Fwog and co.

[shell inlet]

highlights are out of the equation because those are not uniform obviously

[long robin]

I think it's about human perception/dynamic range

e.g., screens appear to become dark in direct sunlight

[shell inlet]

https://tenor.com/view/confused-what-monkey-gif-10997823

[long robin]

It's talked about in this presentation
https://gpuopen.com/wp-content/uploads/2016/03/GdcVdrLottes.pdf

[shell inlet]

I have no time to intake that much info currently

what is the takeaway?

[long robin]

First half: less ambient light = monitor effectively has more dynamic range
Second half: designing a tonemapper for HDR output (I can't tldr this)

I guess it's less of a perceptual thing, and more that all monitors reflect some light

[signal widget]

If you are curious: #1067777224528375858 message

[long robin]

here are some pertinent slides

there must be a human element too. I face a window when I'm on my computer and I notice that the dynamic range on my monitor (and other dark areas of the room) is improved if I use my hand to block outside light from reaching my eyes

[heavy cipher]

real gamers exist in total darkness, i don't understand the relevance of this work

[shell inlet]

seen this

also I've been telling all this time that aces sucks

"The ACES project began its development in 2004 in collaboration with 50 industry technologists."

50 industry technologists walk into a bar

1 tonemapper please

insert punchline here

[final cove]

if you tune your tonemapping to a calibrated monitor but no one else's monitor is calibrated... is your tonemapping really calibrated?

I like that it's got richer dark colors... but it's in the negatives? what is that scale

[signal widget]

Yes, because in practice non-calibrated monitor/TVs are not that far off anyway.
Avalanche did a test for one of their Just Cause game (if I recall) and over 100 TVs tested, most of them were not to badly calibrated. So they didn't have to worry about it.

[long robin]

one might even say true gamers are born in it, molded by it

[shell inlet]

not 100% sure but I think 10 is white point so -10 is... black point ig

even though the LUT maps from reinhard 1/(x+1) so it would make sense to use that same remapping of 0 to inf but maybe it'd be harder to demonstrate on that scale

this image shows gradients more clearly so you can see that hue shifts (Abney effect) do not happen

[golden schooner]

hehe that reminds me of an indian colleages from ages ago

he had his monitor (CRT back then) dimmed down so much it was on that level that when you go one button click further you wont see anything at all, ie black

[daring surge]

can you calibrate mine

[dapper gorge]

Jaker is suddenly going to pop up at your house, calibrate your monitor, will refuse to elaborate and will just leave

[golden schooner]

and hide in the walls again

[shell inlet]

uncalibrated monitors going extinct in jakers neighbourhood

[golden schooner]

hehe

picture him during summer holidays, knocking on all the neighbors doors, as if he wants to mow the lawn 😛

[long robin]

the srgb spec says that the standard viewing environment is an 80 nit display in a dimly lit office, which probably isn't far off from this

[golden schooner]

you couldnt see anything

plus the glas was covered in 29034809493847u9 fatty fingerprints from all the other colleagues

looked like PBR metal texture when you looked from the side

[long robin]

I'm surprised the consumer tvs would have non-shit calibration

Or maybe they had a different standard of "not too badly calibrated"

[digital lion]

@long robin would it make sense to add GL in here https://github.com/JuanDiegoMontoya/FidelityFX-FSR2/blob/master/build/GenerateSolutionsDLL.bat#L55 ? Also to the other batch build script.

[long robin]

huh, I did do that

somehow it's not pushed

[golden schooner]

https://tenor.com/view/krill-issue-skill-issue-krill-issue-gif-25572478

[digital lion]

okok

I've also noticed that this generates some linker errorhttps://github.com/BoyBaykiller/FidelityFX-FSR2/blob/master/src/ffx-fsr2-api/gl/ffx_fsr2_gl.cpp#L612.
Adding backendContext->glFunctionTable. before fixes it

[long robin]

oh crap

thanks for letting me know

can you raise an issue so I don't forgor

speaking of which, I need to close the other issues

[digital lion]

Could I also just make an PR now?

[long robin]

yeah

inshallah it is merged

[golden schooner]

https://tenor.com/view/shreyaghoshal-shukran-allah-shreya-ghoshal-gif-18068037

[golden schooner]

i saw that 😉

[long robin]

ah I see how I ignored the GenerateSolutions.bat changes this whole time

Changes (1096)

[golden schooner]

https://tenor.com/view/hahahaha-hahahahaha-haha-hahaha-fuck-gif-24734748

[digital lion]

in build/.gitignore there is GL/ so it shouldnt track those

[long robin]

Yeah that was before I pulled

[digital lion]

(they pranked us)

next one🙏

[long robin]

I guess the bug was fixed too late to be in that release

But I can confirm that I tested the fixed driver and it works

[digital lion]

oh you could already test the unreleased driver thats cool

[long robin]

when are you getting a job at AMD

[digital lion]

i dont think I'll be working in an IT-job at all

idk

[long robin]

By IT do you mean tech in general

Because there are plenty of non-IT jobs at AMD

[digital lion]

ye

[final cove]

I wonder if you can apply to be a janitor at AMD

[oak garden]

I mean

The floors dont clean themselves

[daring surge]

shrimply go fully remote

then there are no floors to clean

[long robin]

remote janitor

[daring surge]

apply to be an AMD maid

[long robin]

Remote employees get their own catboy maid to clean up after them

[golden schooner]

Intel Open Image Denoiser next?

[long robin]

Doesn't seem like it's for realtime

Also I don't see a Vulkan backend, which means it'll take actual effort to port

[daring surge]

real time dedicated denoisers are a bit

[long robin]

It only has these
OIDN_DEVICE_TYPE_DEFAULT select the likely fastest device (same as physical device with ID 0)
OIDN_DEVICE_TYPE_CPU CPU device
OIDN_DEVICE_TYPE_SYCL SYCL device (requires a supported Intel GPU)
OIDN_DEVICE_TYPE_CUDA CUDA device (requires a supported NVIDIA GPU)
OIDN_DEVICE_TYPE_HIP HIP device (requires a supported AMD GPU)

[daring surge]

they were the final nail in the coffin for me not doing occlusion with area lights

[long robin]

I don't use twitch so idk what that emote means 🤰

[daring surge]

sleepy boy

they make me tired

and exasperated

[long robin]

just svgf homie

then add the 3000 small improvements to it that have been researched over the years

[daring surge]

[long robin]

honestly svgf isn't that bad once you understand that it's fundamentally a temporal reprojection pass and a blur pass (the variance guiding isn't even useful in all contexts, so it can be ignored)

[golden schooner]

ah, oki 🙂

[heavy cipher]

or, in colloquial terms, a monad

[long robin]

added a little cmake flag and define to conditionally enable FSR 2

so our linux homies can run this example

now I can finally merge dis branch

[golden schooner]

https://tenor.com/view/arch-linux-arch-linux-gif-25380879

let me try sis

ah that thing again

cmake/vscode isnt picking up the project again

[long robin]

I didn't push yet

gimme 2 minutes

[golden schooner]

oh

[long robin]

ok changes are pushed to fsr2-test

don't forget that you need to enable FWOG_BUILD_EXAMPLES in cmake to run the exshrimples

and FSR 2 is enabled with FWOG_FSR2_ENABLE

[golden schooner]

aye

something is fucked

vscode isnt pickign up anything

let me try clion

it wont show its typical cmake status bar with all the build targets and actions

cmake's delete & reconficture isnt pulling any deps, with/without deleting 'build' first

[long robin]

hmm lemme try cloning my own repo

[golden schooner]

it only detects fwog and libglad as targets

[long robin]

did you do this

[golden schooner]

oui

[long robin]

hmm

[golden schooner]

in example/cmakeilsts.txt

option(FWOG_FSR2_ENABLE "Enable FSR2 for examples that support it (currently 03_gltf_viewer). Windows only!" TRUE)
option(FWOG_BUILD_EXAMPLES "Build Examples" TRUE)

[long robin]

hmm

[golden schooner]

wait

the 2nd option is in fwogs cmakelist already

and with fsr2 enabled i still get

CMake Error at cmake-build-debug/_deps/fsr2-src/CMakeLists.txt:69 (message):
  Cannot find MSVC toolset version 142 or greater.  Please make sure Visual
  Studio 2019 or newer installed

[long robin]

that's normal, but I'll see if I can deal with it

btw I can run my own thing from a fresh clone

[golden schooner]

ye, im just dumm

[long robin]

# Check MSVC toolset version, Visual Studio 2019 required
if(MSVC_TOOLSET_VERSION VERSION_LESS 142)
    message(FATAL_ERROR "Cannot find MSVC toolset version 142 or greater. Please make sure Visual Studio 2019 or newer installed")
endif()

somehow I doubt that this is necessary if you're just building the gl backend

[golden schooner]

same

we just use glad

[long robin]

can you try going to the cmakelists.txt in question and removing those lines

[golden schooner]

or do you still handroll glbindings in the fsrisms?

japp

[long robin]

I'm using glad.h for the function declarations and constants, but not glad.c

the user provides the function loader

[golden schooner]

line 69 😄

there is another one in
sr2-src/src/ffx-fsr2-api/CMakeLists.txt:46

ah those are also msvc flags it seems, in there

[long robin]

what compiler are you using?

