when you move forward and the pixels get magnified, multiple pixels end up mapping to the same one (or sampling the same area), giving blocky artifacts like that in the reprojection

long robin Nov 2, 2022, 10:14 AM

#

now I need to vary the accumulation rate or something in recently disoccluded areas

#

because right now they accumulate very slowly

#

likewise, stable areas need to have a lower blend factor so they can accumulate more samples

#

maybe if I read the SVGF paper more it will explain these things 😄

#

https://gfycat.com/FelineElementaryBullmastiff

Gfycat

ShareX

▶ Play video

long robin Nov 2, 2022, 11:03 AM

#

https://gfycat.com/MerryAnnualGlowworm

Gfycat

ShareX

▶ Play video

long robin Nov 2, 2022, 11:56 AM

#

changes are pushed if anyone else wants to experience this life-changing opportunity

golden schooner Nov 2, 2022, 12:02 PM

#

i have a feeling my inductors will jump off of the mainboard

#

however, i will ive it a try lat0r

long robin Nov 2, 2022, 12:05 PM

#

it actually allows you to use fewer samples while still looking good

#

so it improves perf at the cost of new and exciting visual artifacts

golden schooner Nov 2, 2022, 12:09 PM

#

on it already 🙂

#

heh we need to fix the tbb bs at some point again, its quite annoying, will see if i can find the commit of that thing, which went somewhere

#

was that gltfviewer?

long robin Nov 2, 2022, 12:11 PM

#

Just run 02_deferred

golden schooner Nov 2, 2022, 12:11 PM

#

ah

#

8 fps out of the box 😄

long robin Nov 2, 2022, 12:12 PM

#

Yeah 03 and above require multithreading

long robin Nov 2, 2022, 12:12 PM

#

golden schooner 8 fps out of the box 😄

Ye I forgor to reduce samples by default

#

You can also reduce the filter

golden schooner Nov 2, 2022, 12:12 PM

#

also the mouse is invisible

long robin Nov 2, 2022, 12:12 PM

#

Yup, press `

golden schooner Nov 2, 2022, 12:12 PM

#

ah

long robin Nov 2, 2022, 12:12 PM

#

aka ~

golden schooner Nov 2, 2022, 12:12 PM

#

oui

#

fortunately i understand quake console-isms 😄

long robin Nov 2, 2022, 12:13 PM

#

I need to change that so the mouse is visible by default

#

Try 1 sample per pixel

#

And 1/1 for the filters

#

It should look decent until you turn the camera 😂

golden schooner Nov 2, 2022, 12:14 PM

#

the "artifacts" you see thre flicker like crazy

#

in the unfiltered view, even with 400 samples, you also still see artifacts a lot

long robin Nov 2, 2022, 12:15 PM

#

Flickering?

golden schooner Nov 2, 2022, 12:15 PM

#

ye like temporal shit is happening

long robin Nov 2, 2022, 12:15 PM

#

Unfiltered should not have those

golden schooner Nov 2, 2022, 12:15 PM

#

the artifacts change every frame

#

true, only filtered

long robin Nov 2, 2022, 12:16 PM

#

Okay

#

Try changing the temporal alpha

#

When it's 0 it shouldn't flicker much

#

When it's 1 it will only take the current frame, which should make it flicker a lot

golden schooner Nov 2, 2022, 12:17 PM

#

no effect at all

long robin Nov 2, 2022, 12:17 PM

#

Hmm

golden schooner Nov 2, 2022, 12:17 PM

#

hmm maybe super slightly

long robin Nov 2, 2022, 12:17 PM

#

I probably goofed when I committed

golden schooner Nov 2, 2022, 12:17 PM

#

but no big differece

#

i was trying to record it but the screencap just closes itself when i do 😄

long robin Nov 2, 2022, 12:17 PM

#

Lol

#

If you don't move the camera at all, does it still flicker?

