Luna Engine - C++ and Vulkan | Graphics Programming | Page 2

true willow Mar 28, 2023, 5:07 PM

#

made this for fun just now hope it helps

#

and I again called the circle a sphere lol this is incurable

west hamlet Mar 28, 2023, 5:10 PM

#

its still a sphere, if you say its a 3d sphere and the plane cuts through it, giving you a slice of it

true willow Mar 28, 2023, 5:11 PM

#

fun fact most 2d functions are actually 3d functions with sulutions lying at an isosurface with z=0

#

2x+2y=0
z=2x+2y
tada it's now 3d

#

that's why marching squares and marching cubes works at all

#

even for implicit functions

west hamlet Mar 28, 2023, 5:15 PM

#

i see what you did there

true willow Mar 28, 2023, 5:30 PM

#

west hamlet i see what you did there

https://www.desmos.com/calculator/lfq9jvfphn

Desmos

extended 3d plot of the circle equation

#

here is how x^2+y^2=1 circle equation looks when extended to 3d

#

you can actually see that the original circle is the intersection with z=0 plane

west hamlet Mar 28, 2023, 5:33 PM

#

yeah

#

we did do a bit of this stuff back in school, very briefly

#

grade 11? or so

#

21 years ago

true willow Mar 28, 2023, 5:34 PM

#

you're a dinosaur

west hamlet Mar 28, 2023, 5:35 PM

#

https://tenor.com/view/toddlers-kids-dinosaur-barney-fight-gif-15758311

Tenor

cloud osprey Mar 28, 2023, 6:15 PM

#

Finding the sum of infinitely many infinitesimal slices of the function

#

I'm sure you already know sigma notation for taking the discrete sum of a function

#

An integral is just that, but for a continuous one

#

E.g., a question we have in rendering is "how much light is hitting this point from all directions". The answer is to find the integral of incoming light across the hemisphere

prisma folio Mar 30, 2023, 3:56 PM

#

Good news: The entire scene is now rendering with one vkCmdDrawIndexedIndirect

#

Bad news: RenderDoc hates me

#

also the FPS has taken a fairly steep dive, down to 47FPS

west hamlet Mar 30, 2023, 4:47 PM

#

Bad news: Its not vkCmdDispatchIndirect

fleet hollow Mar 31, 2023, 6:31 AM

#

prisma folio also the FPS has taken a fairly steep dive, down to 47FPS

So you have a GPU bottleneck? Lack of sorting by depth and or culling causing overdraw?

prisma folio Mar 31, 2023, 8:40 AM

#

fleet hollow So you have a GPU bottleneck? Lack of sorting by depth and or culling causing ov...

I mean to be fair this is 47FPS on an Intel iGPU

#

And there's a bit of overdraw, but it's only rendering Sponza so it's like a max of 3 overdraws

prisma folio Mar 31, 2023, 9:23 AM

#

IBL looks like absolute crap tho

#

this is the diffuse part of the IBL, which is the one sampling from the irradiance map

#

the irradiance map is only 64x64 but why isn't it smoothly interpolating

true willow Mar 31, 2023, 3:04 PM

#

what do normals look like because this doesn't appear to be a problem with the cubemap

#

there is no reason for a cubemap to give such oddly specific pixellations at the edges

#

I've only seen similar artifacts when your g-buffer is smaller than the swapchain and it gets interpolated when sampled, but this is not correct

prisma folio Mar 31, 2023, 3:40 PM

#

The normals on the curtains are definitely bumpy, but I wouldn't really expect that level of blockiness from it

#

Even if I disable normal mapping, there's harsh borders when the normals change sharply

true willow Mar 31, 2023, 3:50 PM

#

can you change the output of the irradiance map sampling to the vector it's sampled with (which should be a normal)?

#

so like output normals in the irradiance map pass

prisma folio Mar 31, 2023, 4:03 PM

#

Virtually identical if I do that

true willow Mar 31, 2023, 4:05 PM

#

no you can see there's an issue

prisma folio Mar 31, 2023, 4:05 PM

#

is the Y supposed to be flipped? because I thought that was a bug and "fixed" it 😅

true willow Mar 31, 2023, 4:05 PM

#

prisma folio Mar 31, 2023, 4:06 PM

#

ah

true willow Mar 31, 2023, 4:06 PM

#

look second screenshot has weird "bevel" at the intersection

#

this causes your jaggies

#

again I suspect your resolution may not match between g buffer and the framebuffer you perform the pass in

#

so the g-buffer pixels get linearly interpolated and cause the artifacts

prisma folio Mar 31, 2023, 4:08 PM

#

no, everything is definitely at 1600x900, renderdoc confirms. and the normals are subpassLoaded so there shouldn't be any interpolation

#

the first screenshot is straight from the final output

#

they both are

true willow Mar 31, 2023, 4:09 PM

#

then I'm out of ideas but you have a lead now

prisma folio Mar 31, 2023, 4:10 PM

#

unfortunately Intel has decided to declare war on RenderDoc today

#

#

according to renderdoc all of my vertex buffers are nan

#

I already had one issue with renderdoc yesterday related to buffer device addresses

#

I guess Intel just really hates them

prisma folio Mar 31, 2023, 4:52 PM

#

So apparently setting it to only sample mip 0 fixes it?

true willow Mar 31, 2023, 4:55 PM

#

this is fragment shader?

prisma folio Mar 31, 2023, 4:56 PM

#

yeah

#

nothing compute yet

true willow Mar 31, 2023, 4:56 PM

#

makes sense then

#

you often get weirdness with mips if you use fragment shader

prisma folio Mar 31, 2023, 4:56 PM

#

wat

true willow Mar 31, 2023, 4:56 PM

#

in compute it's always mip 0

#

because there's no gradients

prisma folio Mar 31, 2023, 4:57 PM

#

you mean because all the barycentric data is lost in the gbuffer?

true willow Mar 31, 2023, 4:57 PM

#

no I mean that rasterization pipeline attempts to calculate mip levels the best it can whenever you sample a texture

#

so you need to be explicit if you want to sample only top mip level

prisma folio Mar 31, 2023, 4:58 PM

#

Why would it suddenly switch mips though? The only "geometry" being rendered is a full screen triangle. What else goes into mip selection?

true willow Mar 31, 2023, 5:00 PM

#

honestly I have no clue what exactly can cause it but mip levels are selected from the gradients, I have no specific cause in mind for a fullscreen triangle

prisma folio Mar 31, 2023, 5:01 PM

#

Weird, I always assumed mips were basically chosen by primitive size and/or depth

#

which is constant in this case

true willow Mar 31, 2023, 5:01 PM

#

depth and depth gradient afaik yes

prisma folio Mar 31, 2023, 5:02 PM

#

well in any case I guess I need some AA now to deal with these fireflies

true willow Mar 31, 2023, 5:04 PM

#

I'm going to read about how mips are selected instead of spreading misinfo

#

because it still doesn't make sense that it would cause the issue if it's depth

#

it has to be uv or something

prisma folio Mar 31, 2023, 5:06 PM

#

but UV is also a smooth gradient, it's a full screen tri

#

unless you mean the sampled UV

true willow Mar 31, 2023, 5:07 PM

#

the uv used to sample the texture

#

across multiple adjacent pixels

#

hmm yeah it seems like it's uv gradients

#

not depth

prisma folio Mar 31, 2023, 5:08 PM

#

that would make more sense, since there is a sharp change in UV on those borders

true willow Mar 31, 2023, 5:09 PM

#

so it projects the sample points to the uv space and checks how much of an area it covers and picks an optimal level that best fits the size of the gradient in texture space

prisma folio Mar 31, 2023, 5:09 PM

#

and I guess that makes sense too, since having a sharp change in UV usually means you're sampling less often, and thus need a lower mip

#

I'm just confusing the mip algorithm by sampling 1 texture across the entire scene

true willow Mar 31, 2023, 5:13 PM

#

I think you should switch to the compute though, raster pipeline will inherently give a slight overhead

#

this is something you really don't need for post processing

#

you work strictly with texture data after all

prisma folio Mar 31, 2023, 5:15 PM

#

I guess I never thought of it as post processing

#

I was thinking of using light meshes next to do spot and point lights though

prisma folio Mar 31, 2023, 8:02 PM

#

visibility buffer progress

prisma folio Apr 1, 2023, 9:52 PM

#

Point lights, tonemapping, and some simple threshold/downsample/upsample bloom

muted aurora Apr 1, 2023, 9:54 PM

#

nice!

true willow Apr 1, 2023, 10:20 PM

#

prisma folio Point lights, tonemapping, and some simple threshold/downsample/upsample bloom

which tonemapping?

prisma folio Apr 1, 2023, 10:20 PM

#

uncharted 2

true willow Apr 1, 2023, 10:21 PM

#

can you try the tony mc mapface

#

it's really good

prisma folio Apr 1, 2023, 10:21 PM

#

wot

true willow Apr 1, 2023, 10:21 PM

#

tony mc mapface

#

it's good

#

can you try it?

prisma folio Apr 1, 2023, 10:28 PM

#

I don't have a way to load dds, uh

true willow Apr 1, 2023, 10:29 PM

#

convert it

prisma folio Apr 1, 2023, 10:30 PM

#

trying

west hamlet Apr 1, 2023, 10:40 PM

#

convert to ktx mayhaps

true willow Apr 1, 2023, 10:44 PM

#

west hamlet convert to ktx mayhaps

it was already done https://github.com/h3r2tic/tony-mc-mapface/issues/2

GitHub

KTX2 LUT texture · Issue #2 · h3r2tic/tony-mc-mapface

Hi! I was experimenting with https://github.com/expenses/ktx2-tools and thought I'd cook up a KTX2 version of the LUT texture as that might be more usable for some folks. I supercompressed it w...

true willow Apr 1, 2023, 10:59 PM

#

@west hamlet wanna try it too?

west hamlet Apr 1, 2023, 11:02 PM

#

not yet

#

need to get some other things working before i touch tonemapping

prisma folio Apr 1, 2023, 11:07 PM

#

[16:06:44] Luna-E: [Viewer] Failed to load LUT texture: Texture type not supported.

