Fwog and co. | Graphics Programming | Page 6

long robin Dec 7, 2022, 12:03 PM

#

tbh though, I think a shrimple screen-space denoiser would work fine for this

#

since the shadow is not sparse at all

golden schooner Dec 7, 2022, 12:03 PM

#

at least the jagged shadow is history

long robin Dec 7, 2022, 12:03 PM

#

once you start tracing actual rays is when a denoiser is needed

shell inlet Dec 7, 2022, 12:06 PM

#

but it's called denoiser not derayer

#

but I understand what you mean, you aren't warranting a variance guided filter because you don't even need to preserve penumbra if you don't even have one

#

it's a penumbruh

#

same width everywhere

#

but you know what

#

actually nvm it's more expensive maybe

long robin Dec 7, 2022, 12:12 PM

#

a 3x3 bilateral filter would probably remove most of the noise

shell inlet Dec 7, 2022, 12:13 PM

#

I was about to suggest cubic for shadow sampler

long robin Dec 7, 2022, 12:13 PM

#

maybe I could even go down to 1spp

shell inlet Dec 7, 2022, 12:13 PM

#

shadow sampler is bilinear so will work

#

maybe it won't be as slow as I think due to locality of taken samples but I don't know really

#

essentially a better 2x2 pcf

#

yes here I go again with my weird ideas

long robin Dec 7, 2022, 12:16 PM

#

you could try it in your local repo

#

just go to ShadeDeferredPbr.frag.glsl and update the shadow function

shell inlet Dec 7, 2022, 12:17 PM

#

no you do it

long robin Dec 7, 2022, 12:17 PM

#

and add your custom cubic thingy

shell inlet Dec 7, 2022, 12:17 PM

#

you have the code

long robin Dec 7, 2022, 12:17 PM

#

I already deleted the shadow sampler on mine

#

it's gone forever

shell inlet Dec 7, 2022, 12:17 PM

#

I am doing homework I am business

long robin Dec 7, 2022, 12:17 PM

#

well I gtg to bed

shell inlet Dec 7, 2022, 12:17 PM

#

goodnight

long robin Dec 7, 2022, 12:18 PM

#

gn

shell inlet Dec 7, 2022, 2:45 PM

#

took a break and tried it out, not much of an improvement

long robin Dec 7, 2022, 8:11 PM

#

Not a bad improvement imo

golden schooner Dec 7, 2022, 9:52 PM

#

looks almost as good as my shadow ;P

long robin Dec 8, 2022, 2:04 AM

#

after the depth bias was removed, ultra low RSM resolutions work fine (this is 128^2)

shell inlet Dec 8, 2022, 2:53 AM

#

nice

shell inlet Dec 8, 2022, 3:11 AM

#

what are those triangles on the green curtain

long robin Dec 8, 2022, 3:12 AM

#

shadow acne

#

the bias numbers I hardcoded for the shadows do not work in all scenarios

shell inlet Dec 8, 2022, 3:14 AM

#

have you heard about slope scale bias

long robin Dec 8, 2022, 3:15 AM

#

this is what I'm doing

  float constantFactor = 0.0002;
  float angleFactor = 0.001 * (1.0 - dot(-shadingUniforms.sunDir.xyz, normal));
  ...
  lightDepth += angleFactor + constantFactor;

shell inlet Dec 8, 2022, 3:16 AM

#

looks like slope scale, except cosine is not linear

#

angle is

#

cosine not

long robin Dec 8, 2022, 3:17 AM

#

what's the correct math for it

shell inlet Dec 8, 2022, 3:17 AM

#

I don't know

long robin Dec 8, 2022, 3:17 AM

#

these numbers work okay most of the time

#

I'm pretty sure there is an analytical way to compute the bias, but I'm not gonna tryhard it that much

shell inlet Dec 8, 2022, 3:17 AM

#

linearly changing value would be acos(dot)/pi

#

but I'm not sure whether it's required now

#

using cosine as a slope scale just feels strange to me because it won't grow properly with slope

long robin Dec 8, 2022, 3:19 AM

#

you're right, it fails at extreme angles

shell inlet Dec 8, 2022, 3:20 AM

#

I need to think what correct slope would be but I need to eat first

long robin Dec 8, 2022, 3:26 AM

#

https://jcgt.org/published/0003/04/08/

shell inlet Dec 8, 2022, 3:29 AM

#

2014

long robin Dec 8, 2022, 3:31 AM

#

https://www.gamedev.net/tutorials/programming/graphics/shadow-mapping-in-depth-analysis-r5424/

#

the second one isn't an analytical approach, but it does talk about some other stuff

#

kinda interesting

#

rip
float bias = 1.0 - clamp(dot(lightFragNormal, lightFragmentCoord), 0.0, 1.0);

#

stabilizing the shadow with temporal filtering is a good idea though

shell inlet Dec 8, 2022, 5:29 AM

#

https://www.desmos.com/calculator/thnub5raww

Desmos

slope scale bias

#

here is what you have now

#

https://www.desmos.com/calculator/kbju5ylz5l

Desmos

slope scale bias

#

this is what would make sense to me

#

inverse trig is banned on GPU tho so maybe there's a way to find acos using identities

#

uhh yeah I think we can find tan by using tan=sin/cos

#

can get sin from cross product and cos is the dot product

#

so it's len(cross(d,n))/dot(n,d)

#

when n,d are normalized

#

actually order of cross doesn't matter if we are taking magnitude of a vector

#

in 2d equivalent would be absolute value

#

https://www.desmos.com/calculator/nce35fulby

Desmos

slope scale bias

#

so here is the thing

#

I'm done

long robin Dec 8, 2022, 5:53 AM

#

btw linking this for later
https://gpuopen.com/wp-content/uploads/2019/08/RDNA_Architecture_public.pdf

shell inlet Dec 8, 2022, 6:02 AM

#

why are you rdna bombing

#

where are you going to need this

shell inlet Dec 8, 2022, 7:16 AM

#

edited graph added automatic bias for a given width

#

I think width would be 1/max(textureSize(shadowmap)) in the code

#

max texel width basically

long robin Dec 8, 2022, 7:20 AM

#

alright, I'll try it

shell inlet Dec 8, 2022, 7:20 AM

#

it does not account for quantization of the depth value itself, but I guess it would be 1/((2^precisionbits)*depthrange)

long robin Dec 8, 2022, 7:21 AM

#

for unorm formats I suppose

shell inlet Dec 8, 2022, 7:21 AM

#

ye

#

integer and ortho

#

perspective shadows are a pain

long robin Dec 8, 2022, 7:38 AM

#

seems like the technique works frogapprove

shell inlet Dec 8, 2022, 7:41 AM

#

peter panning under extreme angles

long robin Dec 8, 2022, 7:41 AM

#

ye

shell inlet Dec 8, 2022, 7:41 AM

#

is it worth it

long robin Dec 8, 2022, 7:41 AM

#

better than horrible acne probably

shell inlet Dec 8, 2022, 7:43 AM

#

bigger resolution should make peter panning smaller

long robin Dec 8, 2022, 7:44 AM

#

ye

#

still peter pans at the most extreme angles

#

I know a trick though

shell inlet Dec 8, 2022, 7:47 AM

#

mr. min()?

long robin Dec 8, 2022, 7:48 AM

#

float cosTheta = max(0.0, dot(incidentDir, normal) - 0.02);

#

just make the surface dark before the shadow starts peter panning hard

shell inlet Dec 8, 2022, 7:49 AM

#

😱

#

this is what variance/moments shadow maps aim to solve

long robin Dec 8, 2022, 7:51 AM

#

tfw leaking

shell inlet Dec 8, 2022, 7:52 AM

#

also that's a hard cutoff isn't it?

#

let me graph actually

long robin Dec 8, 2022, 7:53 AM

#

what's a hard cutoff

#

the cosTheta thing I'm doing is for the shading, not for the shadow bias

shell inlet Dec 8, 2022, 7:54 AM

#

yes I know

#

I mean that it'll go black abruptly

#

it'll reach 0.8 and after that it'll go 0

long robin Dec 8, 2022, 7:55 AM

#

I'm subtracting 0.02 from the whole thing, so it still smoothly goes to 0

#

the peak is being lost

#

here is with -0.05 bias

shell inlet Dec 8, 2022, 7:56 AM

#

oh yeah it'll go to 0 smoothly but it won't reach 1 then

long robin Dec 8, 2022, 7:57 AM

#

could use a smoothstep near 0 or something

shell inlet Dec 8, 2022, 7:57 AM

#

so you need to compensate with 1+0.02

#

1.02*dot(n,l)-0.02

long robin Dec 8, 2022, 7:59 AM

#

I'd actually have to multiply by 1.02040816 to compensate

#

huge difference

#

ah nvm your math is different lel

#

tbh I'm not gonna bother

#

the peter panning is only barely noticeable in this fake scene

shell inlet Dec 8, 2022, 8:02 AM

#

my math is wrong

#

you need (dot(n,l)-cutoff)/(1-cutoff)

#

division is to normalize the range [0; 1-cutoff] to [0;1]

long robin Dec 8, 2022, 8:05 AM

#

i c

shell inlet Dec 8, 2022, 8:05 AM

#

at this point we are straying further away from ~~god~~ pbr

long robin Dec 8, 2022, 8:06 AM

#

I've noticed a different artifact. Look at the base of the red shape

#

it's odd because you can see that the shadow is farther back

#

ah I think it's my fault

#

nvm

#

I see what the problem is

#

the edge of the floor box is casting a shadow

#

tl;dr not an actual issue

shell inlet Dec 8, 2022, 8:47 AM

#

what if you clamp the bias though would it really still cause acne?