golden schooner Nov 2, 2022, 12:19 PM

#

maybe i can employ ffmpeg later and capture it myself

#

yes it flickers at any codnition when filtered

long robin Nov 2, 2022, 12:19 PM

#

What if you increase the samples to 5 or so

golden schooner Nov 2, 2022, 12:19 PM

#

tried to play with the samples and filter options too

long robin Nov 2, 2022, 12:19 PM

#

Increasing the sample count should've helped

golden schooner Nov 2, 2022, 12:19 PM

#

nope

long robin Nov 2, 2022, 12:20 PM

#

Maybe it's just harder for me to notice because I get 240 FPS 😂

#

It's probably not a bug in any case

golden schooner Nov 2, 2022, 12:21 PM

#

very possible

#

ill try to get a video capture of it

long robin Nov 2, 2022, 12:21 PM

#

If you want to hack, you could try changing this variable to .999
https://github.com/JuanDiegoMontoya/Fwog/blob/examples-refactor/example/shaders/rsm/Reproject.comp.glsl#L52

#

That should make it more stable

#

But it will accumulate more slowly

#

It's also quite amazing that 1 sample per pixel hurts your iGPU this bad

#

I gotta sleep though. I'll be back in 6 hours

golden schooner Nov 2, 2022, 12:25 PM

#

gn8

#

ill fiddle with it later, also have to get back to werk

golden schooner Nov 2, 2022, 6:10 PM

#

6hrs gone already?

#

man time flies

long robin Nov 2, 2022, 6:16 PM

#

Time flies like an arrow. Fruit flies like banana

golden schooner Nov 2, 2022, 6:20 PM

#

thats what shrimple life is 😛

long robin Nov 3, 2022, 7:51 AM

#

https://gfycat.com/HandsomeClassicAmericanavocet
https://gfycat.com/NeatBrightBlueandgoldmackaw

Gfycat

ShareX

▶ Play video

Gfycat

ShareX

▶ Play video

#

nearest vs linear filter for reprojected samples

#

the former introduces a LOT of artifacts, which were destroying my mind

shell inlet Nov 3, 2022, 8:46 AM

#

did you look into the SVGF source code?

#

it does software bilinear filtering to reduce these artifacts

#

this is why I said reprojection appears to be deceptively simple to implement while it's actually tricky

long robin Nov 3, 2022, 8:49 AM

#

I've been looking at several things

#

mostly my own code though

#

I am enjoying the adventure of winging it so far

shell inlet Nov 3, 2022, 8:51 AM

#

biggest offender in this whole technique is still variance though

#

if only we could sample using a directional estimator

#

but it's not possible I think

#

actually it's possible but would require marching the shadow map

#

to find the intersections from the given visible point to the point of the intersection of the casted ray in an RSM depth buffer

#

still without any visibility information there

#

but that's overkill in performance loss for just importance sampling

long robin Nov 3, 2022, 8:57 AM

#

sounds brainwormy

shell inlet Nov 3, 2022, 9:00 AM

#

it's as stupid of an idea as the screen space visibility test

#

which I actually made and it performed like poop

#

not to mention there were ugly ahh artifacts

#

inherent to screen space techniques

long robin Nov 3, 2022, 9:10 AM

#

depth rejection is hard

#

using a nearest sampler and a hard cutoff makes it reject distant stuff when you merely turn the camera (blue = samples rejected based on depth heuristic)

#

https://gfycat.com/WateryIllegalDrongo

Gfycat

ShareX

▶ Play video

#

I'm gonna try something

shell inlet Nov 3, 2022, 9:15 AM

#

you can try disabling depth rejection

#

I was wondering are you accumulating samples or filtered results as of what's currently on github?

long robin Nov 3, 2022, 9:17 AM

#

let me check

shell inlet Nov 3, 2022, 9:17 AM

#

I just looked into the branch latest commit and saw no changes to the main rsm shader

long robin Nov 3, 2022, 9:17 AM

#

yeah I made a new shader

shell inlet Nov 3, 2022, 9:17 AM

#

which is weird considering that you want new samples to come in every frame

long robin Nov 3, 2022, 9:17 AM

#

the monolithic shader was getting annoying

#

looks like the github version is accumulating samples

#

this comment even asserts it
https://github.com/JuanDiegoMontoya/Fwog/blob/examples-refactor/example/common/RsmTechnique.cpp#L150

shell inlet Nov 3, 2022, 9:28 AM

#

vec2 noise = rsm.random + textureLod(s_blueNoise, (vec2(gid) + 0.5) / textureSize(s_blueNoise, 0), 0).xy;
can you explain why you allow it to go beyond 1?