#

the official libktx library can't open the ktx file

true willow Apr 1, 2023, 11:08 PM

#

are you using ktxTexture2_CreateFromNamedFile()?

prisma folio Apr 1, 2023, 11:09 PM

#

CreateFromMemory but yes

true willow Apr 1, 2023, 11:09 PM

#

does it not default to the ktx1 by chance?

prisma folio Apr 1, 2023, 11:10 PM

#

Create a ktxTexture1 or ktxTexture2 from KTX-formatted data in memory according to the data contents.

#

doesn't seem to be a way to tell it 1 or 2

true willow Apr 1, 2023, 11:11 PM

#

https://github.com/KhronosGroup/KTX-Software/blob/f47d1a2e73f93c9433298c5de0a616419b567a4b/lib/texture.c#L323

#

it should automatically deduce the version it seems

#

damn this is weird

#

if only you had the ktx built from sources in debug

#

you could tell where it fails

#

not that it would help much

prisma folio Apr 1, 2023, 11:15 PM

#

literally stepping through as we speak XD

true willow Apr 1, 2023, 11:16 PM

#

oh lol

#

well there is also exr version

#

can you load exr?

#

https://github.com/h3r2tic/tony-mc-mapface/files/10809052/hdr_exr.zip

prisma folio Apr 1, 2023, 11:18 PM

#

no idea what a Bdb is but apparently it's not 0

#

the only image loading I have so far is stb_image XD

cloud osprey Apr 1, 2023, 11:24 PM

#

bob data base

#

non-const in pointers
😩

true willow Apr 1, 2023, 11:25 PM

#

data format descriptor

prisma folio Apr 1, 2023, 11:25 PM

#

to be fair This is actually modified

cloud osprey Apr 1, 2023, 11:26 PM

#

hmm does doxygen have an inout concept

#

Possible values are "[in]", "[in,out]", and "[out]", note the [square] brackets in this description. When a parameter is both input and output, [in,out] is used as attribute.

#

they done goofed

prisma folio Apr 1, 2023, 11:28 PM

#

also

@return    KTX_TRUE on success, otherwise KTX_FALSE.

return false;

#

Screen_Shot_2018-12-31_at_1.19.26_PM.jpg

prisma folio Apr 1, 2023, 11:51 PM

#

something tells me I didn't load the LUT right

true willow Apr 1, 2023, 11:52 PM

#

hey so apparently renderdoc can convert images?

#

I don't have it on this pc

#

bruh no ktx option

prisma folio Apr 2, 2023, 12:01 AM

#

Tony vs Uncharted

#

I need to add imgui back so I can play with settings live

true willow Apr 2, 2023, 12:03 AM

#

better, good path to white

#

uncharted appears to die on that purple spot on the helmet

cloud osprey Apr 2, 2023, 12:04 AM

#

muh 100% saturated magenta

prisma folio Apr 2, 2023, 12:04 AM

#

I know the basic concept of what tonemapping is but I have no idea what's good vs bad 😄

true willow Apr 2, 2023, 12:04 AM

#

#

it just got magenta'd

prisma folio Apr 2, 2023, 12:05 AM

#

tbf that purple light is set to 50 intensity right now

#

50 what? fuck if I know XD

true willow Apr 2, 2023, 12:05 AM

#

for reference 1 is max?

#

in terms of output display device range

prisma folio Apr 2, 2023, 12:06 AM

#

vec3 Lradiance = point.Radiance * point.Multiplier * attenuation;

Radiance = color
Multiplier = intensity

west hamlet Apr 2, 2023, 12:06 AM

#

one could say, 👂:slayer.gif: is a fuchsiado

prisma folio Apr 2, 2023, 12:06 AM

#

Radiance is always 0-1 for each channel

true willow Apr 2, 2023, 12:06 AM

#

can you share the LUT texture that you got working

#

if it's ktx

prisma folio Apr 2, 2023, 12:07 AM

#

it's dds

cloud osprey Apr 2, 2023, 12:07 AM

#

prisma folio Radiance is always 0-1 for each channel

that doesn't sound physically based

prisma folio Apr 2, 2023, 12:07 AM

#

I found a single-header DDS loader but I had to convert to RGBA16 to make it load

prisma folio Apr 2, 2023, 12:07 AM

#

cloud osprey that doesn't sound physically based

bold of you to assume I know anything about PBR

cloud osprey Apr 2, 2023, 12:08 AM

#

well irl radiance isn't restricted to some range, unless 1 is mapped to infinity in your thing

true willow Apr 2, 2023, 12:08 AM

#

jaker probably means that you should pass float3 unrestricted

#

unclamped

prisma folio Apr 2, 2023, 12:09 AM

#

problem is I've more or less blindly copied the shader code so I have no idea if the inputs had units or what they are

#

I mean it is

#

I'm not restricting it at all, I'm just only using 0-1 so far

cloud osprey Apr 2, 2023, 12:09 AM

#

tbh the units don't matter as long as they're consistent

#

so you better hope they're consistent bleakekw

prisma folio Apr 2, 2023, 12:09 AM

#

pl.Radiance = glm::vec3(0.36f, 0.0f, 0.63f);

true willow Apr 2, 2023, 12:09 AM

#

they matter if you use real life data

prisma folio Apr 2, 2023, 12:26 AM

#

well we're almost there

#

that's better

prisma folio Apr 2, 2023, 12:57 AM

#

and now I can do fancy things

prisma folio Apr 2, 2023, 2:09 AM

#

obligatory deferred lighting test

#

38FPS though, def needs work

#

Interesting how the scene gets brighter with no point lights enabled; is that the tonemapping?

true willow Apr 2, 2023, 10:12 AM

#

prisma folio Interesting how the scene gets brighter with no point lights enabled; is that th...

like how?

spare otter Apr 2, 2023, 10:52 AM

#

i mean it should just be black

true willow Apr 2, 2023, 10:56 AM

#

the LUT maps from classic reinhard so yes it should be black

prisma folio Apr 3, 2023, 5:07 PM

#

I do have a bit of IBL too

#

but I think it was either a weird perception issue or my monitor doing dynamic brightness

true willow Apr 3, 2023, 7:37 PM

#

what happened to rt?

#

or you want to make a rasterization based rendering too?

prisma folio Apr 3, 2023, 7:38 PM

#

honestly I'm bouncing between things so often even I don't know what my goal is

prisma folio Apr 4, 2023, 7:41 PM

#

I wonder what everyone thinks about dealing with missing data. e.g. a mesh without normals, or a material without emissive. Is it better to keep 1 shader and fill/bind placeholder data, or might it be better to build a shader that has those things removed entirely, and save what little processing time you can?

west hamlet Apr 4, 2023, 7:42 PM

#

i was finking about that too some time ago

#

and i think i ont care actually

#

i assume positions, normals, uvs, tangents, i guess if you have animations going you want to have either a separate vs for your skinning or you also mix it into one

#

then when i load meshes, i fill the meshprimitive with dummy data, if normals or tangents are not present

#

and try to calculate tangents afterwards, when uvs and normals were present but no tangents for some reason

prisma folio Apr 4, 2023, 7:45 PM

#

Yeah as of right now my mesh loader will do as the glTF spec says and generate flat normals and tangents if missing; but I'm thinking about other more esoteric attributes like UV1 or color

#

And is it worth it to build a new shader that doesn't sample a 1x1 placeholder texture?

west hamlet Apr 4, 2023, 7:45 PM

#

atm im fiddling with shadow and lighting..., i render my light volumes with a vao which only has positions for instance, and its shader is also super shrimplified to understand positions only too

#

for things like color i just use something like a base_color attribute in my material

prisma folio Apr 4, 2023, 7:46 PM

#

true but you can have a material color and a color per vertex

west hamlet Apr 4, 2023, 7:46 PM

#

if no texture is present i use that instead

#

yes, if you want to support vertex color, and its a thing for all of your stuff then have it in your main vertex format

#

otherwise making a 2nd inputlayout/vertexforma should be fine too

#

you most likely have at least another inputlayout/vertexformat going, when you do UI, imgui at least, they have vec2 positions, vec2 uvs and some uint color thing

#

maybe yet another one to debug certain things... light positions/probes/bounding boxes with positions only or position and color

true willow Apr 4, 2023, 7:49 PM

#

I always make sure all data is generated if missing

west hamlet Apr 4, 2023, 7:51 PM

#

im sure its also ok to have separate shaders for your formats, if you want/have to

#

or somehow branch inside

#

or #ifdef your attributes and access within the shader

#

like godot/filament do it

true willow Apr 4, 2023, 7:51 PM

#

it is, moreso if you have on demand generation

prisma folio Apr 4, 2023, 7:51 PM

#

I'm just wary of approaching shader permutation hell, where 1 shader can compile 65,536 ways

west hamlet Apr 4, 2023, 7:52 PM

#

you dont have to permutate each and everything

true willow Apr 4, 2023, 7:52 PM

#

that is something like what unreal has

west hamlet Apr 4, 2023, 7:52 PM

#

just the formats you need right away

prisma folio Apr 4, 2023, 7:52 PM

#

And yeah right now my shaders are all on-demand compiled

west hamlet Apr 4, 2023, 7:52 PM

#

lets say you wont have more than 10 input layouts for the vertex shader and vaoisms, and then depending on how complicated your materials are a few shaders + shadows (perhaps more than one for the different algos you want to support?) + effects + tonemap + atmosphereisms + pbr+iblisms + perhaps all the compute shaders for whateverisms

true willow Apr 4, 2023, 7:53 PM

#

realistically it won't compile to such huge numbers

prisma folio Apr 5, 2023, 5:16 PM

#

I've been working on this copy of the engine for about a month

#

I think it's time I add camera movement

prisma folio Apr 5, 2023, 7:50 PM

#

Behold: A different position!