#

min(max_bias, slope_scale_here)

long robin Dec 8, 2022, 8:49 AM

#

Yeah I tried earlier

shell inlet Dec 8, 2022, 8:49 AM

#

hmmm it will I think

#

ok confirmed then

#

what if we shift shadow sample uv along projected normal

#

at best 1 pixel to the side

#

would it solve acne at extreme angles

#

so that we could also reduce peter panning

#

or like check if tangent is higher than some threshold and then instead of adding depth bias shift the uv along the projected normal

#

you get what I mean?

long robin Dec 8, 2022, 8:57 AM

#

shell inlet what if we shift shadow sample uv along projected normal

You mean like projecting the surface normal into uv space?

shell inlet Dec 8, 2022, 8:58 AM

#

yes and using it as offset vector

long robin Dec 8, 2022, 8:58 AM

#

Ye I see how that could potentially help

shell inlet Dec 8, 2022, 8:59 AM

#

projecting it is taking difference (project(p+n) - project(p))

#

huh that is stupid

#

can use inversetranspose

#

transpose(inverse(sunviewproj))

#

I did not think this through it's a crazy idea again

#

so it might not work at all

long robin Dec 8, 2022, 9:03 AM

#

shell inlet Dec 8, 2022, 9:04 AM

#

maybe shifting it towards the inverted projection

#

cuz we want the pixel to be in shadow

#

maybe adding bias AND shifting towards inverted projection

#

this is what'd happen

#

maybe multiply the projection by pixel size also

shell inlet Dec 8, 2022, 9:09 AM

#

long robin

elaborate

#

did it not work?

long robin Dec 8, 2022, 9:12 AM

#

It was a response to your previous message about it being a crazy idea

shell inlet Dec 8, 2022, 9:12 AM

#

ic

#

it's probably stupid

shell inlet Dec 8, 2022, 3:34 PM

#

how did sponza change after adding perfect slope scale bias?

#

and does dithered shadow look better?

long robin Dec 8, 2022, 8:00 PM

#

shell inlet how did sponza change after adding perfect slope scale bias?

It looks pretty much the same overall, except the backside of the curtains are no longer brightly lit

long robin Dec 8, 2022, 8:01 PM

#

shell inlet and does dithered shadow look better?

They look approximately the same as before. That scene wasn't suffering from shadow mapping artifacts much

shell inlet Dec 9, 2022, 6:58 AM

#

tried hqx shadows?

#

https://en.wikipedia.org/wiki/Hqx

Hqx

In image processing, hqx ("high quality scale") is one of the pixel art scaling algorithms developed by Maxim Stepin, used in emulators such as Nestopia, FCEUX, higan, Snes9x, ZSNES and many more. There are three hqx filters defined: hq2x, hq3x, and hq4x, which magnify by factor of 2, 3, and 4 respectively. For other magnification factors, this ...

#

I mean obviously no but what if

long robin Dec 9, 2022, 7:01 AM

#

interesting

shell inlet Dec 9, 2022, 7:03 AM

#

there was a paper out there about vectorization of shadows too, this is similar but less fancy

#

similar in end result

#

this does not require bilinear taps

#

unlike bicubic

#

just 3x3 and a LUT for best results

#

this will likely reduce amount of pixelation

#

not as much as vectorized shadows tho

#

https://graphicsinterface.org/wp-content/uploads/gi2016-10.pdf

#

by the way about specular rsm

#

I think it's possible to do some sort of rejection sampling with reprojection to sample more efficiently

#

oh and probably restir too

#

or ris at least

#

I mean restir is built on top of ris so it's going to be a start

#

though it's going to require yet another temporal pass for both

#

and is probably overkill so maybe it's time to stop

long robin Dec 9, 2022, 7:17 AM

#

ye I'm about done with the svgf stuff

#

I just want to optimize it and gtfo

long robin Dec 9, 2022, 7:36 AM

#

they don't estimate those temporal gradients tho
https://youtu.be/qW6RVaIvZPk?t=118

shell inlet Dec 9, 2022, 9:50 AM

#

long robin they don't estimate those temporal gradients tho <https://youtu.be/qW6RVaIvZPk?t...

how'd you know?

long robin Dec 9, 2022, 9:51 AM

#

I estimated

shell inlet Dec 9, 2022, 9:51 AM

#

you would get penumbras all smudged without them

#

from direct illumination

long robin Dec 9, 2022, 9:53 AM

#

the penumbras do be smudged
https://youtu.be/lZQJU3yek08?t=366

#

well, it just looks smudgy in general

shell inlet Dec 9, 2022, 9:54 AM

#

I don't see it

#

I see temporal lag

#

which is not related to variance estimation

#

perhaps you mean temporal gradient

long robin Dec 9, 2022, 9:54 AM

#

that's wat I sayd

shell inlet Dec 9, 2022, 9:55 AM

#

damn what's wrong with me

#

I'm killing myself now

#

you are absolutely right

long robin Dec 9, 2022, 9:58 AM

#

I wonder why they didn't do it

#

I'd expect these people to use top tier denoising techniques

shell inlet Dec 9, 2022, 10:08 AM

#

there are other methods besides temporal gradients

#

perhaps they used that but it's being overwhelmed

#

look up history clamping in ReLAX

#

https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s32759/

NVIDIA

ReLAX: A Denoiser Tailored to Work with the ReSTIR Algorithm | NVID...

ReLAX originates from SVGF but takes image quality and performance to the next level

long robin Dec 9, 2022, 10:12 AM

#

april 2021 nice

shell inlet Dec 9, 2022, 10:15 AM

#

no heuristic beats temporal gradients imo

#

but temporal gradients either roughly double the cost (trace twice per pixel and keep previous frame around in its entirety), or add bias (pick 1 pixel in 3x3 strata and forward project them to be shaded together with the current frame)

#

ok can also do it in half/third resolution the first way, too

#

but that kills more pixels that may be valid

#

depending on how many pixels 1 gradient pixel covers

#

obviously

long robin Dec 9, 2022, 10:33 AM

#

shell inlet https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s32759/

this is brilliant tbh

shell inlet Dec 9, 2022, 11:16 AM

#

what is, exactly

#

don't tell me whole thing

#

also if you love nvidia so much why don't you work at nvidia

long robin Dec 9, 2022, 11:17 AM

#

da part where they use a fast history buffer

#

and clamp to neighborhood color like in TAA

shell inlet Dec 9, 2022, 11:18 AM

#

tbh I wish I was working at nvidia then I would call you my sworn rival

#

lmao that emote nonovidia

long robin Dec 9, 2022, 11:20 AM

#

I actually made that one lol

shell inlet Dec 9, 2022, 11:21 AM

#

I expected nothing less

long robin Dec 9, 2022, 11:21 AM

#

btw I think the fast history buffer would work great for indirect diffuse

#

also wtf that page turned chinese while I wasn't looking at it

rugged notch Dec 9, 2022, 11:22 AM

#

not very bing chilling of them

shell inlet Dec 9, 2022, 11:22 AM

#

ching cheng hanji moment

long robin Dec 9, 2022, 11:22 AM

#

it must have been autoplay or something because it was a different page than before

golden schooner Dec 9, 2022, 11:39 AM

#

whats the next topic for fwog? after svgtfo?

rugged notch Dec 9, 2022, 11:39 AM

#

finding love 😳

golden schooner Dec 9, 2022, 11:40 AM

#

he found love already, us, frogs

rugged notch Dec 9, 2022, 11:40 AM

#

golden schooner Dec 9, 2022, 11:40 AM

#

speaking of marzipan, are you working on a froge 2.0? @rugged notch

long robin Dec 9, 2022, 11:40 AM

#

maybe I will finally have the willpower to implement clustered light culling

#

nah jk KEKW

rugged notch Dec 9, 2022, 11:41 AM

#

golden schooner speaking of marzipan, are you working on a froge 2.0? <@249293982431969282>

sadly there was not much interest from my collaborator

golden schooner Dec 9, 2022, 11:41 AM

#

too bad

long robin Dec 9, 2022, 11:41 AM

#

https://tenor.com/view/collaboration-with-the-enemy-captain-joseph-piller-the-last-vermeer-collabing-witht-the-enemy-helping-the-enemy-gif-18987239

Tenor

rugged notch Dec 9, 2022, 11:41 AM

#

which is fair, she just did it on a whim and doesn't want the bother I guess

golden schooner Dec 9, 2022, 11:41 AM

#

maybe jumping spider is next 🙂 can you do some inception work

rugged notch Dec 9, 2022, 11:41 AM

#

https://tenor.com/view/spider-insects-gif-18920877

Tenor

golden schooner Dec 9, 2022, 11:42 AM

#

so cute 🙂

long robin Dec 9, 2022, 11:42 AM

#

that froge looks funny

long robin Dec 9, 2022, 12:03 PM

#

btw @shell inlet I think svgf's default of 0.2 for alpha is shrimply crazy

#

it's in the range of what this ReLAX technique considers to be "fast history"

#

#

kinda funny too cuz this path tracer with ReSTIR has insanely good sampling

#

so they don't even need a very low alpha

#

except maybe for specular, idk

shell inlet Dec 9, 2022, 12:05 PM

#

good sampling will leave you a 1spp image that is almost converged

#

so I don't think it's too crazy

long robin Dec 9, 2022, 12:05 PM

#

that's what I'm saying

#

they don't need the crutch of a low alpha like I do

shell inlet Dec 9, 2022, 12:05 PM

#

you overestimate restir

#

restir also needs a few frames to digest samples before it's good

long robin Dec 9, 2022, 12:06 PM

#

maybe the original SVGF paper is indeed being wacky by suggesting 0.2 for alpha

shell inlet Dec 9, 2022, 12:06 PM

#

first few frames are horrible

#

so much so that they need a disocclusion filter

long robin Dec 9, 2022, 12:06 PM

#

yeah

#

ok I think I see

#

so I guess restir is really good after it converges for a few frames

#

which is what you said

golden schooner Dec 9, 2022, 12:09 PM

#

so loading screen at the beginning for it to catch up a few frames think

long robin Dec 9, 2022, 12:10 PM

#

turn around
loading screen

shell inlet Dec 9, 2022, 12:10 PM

#

display loading screen at disoccluded regions

golden schooner Dec 9, 2022, 12:10 PM

#

😄

#

or you show the scene heavily blurred, and blur gets less and less until everything is loaded in

long robin Dec 9, 2022, 12:10 PM

#

they should show this DONT CARE texture from renderdoc

#

man wtf is that font

golden schooner Dec 9, 2022, 12:11 PM

#

hmm we have no :novulkan:

#

looks like some CJK font

#

MS Mincho or so

long robin Dec 9, 2022, 12:12 PM

#

ye

#

the thinness + consoleish typeface is the tell for me

golden schooner Dec 9, 2022, 12:13 PM

#

ye and the odd angles/serifes

#

i always wonder why the latin alphabet looks so shitty in the asian fonts

long robin Dec 9, 2022, 12:14 PM

#

some of them look decent
https://en.wikipedia.org/wiki/List_of_CJK_fonts

#

but I'm guessing the main issue is that it's a lot harder to create these fonts

golden schooner Dec 9, 2022, 12:15 PM

#

and not run into copyright issues 😉 with the other 2934824839845302482 fonts

long robin Dec 9, 2022, 12:15 PM

#

I wonder if there is a way to use different fonts for different parts of unicode

#

so you could use a non sucky latin font + a CJK font

golden schooner Dec 9, 2022, 12:16 PM

#

it is possible

#

you render the text runs shrimperately think

long robin Dec 9, 2022, 12:16 PM

#

you're gonna give me the text runs

golden schooner Dec 9, 2022, 12:16 PM

#

: D

shell inlet Dec 9, 2022, 12:19 PM

#

shell inlet good sampling will leave you a 1spp image that is almost converged

https://www.shadertoy.com/view/ldBcDt
simple scene with good sampling

Shadertoy

#

of course it gets harder as scene complexity grows

#

each bounce becomes harder to sample efficiently and this is where resampling excels

#

but at the price of needing a few samples to get the idea of sampling domain

#

here it's analytically tweaked to importance sample direct illumination but without taking visibility term into account because that's basically impossible to do analytically

#

resampling is the numerical way to do it

#

anyways that's where small history would suffice, and sponza is not a very challenging scene in terms of complexity (most of it is directly illuminated and the rest is illuminated by 1st bounce)

#

if you want a challenging scene, see my path_tracing_nightmare.glb

long robin Dec 9, 2022, 12:26 PM

#

Material getMaterial(int i) {
    if(i==0) return materials[0]; else
        if(i==1) return materials[1]; else
            if(i==2) return materials[2]; else
                if(i==3) return materials[3]; else
                    if(i==4) return materials[4]; else
                        if(i==5) return materials[5]; else
                            return materials[6];
    //return materials[i];    //webGL 2.0
}

#

send help

long robin Dec 9, 2022, 12:26 PM

#

shell inlet if you want a challenging scene, see my path_tracing_nightmare.glb

where

#

it'll probably look like ass with RSM but whatever

shell inlet Dec 9, 2022, 12:27 PM

#

you asked for it already bruh

#

and I sent it

long robin Dec 9, 2022, 12:27 PM

#

oh lmao

#

I wonder where I put it

shell inlet Dec 9, 2022, 12:28 PM

#

#1030797522177888316 message

#

also saw this again, still funny

long robin Dec 9, 2022, 12:29 PM

#

I must have forgotten to copy it to my models folder

shell inlet Dec 9, 2022, 12:30 PM

#

it's designed to progressively get harder as you go deeper into the room

#

because each bounce is harder to estimate

#

blender basically dies there

#

and denoiser gets all splotchy

#

in the furthest corner that is

#

essentially a scene to test your path guiding algo

long robin Dec 9, 2022, 12:34 PM

#

shadows be leaking

#

that bias

shell inlet Dec 9, 2022, 12:34 PM

#

this scene is too hard for rsm

#

looks good if you don't let anyone see what's under the 1st floor ceiling

long robin Dec 9, 2022, 12:36 PM

#

shell inlet in the furthest corner that is

the light gets into this corner just fine KEKW

#

filtered RSM is such an improvement in this scene though

shell inlet Dec 9, 2022, 12:39 PM

#

😎 👍

#

oh by the way

#

can you make sun intensity slider

#

or exposure

#

tbh you maybe wanna some tonemapping now

#

maybe warrants a tonemapper demo?

long robin Dec 9, 2022, 12:40 PM

#

I already got the

  // tone mapping (optional)
  //finalColor = finalColor / (1.0 + finalColor);

doe

#

jk jk

shell inlet Dec 9, 2022, 12:41 PM

#

call it hdr and implement all of those found here: https://64.github.io/tonemapping/

Tone Mapping

long robin Dec 9, 2022, 12:41 PM

#

I think I'm using some slightly better tonemapper for the volumetric example already

#

https://github.com/JuanDiegoMontoya/Fwog/blob/examples-refactor/example/shaders/volumetric/TonemapAndDither.comp.glsl

shell inlet Dec 9, 2022, 12:41 PM

#

well that was just an idea for another example which might be fun

#

I sure would like to try those out some time later for personal project

long robin Dec 9, 2022, 12:42 PM

#

apparently using the ACES tonemapper without a bunch of other stuff is a bad idea (according to other people on this discord)

#

idk what the other stuff is though, and imo it looks fine on its own

shell inlet Dec 9, 2022, 12:43 PM

#

I don't like aces because it's simply compressing everything to ldr

#

I want dynamic exposure change cuz its cool

long robin Dec 9, 2022, 12:43 PM

#

ah

#

I have implemented that twice now

shell inlet Dec 9, 2022, 12:44 PM

#

and overexposed parts that emit a glow with fft bloom

long robin Dec 9, 2022, 12:44 PM

#

maybe I could try this fancy thing sometime
https://www.shadertoy.com/view/XljBRK

#

but not today since I gotta schleep

dire badge Dec 9, 2022, 3:02 PM

#

shell inlet and overexposed parts that emit a glow with fft bloom

Is fft bloom finally being used?

shell inlet Dec 9, 2022, 3:03 PM

#

I don't have an answer to this question

#

the only thing I can talk for is myself

#

if you mean used in absolute general sense, then it never was unused

#

if you mean being used in production with real time graphics such as games - I don't know

#

the fact is, FFT is cheaper than naive convolution when you have large kernels

#

and diffraction patterns are generally very large

golden schooner Dec 9, 2022, 10:03 PM

#

did you take a look at https://github.com/tippesi/Atlas-Engine yet?

#

that link came across my eyes again and i though i show you 😄 then i realised i must have showed it to you already

#

long robin Dec 9, 2022, 10:20 PM

#

Ye I skimmed through the source the first time you showed me

golden schooner Dec 9, 2022, 10:25 PM

#

ah

#

and then you forgor

long robin Dec 9, 2022, 11:03 PM

#

what was I supposed to rembererer

golden schooner Dec 9, 2022, 11:08 PM

#

idk, i was just trying to be funny

#

they have ray and pathtracing going on

long robin Dec 9, 2022, 11:12 PM

#

ye it seems quite legit

golden schooner Dec 9, 2022, 11:15 PM

#

all done in openjayell

shell inlet Dec 9, 2022, 11:52 PM

#

so I installed win10 to play portal rtx

#

can confirm ghosting all around on any settings

#

sometimes even portal gun leaves trails when you move the camera

long robin Dec 10, 2022, 12:10 AM

#

muh broment

long robin Dec 10, 2022, 10:45 AM

#

separable 5x5 vs regular 5x5 bilateral (I adjusted the settings to make the spatial blur super obvious lol)

#

hmm I must have messed something up since the first looks so much smoother

#

or maybe the separable bilateral really do be like that

shell inlet Dec 10, 2022, 10:48 AM

#

is this truly 5x5

long robin Dec 10, 2022, 10:48 AM

#

it's 5 passes of 5x5

#

in the first image, I separated the filter so it alternates between horizontal and vertical passes

#

so it's technically 10 smaller passes

#

wait a sec

#

ah, I'm only using 4 passes in the second pic

#

here is how it should look

shell inlet Dec 10, 2022, 10:52 AM

#

uncanny how pattern around the plant is very similar

#

check this and previous separable bilateral screenshot

#

I almost thought you posted same one twice

#

or did you

long robin Dec 10, 2022, 10:53 AM

#

lol they're different, I promise

#

I put each image in a separate layer in paint.net so I can inspect

#

comparing images in chrome is painful

shell inlet Dec 10, 2022, 10:54 AM

#

they are totally same patterns

#

everywhere

#

except one is darker

long robin Dec 10, 2022, 10:54 AM

#

which pattern? the noise?