long robin Nov 3, 2022, 9:28 AM

#

no

#

is this where I do the magic of dividing it by 2

shell inlet Nov 3, 2022, 9:30 AM

#

you should demand new samples from hammersley instead

#

using frame count or something

#

but then you have N in it

#

vec2 Hammersley(uint i, uint N)

long robin Nov 3, 2022, 9:31 AM

#

yeah I need one of those infinite sequences

shell inlet Nov 3, 2022, 9:31 AM

#

would it work to put an arbitrarily big number in there

long robin Nov 3, 2022, 9:31 AM

#

no

#

because the samples originate from the center and create a spiral as i increases

shell inlet Nov 3, 2022, 9:32 AM

#

that is due to the circle map

long robin Nov 3, 2022, 9:32 AM

#

as shown in my famous visualization
https://www.shadertoy.com/view/7ssfWN

shell inlet Nov 3, 2022, 9:32 AM

#

hammersley is a low discrepancy sequence on a unit square

long robin Nov 3, 2022, 9:32 AM

#

yeah, but the main issue is the fact that it's ordered like that, no?

shell inlet Nov 3, 2022, 9:34 AM

#

a coincidence most likely

#

a fortunate one

#

the goal of lds is to have more efficiency in covering the ground on a unit square

#

or not just square

#

more efficient as opposed to white noise where there is a total chaos

long robin Nov 3, 2022, 9:35 AM

#

you can't choose a random contiguous subset of hammersley and get a uniform sampling of the unit square

shell inlet Nov 3, 2022, 9:35 AM

#

and you can sample same-ish place N times in a row by chance

long robin Nov 3, 2022, 9:38 AM

#

halton sequence seems to give the desired property
https://www.shadertoy.com/view/3dyyR3

Shadertoy

#

you can take any sequence in it and it seems to produce a nice result

shell inlet Nov 3, 2022, 9:39 AM

#

I see what you mean

long robin Nov 3, 2022, 9:40 AM

#

oof halton has a while loop in it

#

float halton(int base, int index)
{
    float result = 0.;
    float f = 1.;
    while (index > 0)
    {
        f = f / float(base);
        result += f * float(index % base);
        index = index / base; 
        //index = int(floor(float(index) / float(base)));
    }
    return result;
}

#

not ideal for GPU evaluation

shell inlet Nov 3, 2022, 9:45 AM

#

at this point maybe just rotate the blue noise

#

cranley patterson again

#

can't ruin blue noise any more than what it already is anyways

long robin Nov 3, 2022, 9:47 AM

#

I'm not sure what's wrong with adding it to the blue noise currently

shell inlet Nov 3, 2022, 9:47 AM

#

random numbers should be zero to one

#

actually yeah there is no difference

#

there is a mod in the end

#

when the actual rotation happens

long robin Nov 3, 2022, 9:48 AM

#

that's what I was hoping would happen to the noise

shell inlet Nov 3, 2022, 9:49 AM

#

I admit I did not think about it as much as I should have

long robin Nov 3, 2022, 9:50 AM

#

np

#

I got to learn something about halton sequences as a result

long robin Nov 3, 2022, 10:53 AM

#

probably gonna need TAA to deal with these flickering subpixel details 😩
https://gfycat.com/RigidOpulentAiredale

Gfycat

ShareX

▶ Play video

#

they flicker a lot with movement as well

#

the flickering is not as bad if I lower rMax

#

as more samples are concentrated into a smaller area

shell inlet Nov 3, 2022, 10:56 AM

#

see svgf

long robin Nov 3, 2022, 10:57 AM

#

fine 😂

#

I'll look at it at work tomorrow

#

I also think I need normal-based rejection on top of depth

#

I get funny "temporal contact shadows"