[golden schooner]

gcc

[long robin]

heck

[golden schooner]

i replace it wif set(FSR2_PLATFORM_NAME x64) for now

it was able to conf the project

wat im getting permission denieds when building 😄

ah

/home/deccer/Private/Code/External/Fwog/build/_deps/fsr2-src/src/ffx-fsr2-api/../../tools/sc/FidelityFX_SC.exe: Permission denied

cmke output

perchance if one could instantiate a wine/proton session and invoke those binaries and redirect their output to some dir outside of wine/proton

ideally not wine but the proton installation which comes with steam when you enable proton support : >

[long robin]

no clue why this is

[golden schooner]

because the path doesnt exist perhaps.. let me confirm

it exists

but linux cant just invoke a random windows binary

[long robin]

wait you're on linux?

fsr2 only runs on windows

[heavy cipher]

mfw (w is windows)

[golden schooner]

so our linux homies can run this example

what kind of bait was this 🙂

[long robin]

added a little cmake flag and define to conditionally enable FSR 2

you have to disable the flag to run it on linux

[golden schooner]

instructions were unclear, patient died

[long robin]

instructions indeed unclear

[golden schooner]

but i can build everything with fsr2 disabled

[long robin]

ye that's the intent

[golden schooner]

ye

[long robin]

fsr2 is windows-only

[long robin]

it even says here 😄

[golden schooner]

unless someone gets those shader binaries to run under proton

well

that could have been some leftover

and i treated it as such

knowing the goofy history 😛

[long robin]

I could try that one fsr2 fork that adds linux support

but I'm sure it wouldn't immediately work

[golden schooner]

probably a big chungus of pain in the ass

[long robin]

exactlee

[long robin]

tbh it probably wouldn't be that hard to add if I actually had a Linux vm set up

[valid oriole]

fsr runs under vkd3d/dxvk I think?

[golden schooner]

no idea, i have not tinkered with either of those nor had the muse to invoke proton/wine from cli (theres something called wstart)

[digital lion]