west hamlet Apr 5, 2023, 7:53 PM

#

🟣 🟠

true willow Apr 5, 2023, 9:38 PM

#

https://tenor.com/view/spamton-dancing-gif-23332491

Tenor

prisma folio Apr 12, 2023, 9:41 PM

#

refactor time is go

prisma folio Apr 13, 2023, 10:59 PM

#

refactor has gotten this far

west hamlet Apr 13, 2023, 10:59 PM

#

looks like you render the spheres properly

#

their basecolor is that same bleu iirc

#

😄

spare otter Apr 13, 2023, 10:59 PM

#

13ms gui NanaStare

prisma folio Apr 13, 2023, 11:00 PM

#

13ms vsync

spare otter Apr 13, 2023, 11:00 PM

#

ahh

west hamlet Apr 13, 2023, 11:00 PM

#

75hz screen?

prisma folio Apr 13, 2023, 11:01 PM

#

I guess so? Intel's only offering me FIFO and Immediate present modes

#

immediate be like

#

...and a massive memory leak

#

uh

prisma folio Apr 14, 2023, 5:11 AM

#

jesus christ, I've been tearing my hair out looking at all of my object pools, memory allocations, destructors, trying to find this

#

I wasn't even progressing to the next frame context, so the deletion queue wasn't emptying

#

it's a miracle this even worked, considering that it wasn't even waiting on the timeline semaphores

#

#

gotta love windows, moving my mouse adds 0.5ms frametime

west hamlet Apr 14, 2023, 10:27 AM

#

so whats the actual frame time?

#

0.8? or 0.3?

prisma folio Apr 14, 2023, 1:42 PM

#

0.8, bumping to 1.3 when moving mouse

west hamlet Apr 14, 2023, 3:44 PM

#

on what hardware are you running this?

prisma folio Apr 14, 2023, 3:54 PM

#

that was running on GTX 3080

west hamlet Apr 14, 2023, 4:35 PM

#

oof

#

"just" 1k fps for 2 imgui windows feels super "slow"

prisma folio Apr 14, 2023, 7:12 PM

#

multiple scene views is go

#

max of 8 right now, might limit to 4 or something reasonable

prisma folio Apr 14, 2023, 8:04 PM

#

and now I've got Vulkan crying about a multithreading violation, but I can clearly see only one thread is accessing the object at a time

rich coral Apr 14, 2023, 9:55 PM

#

@prisma folio why your keeping restarting your engine?

#

I did this lot of time and trust me it just gives you burnout

prisma folio Apr 14, 2023, 9:58 PM

#

this one isn't a restart, just a huge refactor

prisma folio Apr 17, 2023, 6:08 PM

#

[Luna] =================================
[Luna] === FATAL UNHANDLED EXCEPTION ===
[Luna] =================================
[Luna] Exception Code: 0xC0000005
[Luna] Exception Occurred At: 0x00007FF60E633029 (Luna::IntrusivePtr<Luna::Vulkan::ImageView>::operator*) - (C:\Dev\Luna\Luna\Include\Luna\Utility\IntrusivePtr.hpp:178)
[Luna] - Access Violation while reading memory at 0x0000000000000028
[Luna]
[Luna] Backtrace (up to 32 frames):
[Luna] - 0:  0x00007FF60E633029 (Luna::IntrusivePtr<Luna::Vulkan::ImageView>::operator*)     (C:\Dev\Luna\Luna\Include\Luna\Utility\IntrusivePtr.hpp:178)
[Luna] - 1:  0x00007FF60E631E17 (Luna::Vulkan::Image::GetView)                               (C:\Dev\Luna\Luna\Include\Luna\Vulkan\Image.hpp:169)
[Luna] - 2:  0x00007FF60E630225 (Luna::UIManager::Texture)                                   (C:\Dev\Luna\Luna\Source\UI\UIManager.cpp:403)
[Luna] - 3:  0x00007FF60E8C38DB (Luna::ContentBrowserWindow::Update::<lambda_0>::operator()) (C:\Dev\Luna\Luna\Source\Editor\ContentBrowserWindow.cpp:80)
[Luna] - 4:  0x00007FF60E8C3482 (Luna::ContentBrowserWindow::Update)                         (C:\Dev\Luna\Luna\Source\Editor\ContentBrowserWindow.cpp:113)
[Luna] - 5:  0x00007FF60E63A240 (Luna::Editor::Update)                                       (C:\Dev\Luna\Luna\Source\Editor\Editor.cpp:65)

#

fancy custom exception handler, inspired from Worlds 😄

west hamlet Apr 17, 2023, 6:12 PM

#

its quite noisy

#

and reminds me of the time where everyone wrote custom exception handlers 😄

prisma folio Apr 17, 2023, 6:15 PM

#

it gives me what I need

#

means I don't have to switch over to a debugger just to figure out I'm stupid for using a null

west hamlet Apr 17, 2023, 6:18 PM

#

fair

#

it helps to handle null tho 🙂

#

before getting it to crash your stuff at runtime

prisma folio Apr 17, 2023, 6:19 PM

#

that's a future me problem

west hamlet Apr 17, 2023, 6:19 PM

#

heh

prisma folio Apr 17, 2023, 6:20 PM

#

on the brighter side, whee browser

west hamlet Apr 17, 2023, 6:21 PM

#

that looks neat

faint spindle Apr 17, 2023, 8:54 PM

#

i like it 🙂

spare otter Apr 17, 2023, 8:54 PM

#

icons make any engine look so much better

west hamlet Apr 17, 2023, 9:02 PM

#

even the schmaller ones the arrow thingies look neat

muted aurora Apr 17, 2023, 11:52 PM

#

prisma folio ``` [Luna] ================================= [Luna] === FATAL UNHANDLED EXCEPTIO...

woah nice!!

#

getting backtraces is a pain but that would be super handy for me lol

prisma folio Apr 18, 2023, 12:39 AM

#

it really wasn't that bad

#

basically 1 call

#

sec

#

it's really just this one line, and it gives you an array of pointers that you can resolve to symbols the same way as all the others https://github.com/Eearslya/Luna/blob/dev-reorg/Launcher/Launcher.cpp#L111

#

also random question, is virtual address space limited to 6 bytes or can it use all 8?

#

I noticed the top 2 bytes are always zeroes

cloud osprey Apr 18, 2023, 12:51 AM

#

When in doubt, just assume and you'll probably be right

prisma folio Apr 18, 2023, 12:53 AM

#

also FYI, SymInitialize is very hit-or-miss, because apparently it's only legal to call it once per application lifetime, and you can't tell if it's been called by some other program already, so it's literally impossible to guarantee you're doing it right

#

but I've found that symbols usually resolve even if SymInitialize returns false so I took that check away

cloud osprey Apr 18, 2023, 12:55 AM

#

Least scuffed win32 function

prisma folio Apr 18, 2023, 1:01 AM

#

also love how the win32 documentation says "hProcess should be your process's ID, but don't use GetCurrentProcess()" and then I went to look at their official example and

#

However, if you do use a process handle, be sure to use the correct handle. If the application is a debugger, use the process handle for the process being debugged. Do not use the handle returned by GetCurrentProcess. The handle used must be unique to avoid sharing a session with another component, and using GetCurrentProcess can have unexpected results when multiple components are attempting to use dbghelp to inspect the current process.

#

this whole page is just confusing

#

"it must be a unique value, but it doesn't have to be your process's ID. but if it is a process ID, make sure it's the right one" like wtf how is the function supposed to know the difference

west hamlet Apr 18, 2023, 10:23 AM

#

is there no std:: way to unwind exceptions?

#

🇹🇫 language is that 😄

prisma folio Apr 18, 2023, 1:30 PM

#

not until c++23

prisma folio Apr 19, 2023, 10:59 PM

#

prisma folio Apr 24, 2023, 9:55 PM

#

almost back to rendering real meshes

prisma folio Apr 25, 2023, 8:26 PM

#

tfw you forget the depth buffer

west hamlet Apr 25, 2023, 8:40 PM

#

heh

muted aurora Apr 25, 2023, 8:42 PM

#

who needs a depth buffer anyway, just draw things in the right order without resorting to hacks like that smh

true willow Apr 25, 2023, 8:51 PM

#

https://cdn.discordapp.com/attachments/362945838366064651/897821882512998420/unknown.png

prisma folio Apr 26, 2023, 8:09 PM

#

Scene serialization, woo

true willow Apr 26, 2023, 8:10 PM

#

how?

prisma folio Apr 26, 2023, 8:11 PM

#

json

true willow Apr 26, 2023, 8:11 PM

#

wait that wasn't the question but why json

#

it's text

prisma folio Apr 26, 2023, 8:11 PM

#

simplicity and debuggability

#

what was the question

true willow Apr 26, 2023, 8:12 PM

#

in code

#

did you do some majic

#

or was it painful field by field write to json

cloud osprey Apr 26, 2023, 8:12 PM

#

I would've made a new interchange format called nson (nanomachines, son)

true willow Apr 26, 2023, 8:13 PM

#

magic like reflection

#

which to my knowledge isn't in C++

cloud osprey Apr 26, 2023, 8:14 PM

#

maybe a library like cereal

prisma folio Apr 26, 2023, 8:14 PM

#

nah it's manual field-by-field

true willow Apr 26, 2023, 8:14 PM

#

I see

prisma folio Apr 26, 2023, 8:15 PM

#

gotta start somewhere

cloud osprey Apr 26, 2023, 8:15 PM

#

cereal basically makes the process of specifying which fields you want to serialize a bit shrimpler

true willow Apr 26, 2023, 8:16 PM

#

is there a library called milk that you need to use with cereal?

prisma folio Apr 26, 2023, 8:16 PM

#

true willow Apr 26, 2023, 8:16 PM

#

I think this is notepad++

prisma folio Apr 26, 2023, 8:16 PM

#

ye

true willow Apr 26, 2023, 8:17 PM

#

epic

cloud osprey Apr 26, 2023, 8:17 PM

#

how are you writing jsons

prisma folio Apr 26, 2023, 8:17 PM

#

json.hpp

#

seemed the easiest

cloud osprey Apr 26, 2023, 8:17 PM

#

i c

true willow Apr 26, 2023, 8:18 PM

#

do you translate gltfs to some other format?