#

I made sure to fix the seed in both pics (and disable temporal accumulation)

#

so the only difference is the filter

shell inlet Dec 10, 2022, 10:55 AM

#

yeah I thought you just somehow got lucky capturing similar patterns lol

long robin Dec 10, 2022, 10:55 AM

#

dream luck

shell inlet Dec 10, 2022, 10:56 AM

#

by the way I forgot to mention that last time I ran the example there were some issues with random history drops in the distance

#

it's probably related to depth

#

when you move the camera it randomly drops history

long robin Dec 10, 2022, 10:57 AM

#

was it big sections or single pixels dropping history?

shell inlet Dec 10, 2022, 10:58 AM

#

big sections

#

like slices

long robin Dec 10, 2022, 10:58 AM

#

hmm, that artifact happens when the depth phi is too low

#

or anything else that causes excess sensitivity for that weight

#

I'm not seeing it on my machine, at least not in the scenes I tested

shell inlet Dec 10, 2022, 11:20 AM

#

make filter step as low as possible

#

will be easier to spot

long robin Dec 10, 2022, 11:21 AM

#

ye that's what I did

#

I also have a debug mode to spot disocclusions

#

in Reproject2.comp.glsl there is a #define at the top to toggle it

#

if I turn really fast I think I can see what you're talking about

#

shell inlet Dec 10, 2022, 11:25 AM

#

long robin Dec 10, 2022, 11:25 AM

#

ok that's weird

#

when was the last time you pulled?

shell inlet Dec 10, 2022, 11:26 AM

#

maybe you are not fixing depth values after reprojection

#

I think it won't be same depth if you reproject a pixel from the side to the center

long robin Dec 10, 2022, 11:27 AM

#

I wonder if the fact that I'm running at 240hz makes it harder to see it

#

it would make the depth differences smaller

#

hmm yeah I see it better at 60hz

shell inlet Dec 10, 2022, 11:30 AM

#

maybe instead of using depth values use camera distance

#

those should be invariant to fragment locations certainly

#

length(world_pos - view_pos)

#

or distance(world_pos, view_pos)

#

though when camera moves depth will also change

#

for same surface point between two frames

long robin Dec 10, 2022, 11:31 AM

#

it should probably be world pos then

#

I thought I could get away with linear depth, rip

#

btw the separable bilateral pass is ~1.2ms and the old one is ~2.8ms at 1080p

shell inlet Dec 10, 2022, 11:35 AM

#

I remember sultim also having issues with that

#

depth heuristics are a bit of a pita

long robin Dec 10, 2022, 11:42 AM

#

unfortunately the separable filter seems a bit worse in unstable areas

#

seems like there is more flickering in high frequency disoccluded areas like pot edges and leaves

#

I also tried a 7x7 filter for the lulz and there was 0 difference

#

its width probably made the depth phi too low for outer samples

dire badge Dec 10, 2022, 1:03 PM

#

golden schooner and then you forgor

and then you frogger

long robin Dec 10, 2022, 1:27 PM

#

I pushed my changes, including gui options for the sun specifically for void

long robin Dec 10, 2022, 2:05 PM

#

One trick I might try for improving perf/quality is to pre-downsample the RSM so it's cheaper to sample

#

Might even allow for multiple RSMs since sampling each one is relatively cheap

shell inlet Dec 10, 2022, 2:27 PM

#

long robin I pushed my changes, including gui options for the sun specifically for void

"options"?

#

did you add something more than just intensity?

#

another rotation axis maybe?

#

sun color?

long robin Dec 11, 2022, 6:47 AM

#

hacked up some code to make the RSM + denoiser run at quarter (0.5^2) res, now the whole pass takes ~1ms (it takes 1.8ms at full res)

#

it flickers really hard now

#

increasing the sample count is really cheap though, and it can reduce the temporal aliasing

#

though it would probably be ideal to use legit TAA

long robin Dec 11, 2022, 7:16 AM

#

long robin hacked up some code to make the RSM + denoiser run at quarter (0.5^2) res, now t...

tex hit rate goes down a lot in the filter pass, probably because I'm still sampling the full-res gbuffer in that

#

if I pre-downsample those to the target resolution first, perf may improve

#

sadly glBlitFramebuffer seems to be extremely slow. I use a blit when albedo modulation is disabled, and that alone takes 2ms

#

nsight seems to make the blit take a normal amount of time when profiling

long robin Dec 11, 2022, 10:18 AM

#

odd how there is no single-component 32-bit unorm texture format in gl

#

except for D32

shell inlet Dec 11, 2022, 10:20 AM

#

floatBitsToUint

#

uintBitsToFloat

#

wait unorm

#

is this reasonable

long robin Dec 11, 2022, 10:22 AM

#

for what I need, not really

shell inlet Dec 11, 2022, 10:23 AM

#

does float not lose a lot of info that you'd get to have in unorm 32 bit

long robin Dec 11, 2022, 10:23 AM

#

it gets converted to float anyways when you sample

#

what I wanted was to copy a D32 texture to an R32 unorm texture, but R32F will work exactly the same

shell inlet Dec 11, 2022, 10:25 AM

#

to sample?

#

in vulkan you probably could transition depth image to a readable layout and not copy things around

long robin Dec 11, 2022, 10:26 AM

#

you can sample depth textures directly in gl

#

what I'm doing is downsampling depth with a compute shader

shell inlet Dec 11, 2022, 10:26 AM

#

interesting haven't tried it

#

I remember d3d9 not allowing you to do that

long robin Dec 11, 2022, 10:26 AM

#

I can't write to depth formats so I have to use something else

shell inlet Dec 11, 2022, 10:27 AM

#

are you building depth pyramid

long robin Dec 11, 2022, 10:28 AM

#

I'm just seeing if an optimization idea will actually work

long robin Dec 11, 2022, 10:28 AM

#

long robin tex hit rate goes down a lot in the filter pass, probably because I'm still samp...

explained here

shell inlet Dec 11, 2022, 10:29 AM

#

naturally hitting more cached pixels yes

#

I anticipate costs going down

golden schooner Dec 11, 2022, 10:54 AM

#

long robin

i like it, i also like that the blue curtain kind of emits some volumetric looking looks

long robin Dec 11, 2022, 11:04 AM

#

it runs in about 0.7ms after downscaling the relevant buffers

golden schooner Dec 11, 2022, 11:12 AM

#

sounds fast

long robin Dec 11, 2022, 11:12 AM

#

ye but it's flickery AF now

#

I can mitigate it by cranking up the samples tho

long robin Dec 11, 2022, 11:29 AM

#

if I downsample the RSM to 128^2, I can take 40 samples and the whole pass still takes only 1ms

#

quality seems practically identical on sponza too

shell inlet Dec 11, 2022, 11:35 AM

#

you need something less than 30 series to truly test it

#

less even than 20 series

#

maybe some pascal

long robin Dec 11, 2022, 11:35 AM

#

maybe I can get deccer to try it again on his igpu

shell inlet Dec 11, 2022, 11:36 AM

#

was his igpu better than mine 710m

long robin Dec 11, 2022, 11:36 AM

#

idk

#

@golden schooner what was your perf the last time you ran my rsm example?

shell inlet Dec 11, 2022, 11:37 AM

#

what's ur igpu

golden schooner Dec 11, 2022, 11:37 AM

#

it was in the basement

#

intel hd 4600

#

yours is better

#

ill pull

long robin Dec 11, 2022, 11:38 AM

#

wait

#

I haven't pushed yet

shell inlet Dec 11, 2022, 11:38 AM

#

usermememark

long robin Dec 11, 2022, 11:38 AM

#

there is a bug with the regular RSM

#

golden schooner Dec 11, 2022, 11:38 AM

#

oi

#

i like the grid

shell inlet Dec 11, 2022, 11:39 AM

#

where is the bug

golden schooner Dec 11, 2022, 11:39 AM

#

wrong texture slot?

long robin Dec 11, 2022, 11:39 AM

#

I'm using a half res texture somewhere that a full res texture is expected, or vice versa

golden schooner Dec 11, 2022, 11:40 AM

#

building up some pressure

long robin Dec 11, 2022, 11:40 AM

#

sheeeit

#

somehow I went too far in the opposite direction

#

ah

shell inlet Dec 11, 2022, 11:42 AM

#

jaker plz add y axis sun rotation

long robin Dec 11, 2022, 11:43 AM

#

ok, after this push

shell inlet Dec 11, 2022, 11:43 AM

#

I added it some time ago on a keybind

#

but it was discarded

golden schooner Dec 11, 2022, 11:43 AM

#

shell inlet jaker plz add y axis sun rotation

thats a good idea, for my shtuff too

long robin Dec 11, 2022, 11:44 AM

#

ah I accidentally bound the wrong textures somewhere

shell inlet Dec 11, 2022, 11:45 AM

#

btw you can also add moving shadowmap for it to follow camera, with a pixel perfect snap so that it doesn't float as you move around

#

maybe that's more useful for deccer though

#

fwog examples have or will have AABB fit

#

I also think it's more reasonable to place the shadowmap origin in front of the view so that it doesn't render what's not going to contribute to the shading

golden schooner Dec 11, 2022, 11:48 AM

#

i have a button in the debug window which kinda follows the camera

long robin Dec 11, 2022, 11:49 AM

#

alright, bug fixed. now I just gotta write some commit messages

#

alright, pushed

golden schooner Dec 11, 2022, 11:54 AM

#

still examples branch neh?

long robin Dec 11, 2022, 11:55 AM

#

ye

shell inlet Dec 11, 2022, 11:55 AM

#

when are you going to merge to main

long robin Dec 11, 2022, 11:55 AM

#

never

shell inlet Dec 11, 2022, 11:55 AM

#

interesting decision

golden schooner Dec 11, 2022, 11:55 AM

#

8 fps

#

(deferred exshrimple)

shell inlet Dec 11, 2022, 11:56 AM

#

I had 30 fps on 710m

long robin Dec 11, 2022, 11:56 AM

#

when this rsm brainworm is "done" and I'm done with the other example stuff I will merge

long robin Dec 11, 2022, 11:56 AM

#

golden schooner (deferred exshrimple)

make sure to click the button to use filtered RSM

shell inlet Dec 11, 2022, 11:56 AM

#

maybe need to check out new version later since it's probably havier

#

heavier

golden schooner Dec 11, 2022, 11:56 AM

#

ah one sec

shell inlet Dec 11, 2022, 11:56 AM

#

no y rotation 😔

#

y no y

golden schooner Dec 11, 2022, 11:56 AM

#

29-31 it wobbles back and forth

long robin Dec 11, 2022, 11:56 AM

#

shell inlet no y rotation 😔

ye cuz I can't add it that quickly

golden schooner Dec 11, 2022, 11:57 AM

#

shell inlet Dec 11, 2022, 11:57 AM

#

filtered is faster, that is a success

long robin Dec 11, 2022, 11:57 AM

#

golden schooner 29-31 it wobbles back and forth

you can disable vsync by changing line 585 in deferred.cpp to this
auto appInfo = Application::CreateInfo{.name = "Deferred Example", .vsync = false};