After I've builded fwog how can I build the examples?
Is is not

cd example
mkdir build
cd build
cmake ..
```?

[long robin]

you need to also enable FWOG_BUILD_EXAMPLES in cmake

as a neanderthal I only know how to do it from within visual studio's json editor or cmake-gui

[digital lion]

alright, let me try

[long robin]

further enabling FSR 2 for the gltf example requires enabling FWOG_FSR2_ENABLE

[digital lion]

👌

[long robin]

re: bad fsr2 gl perf on nvidia
I looked at the spirv that is actually used in the slow pass and it is basically identical to that used in the vulkan backend (save for some extra OpImage instructions in the GL version due to my use of combined samplers)

I had to delete all the unique identifiers though because that was making my diffing tool think all the lines were different

I am also able to compare the actual™️ binaries used by the driver, but uh

chalking this one up to gl driver magic gone wrong (in the hood)

[oak garden]

lmao

[heavy cipher]

nice try amd employee

[long robin]

If only there were a tool that could let us view shader binaries with readable ISA

[digital lion]

i think fsr2 in opengl is a big deal you could make a reddit post or smth

[dapper gorge]

I also write shader assembly i hebrew

Nice ISA tbh

[long robin]

Ye just gotta document some more stuff I guess

Maybe release the blog post at the same time, idk

The post is mostly an unhinged rant, idk how much value it has to the average user lol

[golden schooner]

doesnt matter

its a nice addition to the blog of yours

[digital lion]

i actually attempted to make bindings but it keeps hard-crashing because of this dispatch https://github.com/JuanDiegoMontoya/FidelityFX-FSR2/blob/master/src/ffx-fsr2-api/ffx_fsr2.cpp#L1025 : (

[long robin]

Oh that's the first pass

Do you have more info about the crash

[digital lion]

unfortunately, not really. I just know that this optional sharpening pass also causes a crash (if I dont exit fast enough I have to restart pc) https://github.com/JuanDiegoMontoya/FidelityFX-FSR2/blob/master/src/ffx-fsr2-api/ffx_fsr2.cpp#L1041

must be something wrong with my bindings I guess. But its all simple blittable structs so idk where to look

[dapper gorge]

Btw Jaker, does Fwog support compressed textures with ktx2

[long robin]

Compressed textures are on my to-do list, but I can expedite support for u

As for ktx specifically, I think you'll still have to do some manual work as fwog doesn't deal with that kind of thing

I haven't touched ktx though so we'll see

[golden schooner]

its also on me list

[long robin]

why is maxPerStageDescriptorStorageImages 1048576 for NV gpus, but then GL_MAX_IMAGE_UNITS a pathetic 8

[golden schooner]

toomuchvoltage works for nvidia, maybe we task him to find out 😛

[long robin]

what do they do there

[golden schooner]

im not exaclty sure but possibly driver dev

[long robin]

that's the first thing everyone assumes when they hear I work at AMD

[robust bough]

@long robin I'm looking at this, what about creating a context struct? Instead of keeping a global state? (with Fwog::Initialize and Fwog::Terminate)

though maybe that's a bad idea

[long robin]

thought about it and decided it's pointless because you can't reasonably have multiple OpenGL contexts

[robust bough]

yeah

[long robin]

I'd rather have the convenience of global access, just like raw GL, than artificially impose a context that must be passed around

[golden schooner]

thats what jaker told me too back then

i had the same question 😛

[long robin]

if I were abstracting any other API, I'd happily make a context object

but only because those APIs aren't as dumb

[long robin]

time to add s3tc to my glad.h

also threw in khr_shader_subgroup because I'm silly

found this blog post: https://themaister.net/blog/2020/08/12/compressed-gpu-texture-formats-a-review-and-compute-shader-decoders-part-1/

where did the name BPTC come from anyways

it's just an alias for BC7 and BC6H

[golden schooner]

wait

you modified glad yourself?

[long robin]

no

I wasn't clear

I just generated a new one

[golden schooner]

https://github.com/deccer/CMake-Glfw-OpenGL-Template/blob/main/src/Project.Library/lib/CMakeLists.txt#L33

[long robin]

kinda annoying that GL_EXT_texture_sRGB was added 5000 years ago in gl 2, but I need it for the srgb dxt formats

[golden schooner]

didnt read the fine print in them GL_*.txt files, hey? 😄

[long robin]

eew it adds compressed luminance formats

I also find it schtrange that opengl adds the super advanced BC6H and BC7 formats in 4.2 core, but not the BC1, 3, and 5 formats

@dapper gorge do you know if it is absolutely necessary to use glCompressedTextureSubImage to update the contents of a compressed texture

I see it adds an imageSize parameter, so I guess the answer is yes

what a pain in the ass

I'm also gonna have to update my scene loader (probably by adding libktx) to actually test these and spot the 3000 incorrect assumptions of Allah

[dapper gorge]

I think so

At least if you want mips you need to copy parts of the image (of different sizes ofc) at different offsets

[long robin]

yeah after thinking about it I realized that it's probably necessary since block compressed formats are a thing

[shell inlet]

now that you have fsr2 your downsampling thing that doesn't really work right is redundant

that one when you render indirect at quarter res or something I don't recall what it was exactly

[long robin]

huh

the quarter res thing is part of the original rsm paper. I don't even use that in the advanced RSM

[long robin]

and yeah it's annoying because it doesn't even modulate albedo properly, so it gets all blurred when you lower the render resolution

[shell inlet]

?
#1019779751600205955 message

[long robin]

oh, that downsampling thing

it works perfectly fine if you don't downsample to absurdly low resolutions like that, so I assumed you were talking about the other one 😄

but yeah fsr2 mostly makes it redundant, so I already disabled it for now

I tried making a runtime slider for changing the rsm downscaling ratio, but it was horribly bugged and I couldn't be bothered to fix it

[long robin]

hmm

I have a feeling that the imageSize parameter of glCompressedTexSubImage*D can be inferred by the format and dimensions of the data

e.g., some bcn format might have 64 bits per 4x4 texels

I also find it odd that you need to specify a format when you upload compressed data (it's already specified when you create the image). surely opengl can't automatically convert between compressed formats

[long robin]

oh god astc uses trits and quints in its encoding 💀
https://themaister.net/blog/2021/08/29/compressed-gpu-texture-formats-a-review-and-compute-shader-decoders-part-3-3/

[long robin]

TIL "POV-Ray SDL" (in ktx-software repo)

[oak garden]

what is that lmao

[long robin]

some haskell ray tracing thing apparently

no clue really

clicking on the thing brings me to a "no results found" page

so it existsn't

[golden schooner]

spaghetti monster lives in people's walls too

[long robin]

http://www.povray.org/

looks old af

POV-Ray turns 30

[rugged notch]

https://github.com/POV-Ray/povray

[long robin]

I'm guessing some files were marked as "pov-ray sdl" due to some github setting

eh

[rugged notch]

#macro Trace(P, D, recLev)
  #local minT = MaxDist;
  #local closest = ObjAmnt;

  // Find closest intersection:
  #local Ind = 0;
  #while(Ind < ObjAmnt)
    #local T = calcRaySphereIntersection(P, D, Ind);
    #if(T>0 & T<minT) 
      #local minT = T;
      #local closest = Ind;
    #end
    #local Ind = Ind+1;
  #end

  // If not found, return background color:
  #if(closest = ObjAmnt)
    #local Pixel = BGColor;
  #else
    // Else calculate the color of the intersection point:
    #local IP = P+minT*D;
    #local R = Coord[closest][1].x;
    #local Normal = (IP-Coord[closest][0])/R;

    #local V = P-IP;
    #local Refl = 2*Normal*(vdot(Normal, V)) - V;

    // Lighting:
    #local Pixel = AmbientLight;
    #local Ind = 0;
    #while(Ind < LightAmnt)
      #local L = LVect[Ind][0];

      // Shadowtest:
      #local Shadowed = false;
      #local Ind2 = 0;
      #while(Ind2 < ObjAmnt)
        #if(Ind2!=closest & calcRaySphereIntersection(IP,L,Ind2)>0)
          #local Shadowed = true;
          #local Ind2 = ObjAmnt;
        #end
        #local Ind2 = Ind2+1;
      #end
      
      #if(!Shadowed)
        // Diffuse:
        #local Factor = vdot(Normal, L);
        #if(Factor > 0)
          #local Pixel=Pixel+LVect[Ind][1]*Coord[closest][2]*Factor;
        #end

        // Specular:
        #local Factor = vdot(vnormalize(Refl), L);
        #if(Factor > 0)
          #local Pixel = 
             Pixel +
             LVect[Ind][1]*pow(Factor, Coord[closest][3].x)*
             Coord[closest][3].y;
        #end
      #end
      #local Ind = Ind+1;
    #end

    // Reflection:
    #if(recLev < MaxRecLev & Coord[closest][1].y > 0)
      #local Pixel = 
    Pixel + Trace(IP, Refl, recLev+1)*Coord[closest][1].y;
    #end
  #end

  Pixel
#end

[long robin]

the "sdl" stands for "Scene Description Language"

[rugged notch]

such a nice programming language

[long robin]

at least it's readable

[golden schooner]

iirc it detects languages by file extensions

[long robin]

yeah can't you set it somehow (for the entire project)