#

native one

prisma folio Apr 26, 2023, 8:18 PM

#

yeah

true willow Apr 26, 2023, 8:18 PM

#

btw

#

what are you planning to do with editor vs game

#

are you going to have to write a separate project for the game?

#

i.e. how do you actually deploy the game to platforms

west hamlet Apr 26, 2023, 8:20 PM

#

there's a typo

prisma folio Apr 26, 2023, 8:21 PM

#

I haven't fully planned it out, but more or less I'm trying to architect it so you just have a single "launcher" exe that you can point at any data pack and have it run

west hamlet Apr 26, 2023, 8:21 PM

#

"Hierarchy" vs "Heirarchy"

true willow Apr 26, 2023, 8:21 PM

#

archy heir

west hamlet Apr 26, 2023, 8:21 PM

#

hairy arch

prisma folio Apr 26, 2023, 8:21 PM

#

damnit I always do that

true willow Apr 26, 2023, 8:22 PM

#

prisma folio I haven't fully planned it out, but more or less I'm trying to architect it so y...

ah like idtech1

#

quake engine that is

#

wait idtech1 is doom right

#

quake engine is quake engine

prisma folio Apr 26, 2023, 8:23 PM

#

I honestly haven't given much thought to the game part since I have little to no interest in making a game, which is...not ideal for someone writing a game engine, but I'm having fun 😄

true willow Apr 26, 2023, 8:24 PM

#

poople aren't usually happy when you can open their game in an editor and just use their assets in a different project on the same engine

prisma folio Apr 26, 2023, 8:25 PM

#

yeah I don't even know where to start with issues like that

true willow Apr 26, 2023, 8:25 PM

#

but alas it happens even on engines that obfuscates and packs assets and builds an executable at deploy time

#

people just make decompilers

#

and deobfuscators

#

same story as with DRMs

#

if it's on the user's machine you're basically only putting effort into making it harder to get immediately, but it's inevitable if there is an effort to get it

prisma folio Apr 26, 2023, 8:27 PM

#

Yeah more or less this project has been me just having fun making the systems, rather than actually wanting to make/ship something usable. It's basically a learning project for me.

#

It's more fun for me to try and implement concepts instead of just reading about them

true willow Apr 26, 2023, 8:29 PM

#

yeah that's OK, actually good you're honest with it being a toy project rather than cryengine/unity/UE killer

prisma folio Apr 26, 2023, 8:30 PM

#

oh god no

#

yeah if I ever make anything close to a game with this it'll be a miracle

#

I just love the programming and getting to play with all the buttons and switches to see how they dance

true willow Apr 26, 2023, 8:31 PM

#

make an eschatos clone

#

https://www.youtube.com/watch?v=d38Ct6Iv_ts

YouTube

Jaimers

エスカトス ~ Eschatos - Original Hard - ALL Clear 57,632,843

Eschatos came out on Steam recently and it's a really good game. I've always regretted not playing it much when it came out on the Xbox 360 way back in 2011 so you could say that this was an old score that I had to settle.

On the surface Eschatos might look a bit bland and standard but the true beauty and what makes the game so good lies in th...

▶ Play video

#

doesn't look hard to make

#

mayhaps deceptive

prisma folio Apr 26, 2023, 10:02 PM

#

#

having a grid already looks so much better than empty void

true willow Apr 26, 2023, 10:04 PM

#

nice aa

#

still moire'd though

prisma folio Apr 26, 2023, 10:05 PM

#

better if I restrict the distance a bit more

true willow Apr 26, 2023, 10:06 PM

#

nah why bother

#

keep it as it was

true willow Apr 26, 2023, 11:13 PM

#

what was it?

prisma folio Apr 26, 2023, 11:14 PM

#

I changed from a quad shader (6 vertices) to a fullscreen tri (3 vertices) and didn't change the draw count

#

https://media.tenor.com/54MsSn8iyMgAAAAC/bartlett-west.gif

#

well after nearly tearing my hair out, the grid is now a fullscreen tri

true willow Apr 26, 2023, 11:19 PM

#

you mean it's a single shader

#

?

prisma folio Apr 26, 2023, 11:21 PM

#

I mean the geometry it draws is a single triangle that covers the entire screen

true willow Apr 26, 2023, 11:21 PM

#

but is the grid drawing a shader?

prisma folio Apr 26, 2023, 11:21 PM

#

yeah it's a single vertex/fragment pair

true willow Apr 26, 2023, 11:22 PM

#

so it's using the fwidth trick?

#

to draw lines with antialiasing

prisma folio Apr 26, 2023, 11:22 PM

#

tbh I have no idea what fwidth is but yes

#

I took it all from here https://asliceofrendering.com/scene helper/2020/01/05/InfiniteGrid/

true willow Apr 26, 2023, 11:23 PM

#

basically dividing the curve by its derivative you get the signed distance iirc which you then smoothstep to have a smooth falloff

prisma folio Apr 26, 2023, 11:23 PM

#

I just changed it to use a fullscreen triangle shader instead of the special quad shader they used

#

also not passing 2 mat4s between the vertex and fragment shader concerned

true willow Apr 26, 2023, 11:24 PM

#

    vec2 derivative = fwidth(coord);
    vec2 grid = abs(fract(coord - 0.5) - 0.5) / derivative;
    float line = min(grid.x, grid.y);

#

read what I said, doesn't that sound like it

#

except this uses something else than smoothstep

prisma folio Apr 26, 2023, 11:25 PM

#

I mean it makes sense but I am still completely green on derivatives in general so I don't fully get it 😅

true willow Apr 26, 2023, 11:26 PM

#

https://www.desmos.com/calculator/fbzats4wpa

Desmos

Approximate contouring of a function via its gradient

#

visual clue

#

well if you don't get derivatives maybe even this isn't clear

#

though you need to be familiar with how any plotting works

#

to understand this

#

basically if you plug xy to the function you get a value, and your curves are functions where solutions are the points on that curve

#

so say solutions for x^2+y^2=1 is a unit circle equation and each point on that circle is a solution

#

but you can rewrite it as x^2+y^2-1 = 0 and then as f(x, y) = x^2+y^2-1, and now you have a scalar field that you can evaluate and all xy that give 0 are solutions

#

so say you want to plot it now, you evaluate f(x,y) at every pixel

#

and check if |f(x,y)| < epsilon

#

this will approximately give you pixels that are close to solutions

#

there's a problem though that the scalar field doesn't give actual distance

#

but if you divide by derivative it will, I think

#

(I haven't proven it but it seems to be somewhat true and this is what you can see in the desmos graph)

#

or maybe not at all, honestly I forgor

#

ah nope no that doesn't sound right

#

would have been too easy if it was true

prisma folio Apr 27, 2023, 10:15 PM

#

prisma folio May 11, 2023, 11:18 PM

#

I'm back, baby

west hamlet May 11, 2023, 11:18 PM

#

woohoo

prisma folio May 11, 2023, 11:37 PM

#

and just like that, more texture

west hamlet May 11, 2023, 11:43 PM

#

emissive neh?

#

my pbr is still somewhat fucked, but not that fucked, but i think im going to try to get saschaw's vulkanpbr example to work in gl

#

helmet looks so neat in his demo

muted aurora May 11, 2023, 11:45 PM

#

eyyy welcome back

prisma folio May 12, 2023, 3:54 PM

#

ah yes, perfect gizmo first try

west hamlet May 12, 2023, 3:55 PM

#

noice

#

i like how it draws over the UI 🙂

#

kinda looks neat

true willow May 12, 2023, 3:57 PM

#

writing to the front draw lists?

#

or just rendering custom gizmo on top of imgui pass

prisma folio May 12, 2023, 3:58 PM

#

imguizmo does foreground by default yeah

#

that's the easy part to fix

true willow May 12, 2023, 3:59 PM

#

is this tray racing?