#

seems like you won't get much higher than 30fps though

shell inlet Dec 11, 2022, 11:58 AM

#

you can add a checkbox for vsync

#

not very useful since you'd normally always want it on

#

otherwise gpu suffers

long robin Dec 11, 2022, 11:58 AM

#

yeah I don't want to clutter to UI too much

golden schooner Dec 11, 2022, 11:58 AM

#

long robin Dec 11, 2022, 11:59 AM

#

interactive framerates

shell inlet Dec 11, 2022, 11:59 AM

#

what is wrong with shadows

golden schooner Dec 11, 2022, 11:59 AM

#

some shadowmapisms with steep sun angles

long robin Dec 11, 2022, 11:59 AM

#

I clamped the bias

golden schooner Dec 11, 2022, 11:59 AM

#

https://tenor.com/view/clam-seafood-cockle-gif-24787566

long robin Dec 11, 2022, 11:59 AM

#

the 🅱️eter 🅱️anning was too intense if I let it run free

golden schooner Dec 11, 2022, 12:00 PM

#

ah it doesnt like the jiff 🙂

long robin Dec 11, 2022, 12:00 PM

#

lol

shell inlet Dec 11, 2022, 12:00 PM

#

discord embeds dying again

#

https://tenor.com/view/jesus-ballin-mars-bars-gif-19910027

Tenor

long robin Dec 11, 2022, 12:00 PM

#

tbf deccer's gif looked kinda phallic

golden schooner Dec 11, 2022, 12:00 PM

#

heavy cipher Dec 11, 2022, 12:01 PM

#

golden schooner Dec 11, 2022, 12:01 PM

#

that made me also chuckle a little

shell inlet Dec 11, 2022, 12:01 PM

#

golden schooner Dec 11, 2022, 12:02 PM

#

if i made #1019722539116802068 NSFW it should have worked

shell inlet Dec 11, 2022, 12:02 PM

#

clam too explicit

shell inlet Dec 11, 2022, 12:03 PM

#

long robin the 🅱️eter 🅱️anning was too intense if I let it run free

what about dither

long robin Dec 11, 2022, 12:03 PM

#

didn't feel like adding it for that scene

golden schooner Dec 11, 2022, 12:03 PM

#

jaker needs to goto bed first

shell inlet Dec 11, 2022, 12:03 PM

#

it's exclusive for gltf viewer?

long robin Dec 11, 2022, 12:03 PM

#

yeah

#

an incentive to look at it

golden schooner Dec 11, 2022, 12:04 PM

#

oh didnt we talk about reusable imgui controls?

#

looks like you are cutting corners again

long robin Dec 11, 2022, 12:04 PM

#

the only duplicated imgui controls atm are sun controls

golden schooner Dec 11, 2022, 12:04 PM

#

ok c:

long robin Dec 11, 2022, 12:04 PM

#

not gonna make a header for that

#

alright, the second sun angle slider has been added

#

somehow it kinda works when the RSM is downsampled to 32x32

#

really splotchy when the sun moves though

shell inlet Dec 11, 2022, 12:19 PM

#

is resolution scaling a slider?

long robin Dec 11, 2022, 12:19 PM

#

no

#

it can only be changed in the source

#

RsmTechnique.cpp line 104 is where the inverse resolution scale is hardcoded

#

2 = half res, 1 = full res

golden schooner Dec 11, 2022, 12:20 PM

#

maye void can coerce you into making it a slider one day, or sends a PR

long robin Dec 11, 2022, 12:20 PM

#

RsmTechnique.h line 102 is where you can change the downsampled RSM size

#

it would be cool, but annoying to add because it requires remaking all the textures

shell inlet Dec 11, 2022, 12:21 PM

#

would be easy to make with preprocessor and recompilation

golden schooner Dec 11, 2022, 12:22 PM

#

or just 2 methods, {Destroy/Create}SizeDependentTextures()

shell inlet Dec 11, 2022, 12:22 PM

#

I used to have a lot of preprocessor stuff and a recompile button in my ogl raytracer

long robin Dec 11, 2022, 12:22 PM

#

I just need to make a function to construct all the textures and it will be trivial to add those sliders

long robin Dec 11, 2022, 12:22 PM

#

shell inlet I used to have a lot of preprocessor stuff and a recompile button in my ogl rayt...

the shaders aren't what need to be recompiled

#

unless you're talking about cursed dll hot loadingisms

shell inlet Dec 11, 2022, 12:23 PM

#

no im talkin bout shader

long robin Dec 11, 2022, 12:23 PM

#

shaders don't need to be hot loaded atm

#

though a button to toggle debug mode would be kool

golden schooner Dec 11, 2022, 12:35 PM

#

Fwog the ultimate RSM testbench

long robin Dec 11, 2022, 12:54 PM

#

the effect looks decent at 0.25^2 res too

#

though I kinda have to fudge the sliders (super low alpha, increase samples to 8, decrease filter width a lot)

#

and the time seems to be dominated by downsampling the gbuffer and rsm at such a small resolution

#

maybe viable if you're on an iGPU 😄

golden schooner Dec 11, 2022, 12:58 PM

#

the shrimpled version really makes a difference, i think before your rehussle i had like 15fps or so

long robin Dec 11, 2022, 12:59 PM

#

if you look around, you can see artifacts from me naively upscaling the illuminance to screen resolution

golden schooner Dec 11, 2022, 1:00 PM

#

which example is that?

long robin Dec 11, 2022, 1:00 PM

#

this is void's path tracer hell scene

golden schooner Dec 11, 2022, 1:00 PM

#

a scene switcher would be neat

long robin Dec 11, 2022, 1:00 PM

#

on gltf viewer

golden schooner Dec 11, 2022, 1:00 PM

#

ah is that also part of the fwog repo now?

long robin Dec 11, 2022, 1:00 PM

#

not yet

shell inlet Dec 11, 2022, 1:01 PM

#

that is part of my engine

#

you can include it in fwog if you want, I wouldn't give the scene if I was against it

long robin Dec 11, 2022, 1:03 PM

#

(1/8)^2

long robin Dec 11, 2022, 1:03 PM

#

shell inlet you can include it in fwog if you want, I wouldn't give the scene if I was again...

I might modify it a little if I do

#

e.g., right now it is not centered at 0

#

or maybe it is

shell inlet Dec 11, 2022, 1:04 PM

#

the origin is somewhere inside of it

#

somewhere convenient I'm sure

long robin Dec 11, 2022, 1:04 PM

#

nvm it is well-centered

#

the problem was that I had a position for the camera hardcoded for sponza

#

how silly of me

shell inlet Dec 11, 2022, 1:05 PM

#

it's a scene measured in meters also so it might be bigger than sponza

#

way bigger I think actually

long robin Dec 11, 2022, 1:06 PM

#

I had to scale this one down 2x as much as sponza

shell inlet Dec 11, 2022, 1:06 PM

#

one cell is most likely a meter iirc

#

#

that's a meter probs

golden schooner Dec 11, 2022, 1:07 PM

#

i like these default textures

long robin Dec 11, 2022, 1:09 PM

#

(1/32)^2 resolution 😄

#

I'm taking 400 samples here and it's still performing 5x better than unfiltered RSM

shell inlet Dec 11, 2022, 1:10 PM

#

bigger rmax should worsen it

golden schooner Dec 11, 2022, 1:10 PM

#

maybe plain RSM is just "shit" compared to all the optimiziation shteps you did and do for the filtered one

long robin Dec 11, 2022, 1:10 PM

#

ah this is actually (1/16)^2

long robin Dec 11, 2022, 1:10 PM

#

golden schooner maybe plain RSM is just "shit" compared to all the optimiziation shteps you did ...

ye

shell inlet Dec 11, 2022, 1:10 PM

#

surprising but plain RSM is coherent but is shit in quality

#

it's faster than dithered sampling

#

the win here is that we only do 1spp and temporally accumulate, now even at low resolution

long robin Dec 11, 2022, 1:11 PM

#

lol 4000 samples is 1.5ms still

#

the whole texture must be in L1

golden schooner Dec 11, 2022, 1:11 PM

#

are you clamping it in code maybe?

long robin Dec 11, 2022, 1:11 PM

#

no

golden schooner Dec 11, 2022, 1:12 PM

#

crazy

long robin Dec 11, 2022, 1:12 PM

#

if I take 40k samples per pixel it goes up to 15ms

#

but these are not true screen space samples

#

I'm sampling from a texture that's 256x smaller than the screen in terms of area

shell inlet Dec 11, 2022, 1:13 PM

#

lol I noticed how you call sun angles

#

sun angle was so good they made sun angle 2

brisk narwhal Dec 11, 2022, 1:13 PM

#

if sun angle is so good why isn't there sun angle 3

#

checkmate

shell inlet Dec 11, 2022, 1:13 PM

#

because we adopt valve ideologies

shell inlet Dec 11, 2022, 1:17 PM

#

long robin I'm sampling from a texture that's 256x smaller than the screen in terms of area

your pixels are so big now they are basically almost surfels

long robin Dec 11, 2022, 1:17 PM

#

hehe

#

I really need to upsample it less stupidly

shell inlet Dec 11, 2022, 1:18 PM

#

you know that gives me another weird idea

brisk narwhal Dec 11, 2022, 1:19 PM

#

can you linearly filter the RSM

shell inlet Dec 11, 2022, 1:19 PM

#

what if we actually use surfels and do voronoi splatting

#

if rsm is so insensitive to resolution cuz it's low frequency

brisk narwhal Dec 11, 2022, 1:19 PM

#

shell inlet what if we actually use surfels and do voronoi splatting

oh that's big brained

long robin Dec 11, 2022, 1:19 PM

#

brisk narwhal can you linearly filter the RSM

I think it would be more ideal to apply a 3x3 bilateral as I upscale so it doesn't run over edges