#

https://gfycat.com/SpeedySimpleAtlanticbluetang

Gfycat

ShareX

▶ Play video

#

I pushed the bs I was working on

#

the main new thing is a history length buffer which helps guide the temporal weight (which is not inspired by any paper so it's probably jank)

golden schooner Nov 3, 2022, 12:08 PM

#

i like these trippy effects

long robin Nov 4, 2022, 12:14 AM

#

@shell inlet I thought this was interesting in the SVGF paper

#

I'll have to try it

shell inlet Nov 4, 2022, 3:12 AM

#

that's like restir before restir

#

spatial reuse kind of

#

except without maths to correct bias

long robin Nov 4, 2022, 3:14 AM

#

They're living on the edge between empirical and physical correctness

shell inlet Nov 4, 2022, 3:15 AM

#

you haven't seen A-SVGF yet where they copypaste 1 random pixel between frames in a 3x3 tile to compute temporal gradients

long robin Nov 4, 2022, 3:16 AM

#

Huh

shell inlet Nov 4, 2022, 3:17 AM

#

technically this could be alleviated via using a separate render target that consists of old pixels to compute new lighting on but that's computationally more expensive

#

so they just add bias by copypasting pixels into new frame to save on performance

#

by pixel I mean all surface info

#

they want to keep every condition same and let only computed illumination change to compare against previous frame's result

#

that way they have a measure of how much lighting changed and how much of temporal history to drop

heavy cipher Nov 4, 2022, 7:22 AM

#

have you considered renaming the main branch to 'inspiration'

#

would automatically up the threat level of any fwog link

shell inlet Nov 4, 2022, 7:23 AM

#

idgi

long robin Nov 4, 2022, 7:33 AM

#

the joke is that martty is frightened by frogs

heavy cipher Nov 4, 2022, 8:33 AM

#

i just want fwog linkposting to be competitive with nabla

long robin Nov 4, 2022, 8:38 AM

#

nabla links gave me an unidentified disease

#

shell inlet Nov 4, 2022, 9:26 AM

#

heavy cipher i just want fwog linkposting to be competitive with nabla

contribute fun things to fwog and this will happen eventually

long robin Nov 4, 2022, 9:27 AM

#

actually I link fwog quite a bit now

#

shell inlet Nov 4, 2022, 9:28 AM

#

by the way I like how seemingly simple technique grew out of proportions into using modern ray tracing approaches to enhance it

long robin Nov 4, 2022, 9:28 AM

#

ye 😄

long robin Nov 4, 2022, 9:29 AM

#

long robin

unfortunately I still have a ways to go

shell inlet Nov 4, 2022, 9:30 AM

#

what if we use rsm for specular

long robin Nov 4, 2022, 9:31 AM

#

honestly quite amazing

shell inlet Nov 4, 2022, 9:31 AM

#

technically should be possible but very cursed

heavy cipher Nov 4, 2022, 9:31 AM

#

shell inlet contribute fun things to fwog and this will happen eventually

i have my own demons, tyvm

long robin Nov 4, 2022, 9:31 AM

#

shell inlet what if we use rsm for specular

my thoughts:

the OG paper said it would require many more samples to have specular
denoising

shell inlet Nov 4, 2022, 9:32 AM

#

more samples because yes

long robin Nov 4, 2022, 9:32 AM

#

yeah I haven't thought about why tbh

shell inlet Nov 4, 2022, 9:32 AM

#

you can turn it into specular by just evaluating specular brdf for each sample, that's it

#

of course most rays will sample outside of the spot where valuable samples are concentrated

long robin Nov 4, 2022, 9:33 AM

#

ah, it probably requires more samples because of that

#

most samples will contribute nothing

shell inlet Nov 4, 2022, 9:33 AM

#

"outside" of the "cone"

#

yeah

#

right now the brdf is 1/pi

#

uniform across the hemisphere

#

so all samples count

long robin Nov 4, 2022, 9:34 AM

#

now imagine a very glossy surface

#

now like 1% of the hemisphere is valid

shell inlet Nov 4, 2022, 9:34 AM

#

yeah lol

long robin Nov 4, 2022, 9:35 AM

#

and we don't even sample a hemisphere, but a weird projected disk

shell inlet Nov 4, 2022, 9:35 AM

#

but with more roughness more samples will count

#

what we sample doesn't matter

#Fwog and co.