#

does imguizmo decompose a matrix?

prisma folio May 12, 2023, 3:59 PM

#

no, there's no ray tracing at all here

#

you give it the view, proj, and matrix to manip

true willow May 12, 2023, 4:00 PM

#

damn that sounds unstable to me

prisma folio May 12, 2023, 4:00 PM

#

wot

true willow May 12, 2023, 4:00 PM

#

unstable

muted aurora May 12, 2023, 4:00 PM

#

it is kinda unstable, i switched away from it because my objects were slowly shrinking when i had the scale gizmo active and not doing anything (due to precision issues with conversions from trs and back)

true willow May 12, 2023, 4:04 PM

#

I actually tried it at some point but noped out and made my own gizmos to operate on TRS and never looked back, it wasn't even hard to make one

prisma folio May 12, 2023, 4:17 PM

#

muted aurora it is kinda unstable, i switched away from it because my objects were slowly shr...

what did you end up using

#

having a hard time figuring out what it wants for matrices

muted aurora May 12, 2023, 4:17 PM

#

ended up writing my own

prisma folio May 12, 2023, 4:17 PM

#

I was afraid you'd say that

#

welp

muted aurora May 12, 2023, 4:17 PM

#

yeah lol

prisma folio May 12, 2023, 4:53 PM

#

well I fixed it

#

iirc if you use the delta matrix instead of letting it manip the original matrix, the instability doesn't happen

prisma folio May 12, 2023, 11:21 PM

#

What method do you guys use for selection/outlines?

prisma folio May 15, 2023, 3:53 PM

#

first foray into inverse-z ala @regal elk's blog post, so far so good, now I just need to figure out how to fix the grid

regal elk May 15, 2023, 4:30 PM

#

tbf not my article, just in my collection of saved crap

prisma folio May 15, 2023, 5:09 PM

#

having a hard time getting the linear depth though, hmm

#

#

    float linearDepth = -(Camera.ZNear / clipSpaceDepth);

regal elk May 15, 2023, 5:11 PM

#

that looks like what it's supposed to be

#

-zNear / gl_FragCoord.z

#

it works for me in my CSM

lone moon May 15, 2023, 5:13 PM

#

I too am having a super hard time converting to infinite reverse Z lol

#

Mainly because of frustum culling plane extraction, I have no idea how to do that

prisma folio May 15, 2023, 5:15 PM

#

the article I read suggests using 0 + epsilon for the far plane, since just 0 will give you an infinite frustum and who knows what that breaks

regal elk May 15, 2023, 5:16 PM

#

what are you using it for

prisma folio May 15, 2023, 5:16 PM

#

me? the linear depth is used to fade the grid out

regal elk May 15, 2023, 5:16 PM

#

then yeah especially weird -zNear / gl_FragCoord.z isn't doing it for you

prisma folio May 15, 2023, 5:16 PM

#

otherwise you get this

#

#

well that's...not right

#

oh shoot

#

Riiiight. Okay so the problem is I'm not actually passing the camera zNear, I'm deriving it

#

which...clearly doesn't work for infinite

#

or it might, if I flip the value correctly

#

okay sanity check, this should be able to get me the zNear from just a projection matrix, right?

auto near = _invProjection * glm::vec4(0, 0, 0, 1);
near /= near.w;
_zNear = near.z;

#

or at least close enough to it

spare otter May 15, 2023, 5:31 PM

#

vec2 near_far_decompose(mat4 perspective) {
    float near = (1.0 + perspective[3][2]) / perspective[2][2];
    float far = - (1.0 - perspective[3][2]) / perspective[2][2];
    return vec2(near, far);
}

this is what I do

regal elk May 15, 2023, 5:34 PM

#

^ that's a probably better general solution but zNear is just [3][2] in the infinite proj matrix

spare otter May 15, 2023, 5:34 PM

#

yeah for infinite projection its different

regal elk May 15, 2023, 5:34 PM

#

#

and in the inverse it seems to be 1 / matrix[2][3]

#

but I actually have the zNear available as the w in my deprojection vec4 thingy

#

so my linearization usually looks like this

prisma folio May 15, 2023, 5:39 PM

#

okay well I have the right values now at least, still working out how to use them

#

debugging a random pixel gives me a linear depth of -6.2108

#

so this is the equation I'm trying to adapt to reverse-inf

float linearDepth = (2.0 * Camera.ZNear * Camera.ZFar) / (Camera.ZFar + Camera.ZNear - clipSpaceDepth * (Camera.ZFar - Camera.ZNear));
linearDepth /= Camera.ZFar;

maybe I just don't fully understand what this was doing

regal elk May 15, 2023, 5:50 PM

#

it was converting linear depth with a different projection matrix, the answer is literally -zNear / gl_FragCoord.z

prisma folio May 15, 2023, 5:51 PM

#

but the original one doesn't match any of the linearization functions on the article, which leads me to believe this isn't really linear depth

regal elk May 15, 2023, 5:51 PM

#

why not

prisma folio May 15, 2023, 5:51 PM

#

I think it's supposed to be in 0-1 range

regal elk May 15, 2023, 5:52 PM

#

ohh you want normalized range?

#

normalized linearized range

prisma folio May 15, 2023, 5:53 PM

#

I guess? It definitely seems to want 0-1 based on the way it's used

    float fade = max(0, (0.4f - linearDepth));

regal elk May 15, 2023, 5:53 PM

#

-zNear / gl_FragCoord.z is true linearized range, in the sense that it's in view space

#

it's linear in that 1 unit of that is 1 unit away from your camera

prisma folio May 15, 2023, 5:53 PM

#

hmn

regal elk May 15, 2023, 5:53 PM

#

what I'd do is just declare an arbitrary fade distance

#

and just do float fade = 1.0 - min(1.0, linearizedDepth / FADE_DISTANCE);

prisma folio May 15, 2023, 6:00 PM

#

okay yeah that's pretty good

#

    _projection       = glm::mat4(0.0f);
    _projection[0][0] = 1.0f / tanHalfFovx;
    _projection[1][1] = -(1.0f / tanHalfFovy);
    _projection[2][3] = -1.0f;
    _projection[3][2] = _zNear;

so there's my projection matrix now, it surprises me how...simple it is

spare otter May 15, 2023, 6:05 PM

#

what are you making rn

#

the grid?

prisma folio May 15, 2023, 6:05 PM

#

I was just working on moving to reverse-Z

spare otter May 15, 2023, 6:06 PM

#

ah

prisma folio May 15, 2023, 7:05 PM

#

I think I'll try mouse-picking/outlines next

lone moon May 16, 2023, 10:34 PM

#

Just passing by, you can reference this very dumb free-list allocator if you want https://github.com/LVSTRI/Iris/blob/master/src/allocator.cpp

#

(Or use an existing library that does it for you)

regal elk May 16, 2023, 10:34 PM

#

also you can save freeing for later

lone moon May 16, 2023, 10:34 PM

#

Yes

prisma folio May 16, 2023, 10:34 PM

#

MDI Pondering:

Use BDA for vertex buffers, references can be passed in uniforms with the transform data
Generate IBO each frame, possibly using a compute shader, which could be the same step as culling

regal elk May 16, 2023, 10:34 PM

#

my freeing is stubbed lol

#

because I just run scenes that load everything at startup anyway

#

are you doing meshlet culling or something?

west hamlet May 16, 2023, 10:36 PM

#

that should still be the same path

#

you provide all your meshes in your indirect buffer, but let a computeshader run over it to cull away invisible meshes and let it cook up a new indirect buffer as the result

regal elk May 16, 2023, 10:36 PM

#

yeah, though going from non-MDI to meshlet/triangle culling seems to really be jumping into the deep end

west hamlet May 16, 2023, 10:36 PM

#

yeah

lone moon May 16, 2023, 10:37 PM

#

That's a 4km jump, yes KEKW

west hamlet May 16, 2023, 10:37 PM

#

if meshlet part1 is, just cull meshprimitives, and part2 is meshlet for visbuffer isms

prisma folio May 16, 2023, 10:47 PM

#

Generating the IBO on CPU means I need to keep them in RAM rather than VRAM, unleeeess I make a compute shader that can do the copying and combine them all into one; that compute shader could also do culling and generate the indirect commands at the same time, hmm

lone moon May 16, 2023, 10:48 PM

#

You need to keep the indices in RAM? Why?

prisma folio May 16, 2023, 10:49 PM

#

well the alternative without compute shader would be a bunch of vkCmdCopyBuffers I guess

west hamlet May 16, 2023, 10:49 PM

#

dont worry too much about it

#

ram should be no problem

#

and a bunch of meshes consuming vram shouldnt be either

#

if that becomes a problem then gltfpack might come in handy, or LOD enters the chat

prisma folio May 16, 2023, 10:50 PM

#

mesh quantization is a whole other can of worms with MDI

west hamlet May 16, 2023, 10:50 PM

#

also also, all those things can be optimized/refactored later(tm)

lone moon May 16, 2023, 10:53 PM

#

By the way, LODs means more data not less nervous

#

Unless you like the 90's aesthetic and just store the highest LOD, which is a valid solution tbh

regal elk May 16, 2023, 10:57 PM

#

I still don't get why your IBOs need to persist on your CPU

#

you copy them into your main MDI buffer and discard the staging buffers

prisma folio May 16, 2023, 11:03 PM

#

because I actually want to implement the freeing

#

actually, hmm

#

I see what you mean

#

so I would just create one big IBO at the start, instead of making one that perfectly fits

regal elk May 16, 2023, 11:05 PM

#

yeah if it helps you at all, "perfectly fitting" is an NP problem (the packing problem)

#

well, helps you not rabbit hole down sometihng every allocator ever has sought for

prisma folio May 16, 2023, 11:06 PM

#

well I meant "perfectly fit" as in generating each frame, exactly how much space the indices need

lone moon May 16, 2023, 11:06 PM

#

regal elk yeah if it helps you at all, "perfectly fitting" is an NP problem (the packing p...