shell inlet Dec 11, 2022, 1:20 PM

#

yes that's basically same as filtering

long robin Dec 11, 2022, 1:20 PM

#

I'm guessing it would fail hard in a scene with more complexity though

#

like with foliage and stuff

#

I mean low res sampling in general

shell inlet Dec 11, 2022, 1:21 PM

#

if you TAA maybe not as much

long robin Dec 11, 2022, 1:21 PM

#

ye

brisk narwhal Dec 11, 2022, 1:22 PM

#

At that point you could just switch to VXGI, it's probably simpler lol

long robin Dec 11, 2022, 1:22 PM

#

should also do subpixel jitter on the low res samples

golden schooner Dec 11, 2022, 1:23 PM

#

jaker needs to take a break from GI after this, and has other things to implement first, so that i can steal them for myisms

shell inlet Dec 11, 2022, 1:23 PM

#

we brought up voxel tracing some time ago maybe worth a try eventually

brisk narwhal Dec 11, 2022, 1:23 PM

#

The amount of optimizations that went in this RSM implementation is so good

#

I learned a lot just by lurking

shell inlet Dec 11, 2022, 1:24 PM

#

turns out fwog thread is entertaining and educational

long robin Dec 11, 2022, 1:25 PM

#

amen

#

not shown, but if I crank the samples up it's not as bad

shell inlet Dec 11, 2022, 1:26 PM

#

I wonder what is causing shimmers

long robin Dec 11, 2022, 1:26 PM

#

reprojection rejection

shell inlet Dec 11, 2022, 1:26 PM

#

history drop for sure but why is it droppong

#

lol typo

long robin Dec 11, 2022, 1:27 PM

#

probably cuz high frequency stuff is sampled too sparsely at low res

#

I'm gonna try someting

#

ah

shell inlet Dec 11, 2022, 1:27 PM

#

mayhaps it's a given that if pixel covers a lot of different surface points it should be less strict in reprojection rules

long robin Dec 11, 2022, 1:27 PM

#

I'm also not using the 3x3 bilateral reprojection heuristic that the original paper has

shell inlet Dec 11, 2022, 1:28 PM

#

I guess you could test if the reprojected pixel is still within the area that it covers downsampled

#

nah that's gonna give you a lot of troubles if you are accumulating the downsampled results

#

you actually need to accumulate full res

#

then it probably won't be as troublesome

#

is that what you're doing?

long robin Dec 11, 2022, 1:30 PM

#

no

#

the reprojection pass would get more expensive if I accumulate at full res

#

possibly worth it though

shell inlet Dec 11, 2022, 1:30 PM

#

but that's to assure it's not failing for the final image

long robin Dec 11, 2022, 1:31 PM

#

at the same time, just adding more samples is incredibly cheap and works well

golden schooner Dec 11, 2022, 1:31 PM

#

would be neat to have a switch/slider somehow to change the resolution to 1/16 to 1/1 or so, then we could also run some experiments on how it performs on all sorts of hardware again

long robin Dec 11, 2022, 1:31 PM

#

I'll do it tomorrow

#

but now I gotta slep

shell inlet Dec 11, 2022, 1:32 PM

#

ok if you slep I homwork

#

before I forget here's what voronoi splatting is: https://www.shadertoy.com/view/NlKGWt

Shadertoy

#

you can also hash it to quickly find closest ones for a given pixel

brisk narwhal Dec 11, 2022, 1:38 PM

#

given that the RSM sits entirely in cache at low resolutions

#

What if we could irregular Z-buffer the RSM

#

At higher resolutions

long robin Dec 11, 2022, 1:45 PM

#

I did the math and a 256^2 RSM is about 5x larger than the L1 cache of a single SM on my GPU (3070), but there are dozens of SMs and each pixel won't need to access the whole RSM

long robin Dec 11, 2022, 1:46 PM

#

brisk narwhal What if we could irregular Z-buffer the RSM

I'd have to look it up after I sleep

long robin Dec 12, 2022, 9:37 AM

#

mfw the original paper for irregular z-buffers notes that they cannot be constructed in real time on current (~2005) hardware, so they propose a GPU that can do it efficiently okey

#

the idea of using an irregular z-buffer for pixel perfect shadows is interesting though

long robin Dec 12, 2022, 10:36 AM

#

I love writing code like this

if (useSeparableFilter)
{
  Fwog::Cmd::BindSampledImage(0, indirectFilteredTex, nearestSampler);
}
else
{
  Fwog::Cmd::BindSampledImage(
    0,
    (5 - int(std::log2f(float(inverseResolutionScale)))) % 2 == 0 ? indirectUnfilteredTex : indirectFilteredTex,
    nearestSampler);
}

#

it's not brittle at all and doesn't lead to many bugs

heavy cipher Dec 12, 2022, 10:43 AM

#

lmao that condition

rugged notch Dec 12, 2022, 10:45 AM

#

long robin I love writing code like this ```cpp if (useSeparableFilter) { Fwog::Cmd::Bind...

get a pipebomb for christmas in these 4-7 simple steps, hard to count

long robin Dec 12, 2022, 10:46 AM

#

ping-pong algorithms with a variable number of iterations give me cancer

#

6 levels of indentation removed for your viewing pleasure

for (int i = 0; i < 5 - std::log2f(float(inverseResolutionScale)); i++)
{
  filterUniforms.stepWidth = (1 << i) * spatialFilterStep;
  filterUniformBuffer.SubDataTyped(filterUniforms);

  // The output of the first filter pass gets stored in the history
  const Fwog::Texture* in{};
  const Fwog::Texture* out{};
  if (i == 0)
  {
    in = &indirectFilteredTex;
    out = &indirectUnfilteredTexPrev;
  }
  else if (i == 1)
  {
    in = &indirectUnfilteredTexPrev;
    out = &indirectUnfilteredTex;
  }
  else if (i % 2 == 0)
  {
    in = &indirectUnfilteredTex;
    out = &indirectFilteredTex;
  }
  else
  {
    in = &indirectFilteredTex;
    out = &indirectUnfilteredTex;
  }

  Fwog::Cmd::BindSampledImage(0, *in, nearestSampler);
  Fwog::Cmd::BindImage(0, *out, 0);
  Fwog::Cmd::MemoryBarrier(Fwog::MemoryBarrierAccessBit::TEXTURE_FETCH_BIT);
  Fwog::Cmd::Dispatch(numGroups.x, numGroups.y, 1);
}

heavy cipher Dec 12, 2022, 10:54 AM

#

also replace int(log2f) with an int log

long robin Dec 12, 2022, 10:55 AM

#

I thought about it

#

is there actually int log in the C++ standard library?

#

I don't see it in <cmath>

heavy cipher Dec 12, 2022, 10:57 AM

#

i think not

long robin Dec 12, 2022, 10:57 AM

#

maybe it's in <bit>

rugged notch Dec 12, 2022, 10:57 AM

#

reinterpret cast your int's address to a non-trivial type

#

you'll get its log

#

sometimes

heavy cipher Dec 12, 2022, 10:58 AM

#

ah right, i remember now

long robin Dec 12, 2022, 10:58 AM

#

https://en.cppreference.com/w/cpp/numeric/countl_zero

rugged notch Dec 12, 2022, 10:58 AM

#

TIL about <bit>

heavy cipher Dec 12, 2022, 10:58 AM

#

its in <iostream>

int x = 6;
std::clog << x;

rugged notch Dec 12, 2022, 10:59 AM

#

https://www.youtube.com/watch?v=KBWuDhIv9B0

YouTube

EditorialDiscussions

Alan Partridge - Great Banter

Great Banter shirts: https://kek.gg/u/36dhq

▶ Play video

long robin Dec 12, 2022, 10:59 AM

#

making my code unreadable for epsilon perf gain smart

heavy cipher Dec 12, 2022, 11:00 AM

#

long robin making my code unreadable for epsilon perf gain <:smart:591864977296588830>

ilog2(x) is not unreadable 🤔

long robin Dec 12, 2022, 11:00 AM

#

ah I actually want this function
https://en.cppreference.com/w/cpp/numeric/countr_zero

long robin Dec 12, 2022, 11:01 AM

#

heavy cipher ilog2(x) is not unreadable 🤔

ilog2 is readable, but countr_zero is not unless you already know what it's being used for

heavy cipher Dec 12, 2022, 11:01 AM

#

well yeah

long robin Dec 12, 2022, 11:01 AM

#

too bad there is no way to fix that

rugged notch Dec 12, 2022, 11:01 AM

#

long robin making my code unreadable for epsilon perf gain <:smart:591864977296588830>

my brother in christ, you wrote the Fwog::Cmd::BindSampledImage(0, (5 - int(std::log2f(float(inverseResolutionScale)))) % 2 == 0 ? indirectUnfilteredTex : indirectFilteredTex, nearestSampler);

heavy cipher Dec 12, 2022, 11:02 AM

#

so don't write countr_zero instead of ilog2

long robin Dec 12, 2022, 11:02 AM

#

rugged notch my brother in christ, you wrote the ` Fwog::Cmd::BindSampledImage(0, (5 - int(s...

inshallah that code is ""temporary""

rugged notch Dec 12, 2022, 11:04 AM

#

long robin inshallah that code is ""temporary""

He will be refactored alhamdullah

heavy cipher Dec 12, 2022, 11:06 AM

#

mfw when HR tells me "declaring jihad on warnings" is not appropriate workplace language

golden schooner Dec 12, 2022, 7:43 PM

#

bigfrog

golden schooner Dec 13, 2022, 1:16 AM

#

GL_GOOGLE_include_directive doesnt work on nvidia-470xx

long robin Dec 13, 2022, 1:16 AM

#

it should be a warning at worst

golden schooner Dec 13, 2022, 1:18 AM

#

it gtfs me out

long robin Dec 13, 2022, 1:18 AM

#

enable: Causes the named extension to work; if the implementation does not support the extension, it only gives a warning. Fails if used with all.