so it doesn't think .fs files are F# if they're really fragment shaders

[golden schooner]

yeah

you might be able to coerce it via .gitattributes but meh

just name your shaders foo.2letterstage.glsl 😄

makes syntaxhighlighting in any editor shrimpler as well

[long robin]

yeah that do be what I do

as of a while ago, I got annoyed at having to change settings on everything 😄

[golden schooner]

.ps if you do hlsl is usually postscript files not pixelshaders 🙂

[long robin]

I thought this said foo.2strangeletters.glsl 😄

I read some legal thing on a gltf document and learned something

OpenGL® is a registered trademark and the OpenGL ES™ and OpenGL SC™ logos are trademarks of Hewlett Packard Enterprise used under license by Khronos.

I didn't know HP had anything to do with any of this

[fiery sorrel]

unironically saved that trivia now to my obsidian note on 'Misc Notes'

it sits next to the screenshot where u taught me that log2 can tell u how many times u can cut an image into two

which in hindsight is kinda embarassing because... yea its in the definition of what a log is i just forgor to connect it

[long robin]

damn, I can't even load gltfs that have basisu texture compression in blender

I should've foreseen that when I saw that you can't export those

[long robin]

I'm afraid tinygltf cannot do ktx 😦

wait

I think this is just enough info to load a ktx texture (bufferView + mimeType)

mashallah I thought I was gonna have to rewrite a ton of my scene loading code

[digital lion]

https://github.com/JuanDiegoMontoya/FidelityFX-FSR2/pull/5 🚀

[long robin]

how did I not see this

hmm I didn't even get an email 🤔

[digital lion]

perhaps because you are not watching the repo

[long robin]

maybe

yeah I wasn't watching it

will have to break my oath and finish implementing this ktx stuff to my loader tomorrow

[fiery sorrel]

I headcanon that jakar is unironically muslim

[long robin]

what gives you that idea habibi

[final cove]

something something fastgltf

though its not like it loads the ktx for you either

[long robin]

I did search ktx in the issues on tinygltf, only to find an issue that Sean commented on

The author of tinygltf made some basic support for ktx literally 400 commits ago

[final cove]

what I've grown to dislike about tinygltf is that for any extension it basically just gives you the JSON and says "you're on your own buddy"

[long robin]

But it doesn't matter, I am gonna use libktx to load it anyways

[final cove]

but fastgltf actually makes an effort to treat extensions like they're actually part of the spec

[long robin]

yeah tinygltf should do something for the common ones

ktx isn't an issue though. It still gets the mimetype, which is all I need

but the extensions that add more stuff are probably a huge pain

[final cove]

I remember adding the specular ext was annoying enough and that was just a few extra fields on materials

just a big chunk of ugly JSON DOM walking to do stuff the parser should

especially since the specular ext adds a few texture fields and you completely lose all of tinygltf's image handling convenience there

basically, use fastgltf

[long robin]

I might, if I encounter more severe annoyances

But for now I'm going to continue not changing everything

[rugged notch]

Praise be to Hajji Jakar Al-Hamdi

[fiery sorrel]

I live in a city where there are non arab muslims (they mostly malay)

Its interesting becos growing up thatd be my default impression and i still dont connect arab n muslim

anyways jakar based islam is the second most gp religion anyways

Muslims all worship around a hello world cube its amazing

[shell inlet]

opengllah

[golden schooner]

allahu akbar

[rugged notch]

Welcome to the opengl madrassah

[long robin]

🙏

[oak garden]

jaker has found the way of allah

[digital lion]

@long robin New FSR 2 patch version🔔

[long robin]

as if I don't have a stream of secret fsr info being pumped into my bloodstream at all times

[digital lion]

then update fork already😭

[long robin]

soon

I made an issue so I don't forget

btw, have you integrated fsr 2 in your engine yet

[digital lion]

no. It keeps crashing on those two dispatches

[long robin]

fook

do you have any leads on the cause

my only guess is that somehow certain gl function pointers aren't being loaded

[digital lion]

no and I dont really know where to look. At this point I am very much certain its not the actual interop code. But I also know the the fsr gl backend should be working (although the one missing depth format made me a bit suspicious)

[long robin]

the missing depth format was because I copied the ones used in the vulkan backend

https://github.com/GPUOpen-Effects/FidelityFX-FSR2/blob/master/src/ffx-fsr2-api/vk/ffx_fsr2_vk.cpp#L556

[long robin]

can you attach a debugger to it

[digital lion]

maybe. I dont have any experience with that kind of stuff though. Too lazy to set that up

[long robin]

aren't you using vs

you can use its native code debugger

[digital lion]

oh you mean that. ye I can

[long robin]

do you have a stacktrace of where it crashed

I'm trying to recall that one pic you sent (I can't find it in discord search)

[digital lion]

when this scheduled dispatch gets executed it crashes I think. Meaning I have to restart pc

I dont have a stacktrace or much info at all really

I need to look deeper into it

[long robin]

oh I see

a normal debugger wouldn't catch that kind of crash

must be API misuse on my part somehow, because UB in CPU code normally doesn't result in that

well, your C# bindings add another element of uncertainty

the gl backend runs fine on my AMD card

[digital lion]

Yep. I keep jumping between the fsr2 libary and my bindings thinking "This must work why I am looking here"

[long robin]

are you bindings on github somewhere

I could take a kwik look (not a C# expert, but I might spot something)

[digital lion]

thats nice I'll make a repo tomorrow and then share link here

[golden schooner]

hmm didnt want to let me take a look, noted : )

[long robin]

I kinda wish there was a unique_ptr that has a size member

I wanna manage these pointers from C APis better

oh well, I guess I have to delete the data() of a span since there is no stl template that satisfies my needs

[rugged notch]

Wait wot

unique_ptr<T[]> don't cut it?

[long robin]

that doesn't explicitly store the size

I'd have to do something terribly laborious like make a pair<unique_ptr<T[]>, size_t> to get what I want

not worth

[heavy cipher]

i believe the stl calls this type 🅱️ector

[long robin]

I'm getting a pointer from a C lib

[heavy cipher]

my condolences

[long robin]

don't wanna copy it to a 🅱️ector, as that would be inefishent

[rugged notch]

Truly a fate worse than death

[long robin]

why is there no move assignment for std::unique_ptr

nvm I'm shrimply blind

[heavy cipher]

today is "not knowing unique_ptr day"

[digital lion]

back then I still had some hope.
You can look now 🙂
https://github.com/BoyBaykiller/FidelityFX-FSR2-CSharpBindings

[long robin]

I get cancer every time I want to make a unique ptr from a raw pointer

[heavy cipher]

unique_ptr<T>(T*)?

[long robin]

Or rather, when I want to assign a raw pointer to a unique ptr

Then the compiler complains and says there is no conversion from T* to unique_ptr<T>

I cry everytim

[heavy cipher]

i'd cry too if i'd be riddled with cancer

[long robin]

oh mfw my issue is just another one of reading comprehension

the ctor I want is explicit, but I got baited by some implicit ctors that take an extra deleter arg

https://en.cppreference.com/w/cpp/memory/unique_ptr/unique_ptr

[valid oriole]

a case for RRA

[dapper gorge]

A plague that will not get better unfortunately

[viral spear]

you could also use reset

[long robin]

yeah

[dapper gorge]

@long robin I know it's 4am but do you mind reading my ramblings in #1090390868449558618 to see if they make sense

[long robin]

Uh can I save it for tomorrow

[dapper gorge]

Sure thing lol

[long robin]

I have mind frog atm, as you can clearly tell

[dapper gorge]

I'm all out of brainpower as well for now, gotta recharge

[digital lion]

btw did you find anything suspicious in my fsr2 bindings yet

[long robin]

I'm looking at em now (somehow didn't see them earlier)

what does Pack = 8 mean in these unsafe structs

[digital lion]

8 byte alignmend. I've found that what the cpp compiler does too

[long robin]

the whole struct or each member?

[digital lion]

each member

[long robin]

that doesn't seem right

only 8-byte types get that alignment in c++

I don't know what you should put to get the same alignment

[dapper gorge]

Just spam alignas 8 on each member

[long robin]

but members aren't aligned like that on most c++ compilers

[heavy cipher]

:misalignment:

[digital lion]

would that be the same as doing #pragma pack(push, 8)

[dapper gorge]

I am not sure, I think that aligns the whole struct to 8 bytes, not necessarily all its subobjects

I may very likely be wrong

[long robin]

I just checked and it doesn't align them all to 8 bytes

[digital lion]

It seems to be on every struct member so should be the same as Pack

Specifies the packing alignment for structure, union, and class members.
And I am using #pragma pack(push, 8)

[long robin]

okay

[digital lion]