I can solve that in constant time and space, get on my level

regal elk May 16, 2023, 11:07 PM

#

ingredients: 1 infinite turing belt

lone moon May 16, 2023, 11:07 PM

#

But yeah generating each frame is not needed

#

You just upload and unload as needed

prisma folio May 16, 2023, 11:07 PM

#

64MB IBO would allow me to have about 5 million tris

lone moon May 16, 2023, 11:07 PM

#

Hopefully vulkan lets you specify offsets into buffers

#

Like glNamedBufferSubData

regal elk May 16, 2023, 11:08 PM

#

prisma folio May 16, 2023, 11:08 PM

#

of course it can

regal elk May 16, 2023, 11:08 PM

#

it does at vkCmdBindVertexBuffers time as well

#

but also in your indirect command

prisma folio May 16, 2023, 11:08 PM

#

I wouldn't have vertex buffers but yeah

#

not bound ones at least

lone moon May 16, 2023, 11:08 PM

#

Yeah, in GL I specify offets in the indirect command

regal elk May 16, 2023, 11:09 PM

#

yeah pretty much all of those members go into the indirect command, and m_meshIndex is the index of the indirect command itself, essentially

west hamlet May 16, 2023, 11:15 PM

#

https://github.com/deccer/Experiment/blob/main/src/Experiment.Engine/Graphics/MeshPool.cs

west hamlet May 22, 2023, 9:14 PM

#

@prisma folio are you slacking again? 🙂

west hamlet Nov 23, 2023, 3:50 PM

#

weird this thread is active all of a sudden, but last activity was in may 😛

prisma folio Dec 6, 2023, 5:45 PM

#

it lives

west hamlet Dec 6, 2023, 5:55 PM

#

welcome back : )

#

whats the plan now @prisma folio?

prisma folio Dec 6, 2023, 6:10 PM

#

Probably going to focus more on the renderer than the scene editor aspect, want to try some of these new fancy techniques

regal elk Dec 6, 2023, 6:16 PM

#

I'm curious what techniques

prisma folio Dec 6, 2023, 6:18 PM

#

I've been looking at #1128020727380054046 with no small amount of envy; so I guess things like VSM, meshlets, compute culling, there's a LOT I haven't done

west hamlet Dec 6, 2023, 6:18 PM

#

: )

#

i was also thinking about adding hzb

#

after watching simondev's latest video

#

he made it sound so simple hehe

regal elk Dec 6, 2023, 6:19 PM

#

it isn't too bad tbh, vkguide is absolutely the best place to get up to speed on it

west hamlet Dec 6, 2023, 6:19 PM

#

i also need proper culling

#

ah another thing i need to attend to, i promised vb to replay the new vkguide2

regal elk Dec 6, 2023, 6:20 PM

#

I've been meaning to upgrade mine to 2 pass culling, replace the single atomic with a prefix sum scan, and change out the MDI for an MDIC

west hamlet Dec 6, 2023, 6:20 PM

#

: )

lone moon Dec 6, 2023, 6:21 PM

#

prisma folio I've been looking at <#1128020727380054046> with no small amount of envy; so I g...

welcome to the meshlet club

#

+1 lads

west hamlet Dec 6, 2023, 6:21 PM

#

yeah, i also want to dive into that a bit

#

make more sense when you want to cull all the shizzle

regal elk Dec 6, 2023, 6:22 PM

#

same, luckily I proved that yours run on my PC lol

west hamlet Dec 6, 2023, 6:22 PM

#

: D

lone moon Dec 6, 2023, 6:22 PM

#

meshlets are for everybody

west hamlet Dec 6, 2023, 6:22 PM

#

as in you dont need a new gpu neh?

lone moon Dec 6, 2023, 6:22 PM

#

I will convert your renderer to meshlets

regal elk Dec 6, 2023, 6:22 PM

#

as in I don't have to do anything particularly crusty to port someone else's solution

west hamlet Dec 6, 2023, 6:24 PM

#

: )

#

lvstri did you write it down somewhere by any chance? a bullet point list of what you need to do exactly?

lone moon Dec 6, 2023, 6:25 PM

#

hopefully I'll get some of you frogs to do the dirty work for me and write a graph partitioner kekkedsadge

west hamlet Dec 6, 2023, 6:25 PM

#

or a blog 🙂 i thought you wanted to blog too

prisma folio Dec 6, 2023, 6:25 PM

#

Well right now I've just got a basic render graph doing imgui so my render architecture is pretty wide open; not sure where to start

lone moon Dec 6, 2023, 6:25 PM

#

west hamlet lvstri did you write it down somewhere by any chance? a bullet point list of wha...

unfortunately no, but frogfood should be a good reference

#

with mesh shaders it's even easier

west hamlet Dec 6, 2023, 6:26 PM

#

something DR cant use unfortunately

regal elk Dec 6, 2023, 6:26 PM

#

I probably would need to invest in some CPU-side meshlet culling and streaming as well, the biggest issue is idk if I have enough device memory (2GB) to actually have meaningful benefits from meshlets

lone moon Dec 6, 2023, 6:26 PM

#

download 4090 bios

west hamlet Dec 6, 2023, 6:26 PM

#

you need to prepare the gltf a little too, neh? and quantize the vbo here and there with meshopt

regal elk Dec 6, 2023, 6:27 PM

#

just run it through meshoptimizer

prisma folio Dec 6, 2023, 6:27 PM

#

I could use mesh shaders on my home PC but I do a lot of dev on an intel integrated

regal elk Dec 6, 2023, 6:27 PM

#

meshoptimizer also does the meshletification

west hamlet Dec 6, 2023, 6:27 PM

#

ah

#

then "all" you need is calculate the aabbs per meshlet and bob is my aunty so to speak

prisma folio Dec 6, 2023, 6:28 PM

#

well I guess step 1 will be to at least get a gltf loaded and rendered, then I can do the fancy

west hamlet Dec 6, 2023, 6:28 PM

#

yeah

#

same

#

i also need proper population of my indirectbuffers

prisma folio Dec 6, 2023, 6:28 PM

#

I do want to reimplement my shader manager first tho, hot reloading is ❤️

west hamlet Dec 6, 2023, 6:28 PM

#

❤️

#

https://tenor.com/view/running-home-bugs-bunny-jump-hiding-gif-8509983

Tenor

prisma folio Dec 6, 2023, 6:30 PM

#

(I say "my" shader manager, but it's almost entirely "inspired" from other projects)

west hamlet Dec 6, 2023, 6:31 PM

#

like all our projects hehe

regal elk Dec 6, 2023, 6:31 PM

#

lol speaking of which, potrick's headway in daxa has made me notice how many things I could improve in my granite-based abstraction

prisma folio Dec 6, 2023, 6:32 PM

#

oh you're going off of granite too

lone moon Dec 6, 2023, 6:32 PM

#

I started off granite and wandered off to daxa as well KEKW

regal elk Dec 6, 2023, 6:33 PM

#

I went roughly off granite and kinda went my own way in places, I think daxa is bikeshedded towards a lot of the same API design goals that I'd want though

prisma folio Dec 6, 2023, 6:34 PM

#

I've often wondered if granite's just-in-time pipeline creation would screw over GPU-driven shenanigans

regal elk Dec 6, 2023, 6:35 PM

#

it definitely wouldn't, you can't GPU-drive that hard

lone moon Dec 6, 2023, 6:35 PM

#

gpu driven shader compilation would be intetesting though

regal elk Dec 6, 2023, 6:35 PM

#

GPU driven rendering just puts your drawcalls proportional to material count instead of object/mesh/texture count

lone moon Dec 6, 2023, 6:35 PM

#

just like gpu allocation of memory

#

why can't we have nice things

prisma folio Dec 6, 2023, 6:36 PM

#

Well yeah I know you can't compile on GPU, I was more thinking about... Isn't a big part of GPU driven that you sort draw calls into bins by pipeline? If the pipeline is JIT-ed, you technically don't know what pipeline is going to be used beforehand, right?

regal elk Dec 6, 2023, 6:37 PM

#

I don't sort by true VkPipeline, materials/material passes/etc are a higher order of abstraction

lone moon Dec 6, 2023, 6:37 PM

#

you hopefully have one pipeline

regal elk Dec 6, 2023, 6:37 PM

#

especially since they don't corellate 1:1 with draw call emissions

lone moon Dec 6, 2023, 6:37 PM

#

ubershaders are great you know

regal elk Dec 6, 2023, 6:37 PM

#

yeah but you still might have multiple

west hamlet Dec 6, 2023, 6:38 PM

#

i stick to naive stuff

prisma folio Dec 6, 2023, 6:38 PM

#

as of right now my project is basically Granite's Vulkan abstraction with a few minor tweaks, plus I also took the render graph

lone moon Dec 6, 2023, 6:39 PM

#

regal elk yeah but you still might have multiple

then you visbuffer your shit up

prisma folio Dec 6, 2023, 6:39 PM

#

All of that stuff seems great, but when I started looking at their actual renderer setup it felt...weird

lone moon Dec 6, 2023, 6:39 PM

#

and do the unreal materialId = depth trick

regal elk Dec 6, 2023, 6:39 PM

#

I took the render graph but stripped out the strings and the single use handles, although those might be somewhat smart for parallelism in hindsight

prisma folio Dec 6, 2023, 6:39 PM

#

shader suites and render queues and stuff, it all felt very overengineered and rigid

#

like it will work for naive calls but not for GPU-driven

regal elk Dec 6, 2023, 6:40 PM

#

yeah granite is also somewhat old by vulkan standards

#

and designed around mobile

prisma folio Dec 6, 2023, 6:41 PM

#

Yeah I've stripped out a lot of the old cruft; like hard requiring sync2, effectively removing fences

west hamlet Dec 6, 2023, 6:41 PM

#

i still need to add CSM :3 VSM is too bigbrain for me for now, maybe next year

regal elk Dec 6, 2023, 6:41 PM

#

yeah timeline semaphores are a big driver for me to do a rewrite

#

I haven't read a write up on why I should care about sync2 barrier types though

lone moon Dec 6, 2023, 6:43 PM

#

finer sync masks

#

nothing much

prisma folio Dec 6, 2023, 6:43 PM

#

It's a bit messy right now but you can clearly see the Granite influence in the repo https://github.com/eearslya/luna/

regal elk Dec 6, 2023, 6:47 PM

#

lol yeah I definitely can

#

one big thing I haven't paid much mind to is parallelism, I deifnitely need to think more about stuff like framegraph building and JIT pipeline compilation from the perspective of trying to parallelize it

prisma folio Dec 6, 2023, 9:28 PM

#

who's got a cool gltf to play with

runic forum Dec 6, 2023, 9:31 PM

#

sample assets not enough?
https://github.com/KhronosGroup/glTF-Sample-Assets

lone moon Dec 6, 2023, 9:31 PM

#

too low poly

prisma folio Dec 6, 2023, 9:32 PM

#

I've used those to death, I want something big and fancy

#

like bistro

regal elk Dec 6, 2023, 9:37 PM

#

https://litter.catbox.moe/8lfh99.7z

#

here is darian's correctly-exported bistro glb

#

this link expires in 1 hour btw

prisma folio Dec 6, 2023, 9:40 PM

#

got it, thanks 😄

#

oh boy, it's ktx textures; wasn't expecting that yet

west hamlet Dec 6, 2023, 10:11 PM

#

libKtx is quite ez, its just 3 or 4 lines of code

#

https://github.com/JuanDiegoMontoya/Fwog/blob/a190998a24c32df63849981c803ec26d6dbbbfb3/example/common/SceneLoader.cpp#L237