golden schooner Dec 13, 2022, 1:18 AM

#

/home/deccer/Private/Code/External/Fwog/build/example/02_deferred                                 ✔ 
terminate called after throwing an instance of 'Fwog::ShaderCompilationException'
  what():  Failed to compile shader source.
0(3) : warning C7508: extension GOOGLE_include_directive not supported
0(118) : error C1020: invalid operands to "/"
0(119) : error C1020: invalid operands to "/"

long robin Dec 13, 2022, 1:18 AM

#

ok so the error is the implicit float conversion thing

golden schooner Dec 13, 2022, 1:18 AM

#

ah ye 🙂

long robin Dec 13, 2022, 1:19 AM

#

deja vu from when void encountered the exact same thing

golden schooner Dec 13, 2022, 1:19 AM

#

indeed

long robin Dec 13, 2022, 1:19 AM

#

however, the driver is incorrect here

#

for binary operators, if one operand is a float type and the other is an integer type, the integer one should be promoted to the float type

golden schooner Dec 13, 2022, 1:20 AM

#

i cannot use a newer driver

long robin Dec 13, 2022, 1:20 AM

#

I guess I'll have to change it so it works on y'alls' crusty drivers

golden schooner Dec 13, 2022, 1:21 AM

#

ye 🙂

#

unless you have us sponsored the latest amd goodness, i wouldnt say no lol

long robin Dec 13, 2022, 1:21 AM

#

hehe

#

I don't think I can let other people use my employee rebate sadly

golden schooner Dec 13, 2022, 1:22 AM

#

nah i was just kidding

#

i dont think i will get me an amd card ;P

#

i was able to get one of my 780tis working again

long robin Dec 13, 2022, 1:24 AM

#

did you bake it

golden schooner Dec 13, 2022, 1:24 AM

#

xD i did

long robin Dec 13, 2022, 1:24 AM

#

amazin

golden schooner Dec 13, 2022, 1:24 AM

#

quite the hassle

shell inlet Dec 13, 2022, 6:38 AM

#

long robin deja vu from when void encountered the exact same thing

Told ya

long robin Dec 13, 2022, 6:38 AM

#

it's a driver bug on god

shell inlet Dec 13, 2022, 6:39 AM

#

It was always there since 200X

long robin Dec 13, 2022, 6:39 AM

#

worksonmymachine

#

it actually works on both of my machines (rtx 3070 and rx 6800 with latest drivers, windows 10)

#

I can even pinpoint the exact part of the spec that says ur driver is bein silly

#

either way, I gotta conform to the lowest denominator 😩

shell inlet Dec 13, 2022, 6:41 AM

#

Would be fubby if i update mine and it stll wont work

#

I have 527.56 apparently

#

it's the latest I think

#

but I had to update nvidia geforce experience just to look it up

#

I'm now going to try your examples

long robin Dec 13, 2022, 6:57 AM

#

I have the previous driver version from that

shell inlet Dec 13, 2022, 7:04 AM

#

why do you have to explicitly specify if gltf is binary or not

long robin Dec 13, 2022, 7:05 AM

#

too lazy to check the extension automatically 🤰

#

well, I shouldn't have to check the extension

shell inlet Dec 13, 2022, 7:06 AM

#

std::filesystem::path::extension

long robin Dec 13, 2022, 7:06 AM

#

I mean there should be a magic number in the file to indicate it

#

only glbs have a magic number though

#

but yeah I guess I should check the extension

shell inlet Dec 13, 2022, 7:10 AM

#

it works

#

so they actually did fix that just in 2022

#

and both amd and nvidia decided to do that in 2022?

#

well can't say for amd

#

but my drivers weren't year old

#

this is one of the reasons spirv is superior

#

compile to ir yourself and don't worry about driver not being able to compile

#

or idk link your app to compiler

long robin Dec 13, 2022, 7:21 AM

#

ironically, opengl drivers are extra buggy when it comes to spirv compilation

shell inlet Dec 13, 2022, 7:22 AM

#

but was spirv made for ogl at all

#

also I wonder if we'll get another ogl version

#

it seems to have been abandoned

long robin Dec 13, 2022, 7:24 AM

#

it ded 💀

shell inlet Dec 13, 2022, 7:25 AM

#

ur a necromancer then

#

at least we're still getting updates that fix stuff

shell inlet Dec 13, 2022, 8:13 AM

#

long robin it's a driver bug on god

btw doesn't deccer have intel

#

so intel too...

long robin Dec 13, 2022, 8:20 AM

#

pretty sure that was actually his nv card

heavy cipher Dec 13, 2022, 8:22 AM

#

eastern thinkers say 'kalpa' - a moment of eternity - passes when deccer buys a new gpu

shell inlet Dec 13, 2022, 9:21 AM

#

if ur so smart why aren't you rich

#

(C) every mom ever

golden schooner Dec 13, 2022, 11:15 AM

#

yep im on intel usually, but i was able to resurrect 1 of the two 780tis last night

long robin Dec 13, 2022, 1:46 PM

#

I just had a thought

#

What if I could dynamically distribute samples like in surfels GI

#

It would be cool to do the RSM at a really low res, but put more samples (or move the sample location) on edges

#

Maybe just something like VRS would work, hmm

#

I want to put jittered samples on edges so high frequency geometry doesn't die

#

I guess aggressive VRS might work

#

Given how we've seen that downsampling the RSM effect to like (1/32)^2 works

#

But then there's still the problem of filtering, which I don't want to do at full res (as VRS still resolves to a full res image)

#

I gotta try the bilateral filter upscaling strategy first though

golden schooner Dec 13, 2022, 6:27 PM

#

sounds like its worth try

golden schooner Dec 13, 2022, 11:51 PM

#

@long robin that Limbosuccessor gdc talk was really cool, i find the game itself utterly boring, but the graphicsisms look intermeresting

#

mayhaps you can post the link in #graphics-resources when you find the time

long robin Dec 13, 2022, 11:53 PM

#

I can't post there

golden schooner Dec 13, 2022, 11:59 PM

#

oi

#

ah

#

its gp-news where everyone can

#

let me allow mvps to post in gp res too

#

should work now?

long robin Dec 14, 2022, 12:01 AM

#

Still can't

golden schooner Dec 14, 2022, 12:01 AM

#

now

#

i forgor to hit save

long robin Dec 14, 2022, 12:03 AM

#

Worked

golden schooner Dec 14, 2022, 12:05 AM

#

no embed?

long robin Dec 14, 2022, 12:06 AM

#

Idk why

#

They also have one about TAA
https://youtu.be/2XXS5UyNjjU

YouTube

GDC

Temporal Reprojection Anti-Aliasing in INSIDE

In this 2016 GDC talk, Playdead's Lasse Jon Fuglsang Pedersen discusses Temporal Reprojection Anti-Aliasing in the context of INSIDE, touching on the process, the initial research, and the pleasant side-effects.

Register for GDC: http://ubm.io/2gk5KTU

Join the GDC mailing list: http://www.gdconf.com/subscribe

Follow GDC on Twitter: https://...

▶ Play video

#

Lol nothing embeds in that channel

golden schooner Dec 14, 2022, 12:07 AM

#

ah wait

#

try again please 🙂

long robin Dec 14, 2022, 12:08 AM

#

Ez

golden schooner Dec 14, 2022, 12:08 AM

#

merci mon 🪑

long robin Dec 14, 2022, 12:09 AM

#

The triangle noise technique was probably my favorite thing from the first talk lol

golden schooner Dec 14, 2022, 12:20 AM

#

yeah

#

reminded me of how you helped me with my shadowisms 🙂

long robin Dec 14, 2022, 2:19 AM

#

I need a scene with high geometric complexity for testing RSM

#

something with foliage, perhaps

shell inlet Dec 14, 2022, 7:45 AM

#

spawh a bunch of randomly colored cubes

#

inside sponza

long robin Dec 14, 2022, 7:59 AM

#

why does the sketchfab's search suck so bad

#

the minimum price in the filter is $3 bruh

#

can't even filter for gltfs bruh

#

ok I think I found the free gltfs
https://sketchfab.com/3d-models?features=downloadable&sort_by=-likeCount

#

I had to use google instead of sketchfab's 💩 search functionality

#

lmao I literally clicked the "free gltf assets" button on their site and was presented with paid non-gltf models

#

blender is taking literal minutes to open a 1gb obj file

#

already using 11gb of ram

#

epic

#

#

shell inlet Dec 14, 2022, 8:39 AM

#

hey I was thinking about morgan MC guire samples with foliage and forgor it had miguel san

long robin Dec 14, 2022, 8:39 AM

#

yeah I resorted to those because the websites for sharing models suck

shell inlet Dec 14, 2022, 8:39 AM

#

does this work as a complex scene?

long robin Dec 14, 2022, 8:40 AM

#

it certainly makes some problems more obvious

#

like the fact that I need to do the 3x3 search in reprojection

shell inlet Dec 14, 2022, 8:40 AM

#

looking at pixellated trees I'd say it's a success

#

the trees have been censored

long robin Dec 14, 2022, 10:42 AM

#

the upscaling filter is something of a success

#

geometry no longer gets censored when only illuminated by indirect lighting

shell inlet Dec 14, 2022, 11:05 AM

#

that is amazing, you win

long robin Dec 14, 2022, 11:06 AM

#

it's way too unstable at that res though, even with the smooth upscaling

#

2x upscaling seems a lot more feasible

#

just gotta perfect the reprojection pass heee

brisk narwhal Dec 14, 2022, 11:07 AM

#

long robin

very pog

long robin Dec 14, 2022, 11:09 AM

#

persistent world-space surfels would be cool for this though

#

that would completely change filtering though

shell inlet Dec 14, 2022, 11:10 AM

#

yes you could also make use of software rt using them

#

cuz less rays would need to be traced for surfels

long robin Dec 14, 2022, 11:10 AM

#

yeah

#

I forgor how surfel lighting is actually applied to the image though. Is it just a lookup into the surfel structure?