Can you name which one so I can manually check the cpp/c# offsets

[long robin]

so hopefully this isn't the issue then

[long robin]

I just did this

  #pragma pack(push, 8)

  struct joe
  {
    int a, b, c, d;
  };
  constexpr int i = sizeof(joe);

sizeof(joe) is 16, which means there is no extra padding after members of this struct

[digital lion]

thats consisent with c# Pack behaviour

[dapper gorge]

I reckon if you alignas(8) before int a,b,c,d you'll get sizeof 32

[long robin]

I tried that and got the same result

I think you can't overpad

overalign

[dapper gorge]

Hmm interesting

I never understood C++'s alignment/padding rules

[long robin]

I must be tripping fr

[heavy cipher]

pack for both of these cases (C++, C#) can only lower alignment

8 would increase the alignment of those ints, so it doesn't do anything

[long robin]

isn't the whole point of alignas to overalign though

[heavy cipher]

alignas isn't pack tho

[long robin]

o

[long robin]

my issue was that I put alignas in the wrong spot, so it compiled but did nothing

I did int alignas(8) a;, but putting the alignas literally anywhere else would have caused it to work

[dapper gorge]

How does that even compile

[long robin]

idk but at least gcc warns

[dapper gorge]

mfw C++ grammar is literally: idk lmao

[long robin]

@digital lion do you have a test program using C# fsr I can run

[digital lion]

yes, project or executable?

[long robin]

project so I can debug pls

[digital lion]

https://github.com/BoyBaykiller/IDKEngine/blob/4e29d30e0f18e737395b621f60abbd876335f14e/IDKEngine/src/Render/TAAResolve.cs#L181
Thats the relevant commit and code section. But can't guarantee it ill work

[long robin]

what is nint and does ref byte match with void*?
https://github.com/BoyBaykiller/FidelityFX-FSR2-CSharpBindings/blob/master/src/FFX_FSR2/GL/GL.cs#L21

[digital lion]

nint should be size_t. And ref byte comes down to void*.
Actually nuint is size_t

[long robin]

I'll try cloning the project at that commit and running it

[golden schooner]

there is a cute racoon outside : >

https://tenor.com/view/raccoon-bread-raccoon-bread-raccoon-with-bread-racco-gif-22120699

[long robin]

it doesn't crash when I run it

do I need to do something to enable fsr2?

I see you have TAA already, but I'm guessing it's not fsr2

[digital lion]

nice. You want to uncomment this https://github.com/BoyBaykiller/IDKEngine/blob/4e29d30e0f18e737395b621f60abbd876335f14e/IDKEngine/src/Application.cs#L124

if you are on amd it likely crashs on nvidia I was just getting weird output

[long robin]

I'm on the latest commit btw

[long robin]

do I want to comment out the RunTAAResolve too or nah

[long robin]

because fsr 2 needs aliased input

[digital lion]

yes comment RunTAAResolve

[long robin]

well that can come later

anyways, I get a crash when it calls GetScratchMemory because the dll is missing

oh

the dll path is hardcoded 😄

[digital lion]

try the older commit

[long robin]

which one

you can just give me the hash too

[digital lion]

wait real quick how did you even got it working on the newest commit. I removed fsr2

[long robin]

I was actually confused

I was on "start work on fsr2" commit

[digital lion]

try 4e29d30e0f18e737395b621f60abbd876335f14e (update dlls)

should fix the hardcoded dll path

[long robin]

yep

now I see this

[digital lion]

uh

[long robin]

(on rtx 3070)

[digital lion]

hm never test on that specific card but works on rtx 3050 ti laptop (and a bunch of others)

this is without RunFSR2 right

[long robin]

oh I uncommented the fsr2 line

so it is WITH fsr2

looks fine without fsr2

I think these results are to be expected

hmm I can't step into fsr2.contextdispatch

maybe I can give it symbols

ah, but you compiled fsr2 with #pragma pack, right?

[digital lion]

the dlls in that specific commit I think dont have it

[long robin]

I need symboools

[digital lion]

the .pdb you mean?

[long robin]

I might be able to provide it with my own symbols

[digital lion]

why does that matter here btw

[long robin]

because I won't be able to just slip in my own dll with symbols already

[digital lion]

ah ok

one sec ill make a new branch containing the symbols

done

git pull
git checkout fsr2-test

[long robin]

Nice, I'll try it out in a minute

[digital lion]

oh, gitignore filtered out the files hehe. Fixed

[long robin]

Lol

ok I pulled

do I need to pull again

[digital lion]

the newest

lol

yes pull

[long robin]

hmm it's already up to date

(on fsr2-test branch)

this is as far into the call stack as I can get

ah wait

[digital lion]

oh ok. I dont know what to do. The symbols are there and build in debug

[long robin]

aha

I needed to enable native code debugging

now it's at least giving me this when I step into the pinvoke

[digital lion]

oh. I never did that stuff

[long robin]

it's trying to use a location on your computer since that's where the source code is

but I can just give it my own

[digital lion]

ok tell me when I should upload more stuff to the branch or fix something

it do be getting kinda late though. Need to see the 🛏️

[long robin]

did you modify the fsr2 source at all

[digital lion]

yes

[long robin]

hm

might mess up the source mappings for me

looks like it's just one line off, presumably from your pragma packs

[digital lion]

yes the pragma and two print statements in ffx_assert.cpp

should I update the repo dlls to cotain non edited fsr2 source files?

[long robin]

nah it's okay right now

I can step through the fsr2 source from your code

being off by 1 line is still usable

I'm mainly looking to see if the arguments are horribly corrupted

[digital lion]

yep good idea. Can you send a screenshot of the native code debugging your are doing. Just for my understanding because I never did that before. How it looks

[long robin]

I'll do u one better

[digital lion]

yo that is cool

i gotta go sleep now I can already hear the birds chirping. Thanks for taking the time

(i expect it to be fixed when I wake up)

[long robin]

gn

well so far I haven't found any issues

all the parameter structs seem fine

the gl backend struct seems fine

[digital lion]

good to have confirmation from the pros

[long robin]

oh jeez I can't renderdoc this because it requires bindless textures

rip, nsight also crashes

"unknown assertion" epic

issue was wrong working directory, but now I can't capture a frame for some reason

at least I could generate a useful trace with apitrace

[long robin]

kinda hard to debug since there are either no frame markers or apitrace doesn't show them

[daring surge]

nsight is really weird for me

multiple independent programs i've written have not worked with it

[long robin]

ye it usually doesn't give an obvious reason for why it failed (which I guess makes sense if nsight itself crashed lol)

[golden schooner]

heh

on loonix i have to tell nsight to use glfinish as the separator, it wont pick up glxswapbuffers

[long robin]

for some reason, tinygltf is not giving me correct indices for ktx textures

ah

this is why

source is normally a property of the texture itself, not the extension

krill me

time to do a little json parsing

oh, tinygltf is kind enough to store unknown extensions in an extension map

mashallah

[golden schooner]

good froge

[long robin]

oh jeez, tinygltf uses std::map unironically

[golden schooner]

why not 🙂

[long robin]

unnecessary ordering guarantee (for most cases)

[golden schooner]

when is this actually a problem?

[heavy cipher]

btw i remembered the thing

https://github.com/search?q=repo%3Amicrosoft%2FSTL _LOCK_DEBUG&type=code

std::map acquiring a global lock when iterator debugging == 2 in MSVC

[long robin]

epic

I mean, good for debugging, but dat 🅱️erf 📉

[golden schooner]

is it really this bad when you add a few things to that specific map?

in case of tinywtf

[heavy cipher]

global lock my deccy

[golden schooner]

yes

[long robin]

oh, it's global? 💀

[heavy cipher]

means you can multi🍞 tinygltf but it serializes nicely under iterator debugging

[long robin]

why is it global

I have separate maps, why must gates punish me

[heavy cipher]

unspecified transgressions of the opengl kind

[golden schooner]

thats why everyone uses unordered_map instead?

eh nevermind, i feel like getting slapped in the next 5s for asking 🙂

[dapper gorge]

Perhaps

Truth is all maps are garbage in C++

[long robin]

btw this is the chain I need to go through to access the texture source index in tinygltf now

[dapper gorge]

std::map is just especially bad

[long robin]

      int textureSource = texture.source;
      if (auto it = texture.extensions.find("KHR_texture_basisu"); it != texture.extensions.end())
      {
        struct Hack : tinygltf::Value
        {
          Hack(const tinygltf::Value& basisuExtension)
            : Value(basisuExtension) {}

          int GetTextureSource() const
          {
            return this->object_value_.at("source").GetNumberAsInt();
          }
        };
        textureSource = Hack {it->second}.GetTextureSource();
        //printf("source: %d\n", textureSource);
      }
      const auto& image = images[textureSource];

I want 2 die

I have to use a struct because the value I want is a protected member of tinygltf::Value

[golden schooner]

; (

[long robin]