#

https://github.com/JuanDiegoMontoya/Fwog/blob/a190998a24c32df63849981c803ec26d6dbbbfb3/example/common/SceneLoader.cpp#L291-L309

#

https://github.com/deccer/EngineKit/blob/91ae95dda8e79c79610888072902744c7662c76e/src/EngineKit/Graphics/GraphicsContext.cs#L332-L353

prisma folio Dec 7, 2023, 12:26 AM

#

step 1 accomplished

#

Might have to start using release mode though, the load time on this glb hurts 😄

cloud osprey Dec 7, 2023, 12:30 AM

#

You can parallelize various aspects of gltf loading

#

I did it with std::execution and now it's bearable even in debug

prisma folio Dec 7, 2023, 12:31 AM

#

I'll have to get tracy in to see exactly where the holdup is

cloud osprey Dec 7, 2023, 12:31 AM

#

The big one for me was parallelizing texture decoding
#questions message

west hamlet Dec 7, 2023, 12:31 AM

#

i wonder if one could turn a big boi mesh like that into smaller chunks (ie breaking the model into various models and en load them in paraalallel as well

prisma folio Dec 7, 2023, 12:32 AM

#

I haven't even done textures yet so this has to be either file I/O, fastgltf, or me building the vertex buffers

regal elk Dec 7, 2023, 12:32 AM

#

yeah, generally the big hitters are texture stuff, anything that allocates, and just reading the file in

west hamlet Dec 7, 2023, 12:32 AM

#

indeed

#

our plugin extension will most likely hit blender in 4.1

cloud osprey Dec 7, 2023, 12:32 AM

#

The vertex buffer loading and conversion can also be parallelized

regal elk Dec 7, 2023, 12:32 AM

#

west hamlet i wonder if one could turn a big boi mesh like that into smaller chunks (ie brea...

meshlet streaming 👀

prisma folio Dec 7, 2023, 12:33 AM

#

file I/O I kinda doubt is the issue since I'm using windows file mapping

regal elk Dec 7, 2023, 12:33 AM

#

I've profiled it before, on windows it's actually slower to file map than to just dump it into a malloc'd buffer

west hamlet Dec 7, 2023, 12:33 AM

#

regal elk meshlet streaming 👀

ah you mentioned that earlier as well, sounds like something worth to be tried : )

regal elk Dec 7, 2023, 12:33 AM

#

with the malloc time counted in the profiles

#

windows file mapping isn't meant to be used like mmap on linux, I forget why it's there but it's not the fastpath to reading large files in memory

prisma folio Dec 7, 2023, 12:34 AM

#

welp

regal elk Dec 7, 2023, 12:34 AM

#

also with this same test, mingw gcc's fopen beat OpenFile or whatever the winapi function is, 0 clue why, maybe buffered reading or some fancy syscalls or better flags than I picked for my test

#

but literally the age old

fseek(file, 0, SEEK_END);
long size = ftell(file);
rewind(file);
char* data = (char*)malloc(size);
fread(data, size, 1, file);

is the fastest possible

lone moon Dec 7, 2023, 12:36 AM

#

just map the file bro

west hamlet Dec 7, 2023, 12:36 AM

#

make sure to buy more ram before that

lone moon Dec 7, 2023, 12:36 AM

#

use the god given 64 bit address space you have

regal elk Dec 7, 2023, 12:36 AM

#

the data is probably paged in/out either way, but for some reason mapping is slower on binbows

regal elk Dec 7, 2023, 12:36 AM

#

lone moon use the god given 64 bit address space you have

see my comment above

prisma folio Dec 7, 2023, 12:36 AM

#

prisma folio file I/O I kinda doubt is the issue since I'm using windows file mapping

yes, I am XD

lone moon Dec 7, 2023, 12:36 AM

#

classic windows L

west hamlet Dec 7, 2023, 12:36 AM

#

time to switch to Lunix

lone moon Dec 7, 2023, 12:38 AM

#

rewrite windows fs, burn it down

#

or switch to linux yes KEKW

prisma folio Dec 7, 2023, 12:39 AM

#

I'll deal with ktx tomorrow, good progress today

lone moon Dec 7, 2023, 12:39 AM

#

no meshlets? :(

prisma folio Dec 7, 2023, 12:42 AM

#

I don't even know how to start with those

cloud osprey Dec 7, 2023, 12:44 AM

#

step 1: kindly ask lvstri to implement them kekwfroggified

regal elk Dec 7, 2023, 12:47 AM

#

I found my old benchmark (and cleaned it up a bit)
https://github.com/forenoonwatch/file-load-benchmark

lone moon Dec 7, 2023, 12:47 AM

#

I should write something about software meshletisms

regal elk Dec 7, 2023, 12:47 AM

#

cloud osprey Dec 7, 2023, 12:48 AM

#

You should put those numbers in the readme

#

And your setup ofc

regal elk Dec 7, 2023, 12:48 AM

#

I've rerun it a few times, but usually the result is about the same, windows and memory map are roughly equal (but they go up and down between test runs), but stdio always beats it

lone moon Dec 7, 2023, 12:48 AM

#

crazy

#

also slight tangent

#

new discord android mobile app fucking sucks

#

I send a text and the textarea doesn't clear

#

I can't even edit my msgs

#

discord™️

cloud osprey Dec 7, 2023, 12:51 AM

#

The app is already a buggy, laggy POS. How could it get worse bleakekw

lone moon Dec 7, 2023, 12:55 AM

#

anyways, one day I'll write a gist on software meshletisms, the only resource is that garbage bogus blogpost from tellusim

regal elk Dec 7, 2023, 12:56 AM

#

awesome

prisma folio Dec 7, 2023, 2:16 AM

#

Guess I'll go and look at Fwog for now then

cloud osprey Dec 7, 2023, 2:17 AM

#

frogfood is the one with meshletisms btw

#

https://github.com/JuanDiegoMontoya/Frogfood

#

fwog is just my opengl wrapper which I assume you don't care about 😄

rich coral Dec 7, 2023, 6:30 AM

#

Isn't gltf loading part of the asset pipeline process? The runtime usually loads a light file format usually and textures are pre processed as well

prisma folio Dec 7, 2023, 7:41 PM

#

meshlets are going pretty well

west hamlet Dec 7, 2023, 7:51 PM

#

: D

prisma folio Dec 7, 2023, 8:16 PM

#

minor improvement?

west hamlet Dec 7, 2023, 8:16 PM

#

now it looks like some index issue

runic forum Dec 7, 2023, 8:19 PM

#

Is this done with mesh shaders?

prisma folio Dec 7, 2023, 8:21 PM

#

nope

lone moon Dec 7, 2023, 8:44 PM

#

may I suggest starting with a cube

#

also, meshoptimizer's build meshlets func's arguments are tricky to get right, make sure they're good

west hamlet Dec 7, 2023, 8:48 PM

#

lone moon also, meshoptimizer's build meshlets func's arguments are tricky to get right, m...

do you remember posting them in iris or frogfrood by any chance?

lone moon Dec 7, 2023, 8:49 PM

#

the correct™️ ways are in frogfood

#

iris is (sadly) outdated

west hamlet Dec 7, 2023, 8:49 PM

#

froogster

lone moon Dec 7, 2023, 8:50 PM

#

https://github.com/JuanDiegoMontoya/Frogfood/blob/main/src/SceneLoader.cpp#L936-L954

west hamlet Dec 7, 2023, 8:50 PM

#

ah

#

i totally forgot meshopt is also a lib 😄

#

i was seeing cli parameters for some reason hehe

prisma folio Dec 7, 2023, 8:51 PM

#

well a triangle works 😅

lone moon Dec 7, 2023, 8:52 PM

#

the first time I did meshlets, I hit the first wall at a cube

#

i.e the cube was bogus bleakekw

#

so the path I followed was, tringle -> cube -> deccer cubes -> 2 balls -> sponza

prisma folio Dec 7, 2023, 8:59 PM

#

don't have a cube at hand but boombox is unhappy

#

oh wait gltf samples has a cube

#

I've immediately noticed a problem... I'm being given 64 indices per meshlet? Which...isn't a whole number of triangles? Is that normal?