shell inlet Dec 14, 2022, 11:11 AM

#

there is a paper by ea called gibs

#

global illumination based on surfels

long robin Dec 14, 2022, 11:11 AM

#

I saw the talk

#

didn't know there was a paper

shell inlet Dec 14, 2022, 11:12 AM

#

well you gotta splat them

#

some edge aware filter I'd imagine

#

wide one

long robin Dec 14, 2022, 11:13 AM

#

I thought they just used the world space pos of the surface being shaded to look into the surfel acceleration structure

#

I think they had an improvement on top of that tho

#

some kind of interpolation iirc

shell inlet Dec 14, 2022, 11:13 AM

#

I didn't look into how they do it specifically so I don't know

long robin Dec 14, 2022, 11:15 AM

#

I think this is it
https://youtu.be/h1ocYFrtsM4?t=668

#

they fetch N surfels from the cell and weight by distance and normal

shell inlet Dec 14, 2022, 11:16 AM

#

well isn't it a voronoi splatting

long robin Dec 14, 2022, 11:16 AM

#

the surfel locations?

shell inlet Dec 14, 2022, 11:17 AM

#

weighting by distance is essentially saying mixing with a convex combination

#

not sure what's with normal but I guess in the end it's all for a single weight in a convex combination

#

and if you have a set of points that you interpolate based on distance that's basically voronoi

long robin Dec 14, 2022, 11:18 AM

#

I see

#

when I googled voronoi splatting the first time you mentioned it, your shadertoy thing was the only relevant result lmao

shell inlet Dec 14, 2022, 11:19 AM

#

I guess I coined the term

#

but there really isn't any other way to call it that makes sense

long robin Dec 14, 2022, 11:20 AM

#

it do be like that

#

convex combination is another one of those scary terms

shell inlet Dec 14, 2022, 11:21 AM

#

idea is simple though

long robin Dec 14, 2022, 11:21 AM

#

ye https://en.wikipedia.org/wiki/Convex_combination

shell inlet Dec 14, 2022, 11:21 AM

#

essentially a constraint for a set such that a + b + c ... = 1

#

a way to achieve is to divide each element by the sum

long robin Dec 14, 2022, 11:22 AM

#

Weighted means are functionally the same as convex combinations, but they use a different notation. The coefficients (weights) in a weighted mean are not required to sum to 1; instead the weighted linear combination is explicitly divided by the count of the weights.

shell inlet Dec 14, 2022, 11:23 AM

#

example of a convex combo is barycentric coords

long robin Dec 14, 2022, 11:23 AM

#

many image filters are convex combinations through the above definition

shell inlet Dec 14, 2022, 11:24 AM

#

well it's a condition for a discrete pdf that it sums to 1

#

wait that's irrelevant

long robin Dec 14, 2022, 11:24 AM

#

the term "convex combination" has a geometric implication imo

shell inlet Dec 14, 2022, 11:24 AM

#

kernels just have the same condition

#

blur kernels

#

you can have whatever kernel you want for a convolution

long robin Dec 14, 2022, 11:25 AM

#

yeah

#

the bilateral filters I'm using need to sum to 1

shell inlet Dec 14, 2022, 11:26 AM

#

random fact did you know that we divide by pdf because estimators otherwise converge to it if you integrate 1

#

I guess you can numerically get a pdf using this property

long robin Dec 14, 2022, 11:27 AM

#

that's a neat fact

#

I'll try to remember that the next time I need a pdf

shell inlet Dec 14, 2022, 11:28 AM

#

it's not really practical in realtime

#

maybe not at all?

long robin Dec 14, 2022, 11:29 AM

#

I mean just making a fake scene to derive the pdf, then plugging in the number into the code

shell inlet Dec 14, 2022, 11:29 AM

#

maybe RIS is actually exploiting this property though 🤔

#

I remember they have some weight using expected value, but maybe it's unrelated

long robin Dec 14, 2022, 11:29 AM

#

if your pdf is not constant then my idea does not work

shell inlet Dec 14, 2022, 11:33 AM

#

you could make a LUT for numerically estimated pdf

#

but the dimension of the LUT depends on the number of parameters of your sampling function

#

maybe if some parameters are degenerate you could exclude a dimension

long robin Dec 14, 2022, 11:34 AM

#

I'd only consider it if the lut was 0-dimensional tbh

shell inlet Dec 14, 2022, 11:34 AM

#

that would be an approximation that causes bias for a difficult continuous pdf

#

depending on the LUT resolution

#

well not to mention that fat LUTs are lethal for performance

#

otherwise I'm sure we'd be using LUTs for BRDFs

#

there are captured real world BRDF LUTs out there

#

what was it called I can't remember

#

basically 100 materials library

long robin Dec 14, 2022, 11:37 AM

#

idk I'm not an artist

shell inlet Dec 14, 2022, 11:37 AM

#

MERL BRDF database

long robin Dec 14, 2022, 11:37 AM

#

yw

shell inlet Dec 14, 2022, 11:37 AM

#

https://www.merl.com/brdf/

MERL – Mitsubishi Electric Research Laboratories

MERL - Mitsubishi Electric Research Laboratories

#

those are very very fat

#

the only use is pretty much research

#

to compare your analytical model against

long robin Dec 14, 2022, 11:39 AM

#

neat

#

wow that 3x3 reprojection filter makes a difference

shell inlet Dec 14, 2022, 11:44 AM

#

you have mastered reprojection?

long robin Dec 14, 2022, 11:45 AM

#

I still haven't tried a bicubic filter like in that ReLAX talk

#

I won't try it tbh since I doubt it'll improve this much

heavy cipher Dec 14, 2022, 11:45 AM

#

when did we take damage

long robin Dec 14, 2022, 11:46 AM

#

by looking at the aliasing in the pic

#

it works pretty well even at 4x4 upscaling

#

still flickers a lot since there aren't enough samples placed on different surfaces

#

I want to temporally jitter the samples somehow, but idk if it's feasible

#

foliage hell

golden schooner Dec 14, 2022, 12:14 PM

#

burn it

#

i really dont like that scene either, from an aeschtethical perspective

long robin Dec 14, 2022, 12:18 PM

#

it's kinda ugly, but has useful geometry

#

sometimes there is black and I'm not sure why

shell inlet Dec 14, 2022, 12:19 PM

#

from what pass

long robin Dec 14, 2022, 12:19 PM

#

not sure

#

maybe the new thing I added to reprojection

shell inlet Dec 14, 2022, 12:19 PM

#

is it consistent

long robin Dec 14, 2022, 12:19 PM

#

somewhat

#

it seems to happen when there is a rejection

#

but not always

#

pretty sure I just got a bug somewhere

#

rejection seems to make areas get darker

#

it's odd because the affected areas have a long history

#

actually, no

#

low history seems to be correlated with darkening in general

#

I think it's just a lack of blurring + sparse sampling

#

except it still happens when there are 40 samples

#

that's something

#

probably using an index meant for a larger texture or vice versa

#

something is definitely wack, cuz it happens with 400 samples and when I only move by a pixel

#

I'm guessing some kind of UB cuz it doesn't happen in RenderDoc

shell inlet Dec 14, 2022, 1:11 PM

#

#

real question why can I see albedo when I skip modulation

long robin Dec 14, 2022, 1:11 PM

#

because it still does primary shading on the image

#

the "skip albedo modulation" is just for the output of the RSM pass, which means the indirect light

shell inlet Dec 14, 2022, 1:12 PM

#

makes sense

#

another thing is that it feels like sun intensity somehow doesn't make sense with what I see on screen

#

feels like it should be way more overblown for 40

long robin Dec 14, 2022, 1:15 PM

#

idk

shell inlet Dec 14, 2022, 1:15 PM

#

btw

#

do you 1/pi for direct

#

I think you don't

#

well doesn't matter for first bounce

long robin Dec 14, 2022, 1:16 PM

#

yeah direct is horribly not-PBR atm

#

btw this feels pretty overblown to me

shell inlet Dec 14, 2022, 1:17 PM

#

I'm more interested in trying a tonemapper

long robin Dec 14, 2022, 1:17 PM

#

ok, I can shoehorn one in sometime

shell inlet Dec 14, 2022, 1:17 PM

#

vec3 reinhard_ext_luminance(vec3 v, float max_white_l)
{
    float l_old = dot(v, vec3(0.2126, 0.7152, 0.0722));
    float numerator = l_old * (1.0 + (l_old / (max_white_l * max_white_l)));
    float l_new = numerator / (1.0 + l_old);
    return v * l_new / l_old;
}

long robin Dec 14, 2022, 1:17 PM

#

reinhard listenyoupieceofshit

#

ok, it's the ext version

shell inlet Dec 14, 2022, 1:18 PM

#

luminance

#

only luminance is affected

long robin Dec 14, 2022, 1:18 PM

#

nice

shell inlet Dec 14, 2022, 1:18 PM

#

btw I think you also have reinhard

#

finalColor / (1.0 + finalColor)

long robin Dec 14, 2022, 1:18 PM

#

yeah, it looks like ass though

#

it's commented out

shell inlet Dec 14, 2022, 1:19 PM

#

I don't see any difference with it

#

I am maybe blind

long robin Dec 14, 2022, 1:19 PM

#

should be in ShadeDeferredPbr.frag

#

it affects the image a lot for me

shell inlet Dec 14, 2022, 1:20 PM

#

thats what I'm editing ye

long robin Dec 14, 2022, 1:20 PM

#

here's my previous image with reinhard

shell inlet Dec 14, 2022, 1:23 PM

#

how does it compare to this

vec3 reinhard_luminance(vec3 v)
{
    float l_old = dot(v, vec3(0.2126, 0.7152, 0.0722));
    float l_new = l_old / (1.0 + l_old);
    return v * l_new / l_old;
}

long robin Dec 14, 2022, 1:26 PM

#

moar color

shell inlet Dec 14, 2022, 1:26 PM

#

epic

#

I still prefer autoexposure

#

just crushing doesn't do it for me

#Fwog and co.