I just wanna load compressed textures

[heavy cipher]

not really

its just something unexpected

[long robin]

perhaps the real reason is that the standard mandates that it do be kwite slow (regardless of this)

[golden schooner]

i wouldnt know, and expect map to be a dictionary like structure for when i need a dictionary and assume its well implemented, given its part of a standarlib

[long robin]

it's an ordered map

idk if dicts have ordering normally

[golden schooner]

perhaps sorteddicts 🙂

not assorted dicks

[dapper gorge]

💀

[final cove]

not to be too harsh on the tinygltf guy but a lot of its code strikes me like C++ isn't his first/main language

[long robin]

he writes so many C++ libs though

[final cove]

it sorta looks like he's generally used to something else but kinda knows about the fundamentals and STL containers to make a lib, but not how to really leverage C++

[long robin]

yeah seems like it

[final cove]

ouch!

I sentence him to 100 hours of cppcon talks

[long robin]

yeah I'll probably use a different lib next time I need to load gltfs

[dapper gorge]

I have been using cgltf and can't complain

[heavy cipher]

they are japanese iirc

[long robin]

a billion failed asserts later

[dapper gorge]

s t r e t c h

[golden schooner]

that looks krispy

[long robin]

the textures are surprisingly correct, so I guess uvs are fooked (damn mesh quantization)

[oak garden]

[dapper gorge]

I think you need -noq -vpf

[long robin]

I can't even remember what this does, but I had to uncomment the GL_BYTE case to make it work

auto AddElementsNorm = [&]<typename Vec>()
{
  for (size_t i = 0; i < accessor.count; i++)
  {
    // this doesn't actually work exactly for signed types
    attributeBuffer[i] = Fvec(*reinterpret_cast<const Vec*>(buffer.data.data() + totalByteOffset + i * stride)) / Fvec(static_cast<float>(std::numeric_limits<typename Vec::value_type>::max()));
  }
};

switch (accessor.componentType)
{
//case GL_BYTE:           AddElementsNorm.template operator()<Bvec>(); break;
case GL_UNSIGNED_BYTE:  AddElementsNorm.template operator()<UBvec>(); break;
//case GL_SHORT:          AddElementsNorm.template operator()<Svec>(); break;
case GL_UNSIGNED_SHORT: AddElementsNorm.template operator()<USvec>(); break;
case GL_UNSIGNED_INT:   AddElementsNorm.template operator()<UIvec>(); break;
//case GL_FLOAT:          AddElementsNorm.template operator()<Fvec>(); break;
default: FWOG_UNREACHABLE;
}

otherwise it hit the unreachable path

[dapper gorge]

wtf

[long robin]

ikr

[oak garden]

[long robin]

it's my code for converting gltf vertices to my runtime vertex format

now I remember that snorm requires a more complex integer->float calculation, so I didn't do it

[golden schooner]

i like the roof color contribution at the top of the scene

[long robin]

good eye

[golden schooner]

https://tenor.com/view/golden-eye-pierce-brosnan-james-bond-sean-bean-alec-trevelyan-gif-4997021

(idk if you know james bond: golden eye)

[long robin]

muchas gracias

made the glb a third larger though 💀

[dapper gorge]

when the
when the quantization

[long robin]

I made a function for loading snorm attributes, but it still didn't work, so I tried those flags

try not to cringe challenge (impossible)

auto AddElementsSnorm = [&]<typename Vec>()
{
  for (size_t i = 0; i < accessor.count; i++)
  {
    // max(c / (2^(b-1) - 1), -1.0)
    // The second -1 is implicit due to truncation after division
    attributeBuffer[i] = glm::max(Fvec(*reinterpret_cast<const Vec*>(buffer.data.data() + totalByteOffset + i * stride)) /
                         Fvec(static_cast<float>(std::numeric_limits<typename Vec::value_type>::max() / 2)), Fvec(-1));
  }
};

oh wait that's wrong

[golden schooner]

that looks neat

[long robin]

this is the "correct", "readable" version of it, but the mesh still renders incorrectly without lvstri's flags, so meh

oh wow this function isn't even my issue

uvs are unsigned shorts, no I've no clue why they're turning into doodoo

blender can load the quantized mesh fine

[digital lion]

thats because of this https://github.com/BoyBaykiller/IDKEngine/blob/fsr2-test/IDKEngine/src/Windowing/GameWindowBase.cs#L100
I once put it there for making sure the screen is black on start up

[long robin]

multiplying all unorm attributes by 128 somehow makes it look almost correct

[dapper gorge]

Perhaps you need meshopt_decodeVertexBuffer?

[long robin]

I'm not even using meshopt

is that my problem

[dapper gorge]

I'm not sure, I was looking at docs for gltfpack

https://meshoptimizer.org/gltf/

[long robin]

https://stackoverflow.com/questions/75946773/how-to-use-khr-mesh-quantization-khr-texture-transform-to-import-uv-with-gltf

[dapper gorge]

Ah wait no, that's not right

Yeah it's just KHR_mesh_quantization

https://github.com/KhronosGroup/glTF/blob/main/extensions/2.0/Khronos/KHR_mesh_quantization/README.md

[long robin]

For TEXCOORD attribute, the transform can be specified using the offset/scale supplied by KHR_texture_transform extension. Note that this requires merging existing offset/scale values, if present, with the dequantization transform.

oh boy, struct Hack2 : tinygltf::Value let's go

[dapper gorge]

Hack was so good they made Hack2

[long robin]

btw I ctrl+f'd "hack" and found this line in my loader
texcoords.resize(positions.size()); // TEMP HACK

quite amazing

I think that's for meshes with no texcoords

[dapper gorge]

Hack 3? :pog:

[golden schooner]

no worries, i do the same

[long robin]

why is the texture transform a property of the texture

hmm, I guess it helps keep your mesh quantized in a small range

me after 6 hours of compressed textures and gltf self-flagellation
https://youtu.be/2eEcAvzR0FI

[golden schooner]

heh

time to rest your heavy ass

sleep well https://www.youtube.com/watch?v=7hewSKpkgSM

[long robin]

a lullaby to put me to sleep, nice

[final cove]

you could've just used fastgltf

[fiery sorrel]

its not ramadan why would he need to fastgltf

[golden schooner]

that was a good one 🙂

anyway, belated كل عام و أنتم بخير

[shell inlet]

https://www.youtube.com/watch?v=oUuErZ6VwpA&ab_channel=Ali

jaker lore

[long robin]

I don't get how KHR_texture_transform can be a "used extension" but not a required one

what is the line that separates required extensions from "unnecessary" ones?

like, KHR_texture_basisu is required because you can't properly load textures if you ignore it

but ignoring KHR_texture_transform will screw up your UVs horribly if you ignore it

[shell inlet]

it might be an oopsie from the exporter

[long robin]

perhap

the spec only says this

All glTF extensions required to load and/or render an asset MUST be listed in the top-level extensionsRequired array

it's a pretty weak specification

they even use KHR_materials_sheen in the example of a non-required extension, and I'm pretty sure you'd need to support that to render the asset properly lmao

[daring surge]

uhhhh

that seems smart to me

required vs non required is like... should an engine reject this asset entirely if it doesn't support this extension

yeah?

[long robin]

and how should that be determined?

[daring surge]

subjectively

[long robin]

that is evidently how they decided this

[daring surge]

i think texture transform can be optional in some cases

like if you're using it for detail textures, you don't necessarily need support for it

and because you need to favor optional usecases over failing use cases, it fits into that category

[long robin]

pretty much everything can be optional in some circumstances with that reasoning

I guess gltfpack doesn't use any logic to determine whether texture transform is "actually needed"

[daring surge]

personally i'd just assert out on any that i don't expect

[long robin]

yeah I'm doing that now heh

I'll just ignore requiredExtensions

[long robin]

all four newsponza models now fit in under 4gb of vram

[long robin]

tree

[daring surge]

this server needs a frogchamp emote

[shell inlet]

I thought the RSM would make it more colorful

can barely see any bleeding

[long robin]

yeah newsponza curtains have a big gap in the middle which means more multicolored light from the wall can hit the floor

[shell inlet]

I just realized that it actually makes sense

most of the DI covers bland gray walls not colorful curtains

so indirect is largely gray too

[long robin]

yeah, this is the best I can focus the light on the curtain

[shell inlet]

k I see the blue thing now

[daring surge]

i feel like the discrete curtains colors were a big part of sponza, you could see the colored bounces so clearly

[shell inlet]

and orange too below

have you tried any fun scenes besides sponza and miguel san?

though maybe you need more bounces to truly bring out the color bleeding

[long robin]

don't look too closely

multi-curtain drifting

[long robin]

when I attempt to load textures as BC3

lol just realized that I named my compressed formats incorrectly

I just did DXTn -> BCn when converting from my API enums to opengl

turns out DXT3 = BC2

then opengl randomly calls BC4 and BC5 "RGTC"

https://registry.khronos.org/OpenGL/extensions/ARB/ARB_texture_compression_rgtc.txt

my bad, it's not BC4 and BC5. It just happens to share an identical encoding with this extension

[long robin]

ye no more corruption now

[heavy cipher]

i love compressed formats, you fiddle with them for hours but then its so obvious when you look at the result

[long robin]

khronos should put the entire format encoding spec in the compressed format enumerators so we don't forget

the parkour between different naming conventions hurts

anywhomst, fwog now supports the full spectrum of BC formats

though I still don't know what is supposed to happen when you copy or download compressed images

[heavy cipher]

instant defenestration

[dapper gorge]

BC1 = S3TC_DXT1
BC2 = S3TC_DXT3
BC3 = S3TC_DXT5
BC4 = RGTC1

OpenGL™️

[long robin]

also notice how all the dxt ones come from an extension

well actually, I have to use yet another extension to get the sRGB versions of the DXT formats

[golden schooner]

that looks neat

[long robin]

btw @digital lion I updated my fsr2 fork with the latest upstream changes (2.2.1) and also changed the interface slightly (removed unused parameters in the ffx_gl header)

so I'm afraid you'll have to update your bindings if you want the latest fsr 😳

actually that's not even true, you can just checkout this commit to avoid changing your bindings
https://github.com/JuanDiegoMontoya/FidelityFX-FSR2/commit/25cf567f7c365a923b1f71095c322e834e5f903b

[digital lion]

wunderbar

[long robin]

secured myself one of them newfangled usernames

[snow sun]

oh dang

gpgpu is strong

[long robin]

there are a surprising amount of four letter usernames left, like fvec, bsdf, and vec5 (random ones I tried)

[dapper gorge]

Your account is now worth 5 billion dollars

@long robin did you figure out quantization in the end

[long robin]

KHR_mesh_quantization and KHR_texture_transform?

[dapper gorge]

Just the first one

[long robin]

somehow I didn't have a problem with that ext specifically (I think), even though I want using the quantization transforms

though, I'm thinking about upgrading to fastgltf now that I looked at how clean the code is in #1019965526434394173

and it supports all these extensions (and more)

[dapper gorge]

I'm pretty content with cgltf, I don't think I understand how strides work

[long robin]

what are you conchfused about

[dapper gorge]

I have quantized positions in uh

one sec

Ah yes, quantized for 14 bits, I forgor

2^14 is 16384 and it shows as maximum in the accessor, all is good

[long robin]

hmm

[dapper gorge]

Except the accessor tells me that the component type is vec3

so is the stride sizeof(uint16) * 3?

[long robin]

Note that to comply with alignment rules for accessors, "VEC3" accessors need to be aligned to 4-byte boundaries; for example, a BYTE normal is expected to have a stride of 4, not 3.

btw fastgltf would automatically handle this (I think)

||let the shilling begin||

[dapper gorge]

https://tenor.com/view/me-looking-for-who-asked-looking-for-who-asked-who-asked-me-looking-gif-20318322

jk

I may or may not use fastgltf for the vulkan thing

I don't feel like rewriting model loading in GL

[long robin]

I wonder how to make gltfpack export a gltf with mesh quantization but not texture transform

I don't want to have to change my shaders

[dapper gorge]

-v -tc -kn -km -kam -ke -ac -vpi -vp 14 -vt 12 -vn 8 -vc 8 -tj 24 -tq 10```

ya welcome

Ah one of these flags is an impostor

[long robin]

which one turns off texture transform optimization

[dapper gorge]

It's a custom flag I coded myself into gltfpack, it avoids nodes merging

[dapper gorge]

Great question

[long robin]

why would you not want this

[dapper gorge]

Occlusion culling dies (frustum too)

[long robin]

rip in pepino

[dapper gorge]

Alright if everything works well

I should've compressed vertices from 64 bytes to 20

Now if this doesn't make SM throughput the TOP sol I'll go to NVIDIA's HQ for, uh, talking yes, just talking

https://tenor.com/view/doom-gif-25908805

Just for talking

[golden schooner]

its ackchually quite impressive what you do already my spaghetti dude

[long robin]

I don't see how smaller vertices would increase occupancy

[dapper gorge]

I'm limited by VRAM (also by register allocation)

[long robin]

Ah I hallucinated what you said

[golden schooner]

concinve jaker to have him ship one of them workstation 48gb cards to you

[long robin]

Yeah it should help with vram, probably not worth registers though

[dapper gorge]

Hmm yes, do that please Jaker

[long robin]

My team doesn't use those cards

We only use consumer cards

[golden schooner]

https://tenor.com/view/just-make-it-happen-curtis-payne-tyler-perrys-house-of-payne-just-do-it-do-what-you-must-gif-19167311

hehe

and/or you lustri you should write about it

and share it

gloat that your shit is better than GL_NV_mesh_shader

the peeps in the industry might see it

[dapper gorge]

Mesh Shaders have a huge advantage though

[golden schooner]

its a smol world

[dapper gorge]

Which is not mesh shaders

It's task shaders, they can basically dispatch work onto the GPU

So the GPU basically dispatches work to itself, which is huge

[long robin]

mdic is the same thing 🤰

[dapper gorge]

You can do culling and unpacking there, dispatch mesh shaders and render only what's visibile

[dapper gorge]

Kind of yes, but with task shaders you can dispatch arbitrary work

[digital lion]

20 bytes👀

show vertex struct please

[dapper gorge]

struct meshlet_vertex_format_t {
    alignas(4) glm::u16vec3 position = {};
    alignas(4) glm::u8vec3 normal = {};
    alignas(4) glm::u16vec2 uv = {};
    alignas(4) glm::u8vec4 tangent = {};
};```

[long robin]

struct { char data[20]; };

[dapper gorge]

It's smol, doesn't have a lot of attributes

[long robin]

normal could be u8vec2 if you use oct encoding

and I think you can get away with u16 (or even nothing) for tangent if you go full tryhard

actually just ignore me and read this
https://www.jeremyong.com/graphics/2023/01/09/tangent-spaces-and-diamond-encoding/

[dapper gorge]

I will be going full tryhard

This vertex struct shall occupy only 1 bit

Until then I won't stop

[long robin]

you could technically get it below 1 bit in some non-practical cases

like if you apply lz4 on your vertex data and it happens to repeat a lot, but then you have to decompress in the vertex shader which is 💀

[dapper gorge]

I like to imagine at one point the vertex shader becomes somewhere you do literally anything except transforming vertices to clip space

[final cove]

it would probably run pretty poorly but that's definitely worth being able to brag you have 0.5 bits per vertex

[long robin]

There is gdeflate, which you could use to decompress data on the GPU

[heavy cipher]

--bababooey

[heavy cipher]

1 bit vertex buffer => if bit is 1, then lookup gl_VertexIndex'th vertex of Sponza from a hardcoded array

i have yet to see any other use of engines than rendering the sponz, so this seems like a very applicable compression scheme

[dapper gorge]

Jaker, could you peek at my issue in me thread

[long robin]

I'm tempted to try that, but I'm guessing the shader just won't compile

I looked, but

[dapper gorge]

Hmm, I'll try with the trusty default blender cube™️

[dapper gorge]

@long robin is Fwog texture compression up on GitHub?

[long robin]

Yes

[long robin]

read it
https://juandiegomontoya.github.io/porting_fsr2.html

[dapper gorge]

this is not an option

[long robin]

btw there is a todo on that page. I am accepting memes to put there

there also should be no editing mistakes in there because I used ms word to check

[oak garden]

Todo: insert hilarious meme here

Good read

[stray warren]

GP slang leaking

[hexed bay]

F

[stray warren]

I actually nearly spat out my coffee

[long robin]

lol

I nearly removed the "ass", but now I'm glad I kept it

[rugged notch]

just put a link to this and don't explain it https://www.youtube.com/watch?v=cHqfk5lMw-E

[long robin]

man wtf is this

[golden schooner]

i watched it, and then i read on, and i have the same question

"i would come in your dream"

[long robin]

is that why I wake up "sweaty" every night

[golden schooner]

😄

[stray warren]

Really liked the article :) Also, TIL about GL_EXT_shader_explicit_arithmetic_types

[digital lion]

it seems like when feeding opengl with spirv you get access to all these typically vulkan-glsl extensions like this one?
Because as far as I know nvidia doesnt support it when using raw-ass glsl

[long robin]

yeah it's kinda sus

I don't think they're "intentionally" supported

but there is also no way to query supported shader extensions, so meh

I guess they added that one feature to query spirv extensions, but there are barely exts to query

definitely fewer than what exists in reality

[stray warren]

OpenGL GLSL extensions have been the death of me as a NV extension silently works on my AMD laptop but with small artifacts. Honestly didn't know that GL ingested spirv so it's worth looking into

Side note I like how hyper specific the froge meme is

[long robin]

I stole it from #frogshed-𓆏

[digital lion]

yeah the spirv situtation is weird

[digital lion]

current amd drivers are especially cursed. Your problem could be related to this:
https://community.amd.com/t5/opengl-vulkan/opengl-bug-preprocessor-names-of-extensions-exposed-which-arent/m-p/561518#M4415.
Also can I know which particular NV ext you mean?

[long robin]

I also experienced some bugs on RDNA 2 (or maybe it was Pascal) after uploading all the shaders in my voxel engine as spir-v, so I just undid that change

the fsr2 shaders seem to work on both vendors as spir-v, albeit nvidia pessimizing two of the passes severely

[stray warren]

On phone rn but it literally is this exact situation, just with a different NV extension. Will follow that issue as well thanks 🙏

[digital lion]

wait jaker do you have read access to the internal bug ticks from the opengl driver team

[long robin]

yes, and write access even

updoot
https://www.reddit.com/r/opengl/comments/148dcwh/opengl_port_of_fsr_2/

[heavy cipher]

reddit username reveal 😳

[long robin]

Nah, that's just some guy who copied me

[heavy cipher]

are samplers and textures different objects in gl tho?

[long robin]

Yes

Textures have built-in sampler state, but you can make separate sampler objects that you can bind to override that

[heavy cipher]

til

nice job

[fiery sorrel]

https://github.com/JuanDiegoMontoya/Fwog/pull/84

yes ik its silly and trivial asf but well just my way of showing I still care about the fwog 😅

[long robin]

mmm yeah I need to put an assert for that in more places, thx for the reminder

[fiery sorrel]

😅

I remembered that idea of making a Fwog version of CMake-Template

then I realized after forking that it shouldn't be a fork persay

[long robin]

what is this

[fiery sorrel]

I wanted to adapt https://github.com/deccer/CMake-Glfw-OpenGL-Template but with Fwog example and inside by default

because why not lol plus it'd be a good way to start new Fwog projects for ppl with a template

since the examples aren't really a template (I did use them as such naively my first time)

originally made it as a Fork then I realized 'wait, but its not intended to ever be merged into the original thats not what a Fork is supposed to be'

[long robin]

I didn't even know about the virtual assistant

[fiery sorrel]

yea same.

[long robin]

why can't you just detach your fork with a button

what a weird thing

[fiery sorrel]

https://stackoverflow.com/questions/38831301/how-to-un-fork-the-github-repository/66470086

I only found it from here.

[long robin]

interesting

[golden schooner]

@fiery sorrel describe in your own words wth you are actually trying to do

[fiery sorrel]

A cmake template based off yours but with a Fwog scene inside already

So ppl can use it to code something with fwog straight away in new projects

I can choose not to make it public if you prefer it not existing

So Deccer Cmake Template + Fwog

[golden schooner]

that i got, i just dont understand the unfork nonsense

[fiery sorrel]

https://github.com/ClementineAccount/Fwog-CMake-Glfw-OpenGL-Template

It is Deccer's CMake Template but Fwog version.

[long robin]

nice

btw the link to the pic in the readme is broken

[fiery sorrel]

why 😢

[long robin]

it dieded

[fiery sorrel]

ok i try and fix

[long robin]

epic

[fiery sorrel]

some stuff kinda messy in da code but its fine so far. I'll improve it later

I see cube. It has example of using samplers and texture. honestly just needs a skybox example and itd be a good 'starter scene' for anyone ig

[robust bough]

I think Deccer cubes would be good

[fiery sorrel]

prob should go with a less harsh shade of green for the screenshot

[long robin]

[0, 1, 0] is a nice, pleasing color

[golden schooner]

0.1, 0.2, 0.3

[fiery sorrel]

maybe as an extra scene. I actually used Deccer's Cmake Template to do my Deccer's Cube and I had to like block my eyes while I deleted the Deccer solution as it was a spoiler

so why did I make this? Just felt like it because as said no actual template for Fwog but turns out it was a decent way to refactor some Plane Game code indirectly

there we go

(yea I flipped the 0.2 and 0.3 because i figured fwog should be green)

oh yea and this template is public so if @long robin or @robust bough (who hates the globalist) wanna refactor the example to have better practices I'd accept the PRs

i can use the feedback too.

I tried to, unlike the example with the triangle in Fwog examples, use these scary things called... classes...

oh they ended up being structs nvm

[golden schooner]

only thing left is have fwog use fetchcontent for glad

[fiery sorrel]

So urm

Templates dont have an Issues board

oh no they do

i just need to enable them somehow?

there we go

[golden schooner]

theres one more flaw in the current fwog template

the ui is with srgb on

[fiery sorrel]

oh

how do I turn that off

[golden schooner]

check fwog's exshrimples

[fiery sorrel]

ok

Found it

[golden schooner]

hmm 2 spaces indent looks weird

is that fwog or your template?

[fiery sorrel]

https://github.com/ClementineAccount/Fwog-CMake-Glfw-OpenGL-Template/issues/3

[fiery sorrel]

@golden schooner that is from: https://github.com/JuanDiegoMontoya/Fwog/blob/main/example/common/Application.cpp#L241

[golden schooner]

you can post links without embed with < > arond the link

[golden schooner]

need to have a word with jaker about that 😛

[fiery sorrel]

ironically did that for the second one 😅

but forgot for the first haha

[long robin]

2-space indents are based 😤😤😤😤

[fiery sorrel]

I keep forgetting if I am setting up my 'local indent/spacing' style stuff properly

so that when I can push and stuff it can be to some standard which dont look nice on my screen but is standardized

then on my IDE or editor I can do whatever I want

[long robin]

actually idc about the width as long as it's 2 or 4 spaces, I just don't like tabs because it makes consistently formatting stuff across devices harder

[fiery sorrel]

isn't there a way for my visual studios and visual code to auto convert my tabs to two spaces

that way i can use my pinky but still be in the space lifestyle

[long robin]

yeah it should show at the bottom of the text editor which mode it's in

or use a .clang-format file

[fiery sorrel]

oh sweet didn't know that

that means ive been living a space lifestyle without realizing it

This must be what those kids whose parents fed them brownies that had spanish baked in them felt like

And now in a weird cycle of life, I can go back and put Fwog Template stuff inside Alberquerque's codebase as a project, even tho its based off refactoring Alberquerque's Plane Game code 😅

[golden schooner]

yep, tab is a nono, tabs as spaces should be a setting set in any editor of yours, be it the editor itself or something like .editorconfig/.clang-format etc

vs defaults to tabs as spaces aka "smart tabs"

[long robin]

I use 2-space indents because that's what my first programming professor forced us to use (our assignments would be checked for formatting and would insta-fail if we didn't adhere to conventions lol)

[golden schooner]

[long robin]

we were given a tool to check formatting so it was 100% our fault if it wasn't up to par

tons of people STILL managed to submit code without checking formatting and failed assignments as a result

[golden schooner]

ye 😦

but there is also .clang-format AND .editorconfig and resharper has its projectName.DotSettings (but also supports .editorconfig)

[long robin]

we were also given a driver program to use to test our assignment and people wouldn't run it and just submit a wrong assignment instead

[golden schooner]

oh man

[long robin]

people are lazy/dumb

the prof was strict, but 100% fair

[golden schooner]

his way of weeding out the dum dums

[long robin]

a bunch of people got caught cheating on the early assignments too

which were trivial things like fibonacci or sieve of eratosthenes

[fiery sorrel]

oh we had to do sieve of eratosthenes in x86_64 assembly for an assigment once

oh wait no it was finding square roots of an input nvm 🙃

[long robin]

just use fsqrt

[fiery sorrel]

heh didnt know that was a thing. ofc the point was to just use only mov, add and stuff. had to keep track of a stack for the parameter

[long robin]

the trick I used was googling x86 sqrt

[fiery sorrel]

Any preferences for the skybox? Else I'd just make it numbered and draw my own crudely.

[long robin]

"numbered"?

[fiery sorrel]

Labelled*

Front, Back, Top, Down in the file

and in the image

but if u have an actual skybox u wanna use and have license for i can put it as example

[long robin]

ah

I don't mind the green, personally

[fiery sorrel]

ok. I'll keep the 'green' but have an example 'skybox' thats just labelled 'front' and stuff on a transparent png

and then its basically 'done'

I just wanted to have a skybox example as Fwog doesn't have any and it'd be cool to have one as making skyboxes in fwog is easy

[golden schooner]

😄

or this one

[long robin]

Let me help

[fiery sorrel]

Peffecto. I can include both and allow user to switch in the dear imgui.

Gonna let Jaker decide tho since its still representing his framework at end of the day

[golden schooner]

i have miramar ready as well

or the red sky without the letters

[long robin]

Here you go

[golden schooner]

😄

[long robin]

You're the one making something, therefore you choose

Make it organic instead of something that I have my grubby fingers deeply embedded within

[fiery sorrel]

oki! I'd prob use this one

u got the one without the letters too

then i can make it so the letters are a toggle

[golden schooner]

perhaps its not green-ey

also have those

rye field sunset