#

I am passing 64 max indices as per meshopt's recommendation, but I didn't expect it to actually give me a partial triangle

#

Is this backwards? It says it recommends 64 indices and 124 triangles as max. How can you get to 124 triangles with 64 indices?

lone moon Dec 7, 2023, 9:10 PM

#

the same vertices can be part of multiple triangles

#

one of the main points of meshlets is high vertex reuse

prisma folio Dec 7, 2023, 9:11 PM

#

wait so I'm not supposed to use the indices as actual indices? concerned

lone moon Dec 7, 2023, 9:11 PM

#

it's a bit tricky

#

so meshoptimizer gives you two buffers in output

#

meshletIndices and meshletPrimitives

prisma folio Dec 7, 2023, 9:11 PM

#

yeah I haven't touched the primitives bit

lone moon Dec 7, 2023, 9:12 PM

#

meshletPrimitives is an array of uint8, because it assumes that a meshlet must not have a triangle count greater than 255, the values in this array serve as an index into meshletIndices

#

meshletIndices is an array of uint32 and it's basically your "index buffer", because those indices will be used to index the vertex buffer

#

so to get a vertex you must do: vertices[meshletIndices[meshlet.indexOffset + meshletPrimitives[meshlet.primitiveOffset + index]]]

#

where index = gl_VertexIndex, gl_LocalInvocationID.x, whatever else

prisma folio Dec 7, 2023, 9:14 PM

#

Ah. Shoot. Haven't done vertex pulling yet...

lone moon Dec 7, 2023, 9:14 PM

#

It's kind of a requirement with meshlets

prisma folio Dec 7, 2023, 9:17 PM

#

So wait, what would I actually pass to draw indexed then?

lone moon Dec 7, 2023, 9:19 PM

#

good question

#

ok so you can do this

#

you can generate an index buffer on the CPU

#

like this

west hamlet Dec 7, 2023, 9:20 PM

#

you can also export the starter cube from blender

#

ah it scrolled

lone moon Dec 7, 2023, 9:21 PM

#

vector<uint32> indexBuffer;
for (meshlet in meshlets) {
  for (int i = 0; i < meshlet.primitiveCount * 3; ++i) {
    indexBuffer.emplace_back(meshletIndices[meshlet.indexOffset + meshletPrimitives[meshlet.primitiveOffset + i]];
  }
}```

#

then you bind this as a regular index buffer

#

no vertex pulling required I think

#

uhhh

#

no actually you need vertex pulling for this too nevermind bleakekw

#

if you have more than one mesh from which you generate meshlets then you need vertex pulling

#

because meshlet indices and offsets are local per mesh

west hamlet Dec 7, 2023, 9:27 PM

#

: ) that is quite the little rabbithole... but im also taking notes

prisma folio Dec 7, 2023, 9:28 PM

#

vertex pulling, perfect on the first try

#

added scalar layout, back to where we were 👍

#

so I'm still a little confused as to what draw command I would execute for each meshlet

lone moon Dec 7, 2023, 9:30 PM

#

there's just a lot of gotchas

prisma folio Dec 7, 2023, 9:31 PM

#

even with vertex pulling, it's still 64 indices, no?

lone moon Dec 7, 2023, 9:31 PM

#

it can be less than 64 indices

#

remember, the 64 and 124 you set previously are upper bounds

prisma folio Dec 7, 2023, 9:31 PM

#

yeah but shouldn't it be a multiple of 3?

lone moon Dec 7, 2023, 9:31 PM

#

yesn't

prisma folio Dec 7, 2023, 9:31 PM

#

lone moon Dec 7, 2023, 9:32 PM

#

the reason is cause if you have 124 primitives you need memory to hold 124 * 3 = 372 vertices

#

NV and AMD use this to reserve 4 additional bytes of memory to write about other stuff

#

this is only relevant to mesh shaders though

#

you can realistically set your upper bounds to whatever you want really, since you're not doing mesh shaders

#

I personally recommend and use 64/64, it offers good vertex reuse and excellent culling quality

prisma folio Dec 7, 2023, 9:35 PM

#

Right, so what else needs to change without mesh shaders? Because I'm sure vkCmdDrawIndexed doesn't deal with partial-triangles very well

lone moon Dec 7, 2023, 9:36 PM

#

oh right

#

I forgor about that

#

set index count to primitive count * 3

prisma folio Dec 7, 2023, 9:39 PM

#

Aha. And what do I use for the actual bound index buffer? Would that be the normal index buffer for the mesh?

lone moon Dec 7, 2023, 9:40 PM

#

lone moon ```cpp vector<uint32> indexBuffer; for (meshlet in meshlets) { for (int i = 0;...

this will do for now

#

if you feel adventurous you can use the frogfood method (patented by yours truly)

prisma folio Dec 7, 2023, 9:40 PM

#

lone moon so to get a vertex you must do: `vertices[meshletIndices[meshlet.indexOffset + m...

you mean this one?

lone moon Dec 7, 2023, 9:41 PM

#

nono, if you use the index buffer I showed you then that becomes vertices[gl_VertexIndex]

prisma folio Dec 7, 2023, 9:41 PM

#

ah okay so the frogfood method is a lot more complex then

lone moon Dec 7, 2023, 9:41 PM

#

a bit

#

right now the frogfood method is huge

#

but if you go to the early commits, you can find a shrimplified version

prisma folio Dec 7, 2023, 9:54 PM

#

hmmm, closer, but still screwing something up..

#

lone moon Dec 7, 2023, 9:56 PM

#

make sure that the vertex buffer you gave to meshoptimizer is exactly the same as the vertex buffer you are using to pull vertices from

#

also perchance renderdoc may help here

prisma folio Dec 7, 2023, 9:57 PM

#

got a feeling it's an offset problem

#

meshlet 0

#

meshlet 1

#

ohh wait I know what it is

lone moon Dec 7, 2023, 9:58 PM

#

what are you setting baseIndex of vkCmdDrawIndexed to

prisma folio Dec 7, 2023, 9:58 PM

#

yeah that's exactly it

#

ayy

west hamlet Dec 7, 2023, 10:04 PM

#

noooooice

lone moon Dec 7, 2023, 10:05 PM

#

let the meshletization commence

prisma folio Dec 7, 2023, 10:07 PM

#

and with a little extra shader magic

#

there we go, much more distinct

#

now to try bistro again...my pc was crying last time

lone moon Dec 7, 2023, 10:10 PM

#

try sponza first

#

there is one last thing to take care of my frog

#

vertex offsets and mesh offsets

prisma folio Dec 7, 2023, 10:11 PM

#

oh no

lone moon Dec 7, 2023, 10:11 PM

#

you can render goodfroge though

prisma folio Dec 7, 2023, 10:13 PM

#

something tells me Intel does not appreciate multiple thousands of draw calls

lone moon Dec 7, 2023, 10:13 PM

#

oh

#

yeah I have one more bad news to tell you

#

and this one is really bad

#

with the frogfood method, you can only render scenes whose primitive count is less than one million

#

this is only true on intel iGPUs

prisma folio Dec 7, 2023, 10:15 PM

#

what did intel do

lone moon Dec 7, 2023, 10:15 PM

#

limit indexCount to 4 million

prisma folio Dec 7, 2023, 10:15 PM

#

honestly this is better than I expected

prisma folio Dec 7, 2023, 10:17 PM

#

lone moon limit indexCount to 4 million

Would it not be possible to just try and batch the draws?

lone moon Dec 7, 2023, 10:17 PM

#

yes

#

but you need to dispatch more workgroups and send more info with relation to offsets and what not

#

big pain

prisma folio Dec 7, 2023, 10:18 PM

#

hello sponza foliage, you're looking extra spaghetti today

#

I have a 3080 at home I could be doing this on, but work is often slow so I've just taken to doing this whenever I can

#

Though at the same time, optimizations and culling are even more important to implement on this PC 😄

west hamlet Dec 7, 2023, 10:25 PM

#

time to setup your home machine for remote work my man

#

then you can use your schlepptop as a monitor

prisma folio Dec 9, 2023, 12:15 AM

#

Haven't implemented the culling, but now fully set up on the compute-generated index buffers

#

disco boombox

west hamlet Dec 9, 2023, 12:21 AM

#

im jealous

prisma folio Dec 13, 2023, 12:02 AM

#

meshlet frustum cull

west hamlet Dec 13, 2023, 12:02 AM

#

can you also visualize the frustum?

prisma folio Dec 13, 2023, 12:58 AM

#

sorta

#

frustum visualization is a bit weird because the far plane ends up culled

#

since it's the same far plane as the matrix drawing the lines

west hamlet Dec 13, 2023, 1:13 AM

#

hehe

lone moon Dec 13, 2023, 1:28 AM

#

next up, render 4 billion triangles

prisma folio Dec 13, 2023, 2:13 AM

#

how many does bistro have

cloud osprey Dec 13, 2023, 2:15 AM

#

uh like a couple million iirc

#

it has about 1 million "primitives" (meshlet indices) according to one random vid I posted

prisma folio Dec 13, 2023, 2:17 AM

#

drawIndirect.instanceCount = 400'000;

cloud osprey Dec 13, 2023, 2:18 AM

#

About 70k meshlets in mine

prisma folio Dec 13, 2023, 2:18 AM

#

I'll be curious to see if my intel cpu can handle the meshlets once I finish culling

#

Do you guys do the "cone" culling too? idk what the real term is

#

where it determines if a meshlet is entirely back-facing

cloud osprey Dec 13, 2023, 2:20 AM

#

I don't do cone culling since I heard it barely helps

#

I guess it's basically free at runtime though

#

Idk if it makes meshlet building take longer though

prisma folio Dec 13, 2023, 2:24 AM

#

https://github.com/zeux/meshoptimizer/blob/master/src/clusterizer.cpp#L712
seems to be a sqrt and a couple dozen flops per triangle, probably not too bad?

#

could easily be paralelled too

cloud osprey Dec 13, 2023, 2:26 AM

#

Oh so they compute the normal cone no matter what

west hamlet Dec 13, 2023, 2:27 AM

#

i rewatched that video about the cones 2 days ago

prisma folio Dec 13, 2023, 2:27 AM

#

happens alongside the AABB too so

cloud osprey Dec 13, 2023, 2:28 AM

#

So the parameter I'm thinking of is related to making the meshlet based on the normal angle

prisma folio Dec 13, 2023, 2:30 AM

#

and yeah at runtime it's practically free

if (dot(normalize(cone_apex - camera_position), cone_axis) >= cone_cutoff) reject();

cloud osprey Dec 13, 2023, 2:31 AM

#

I wonder how the cone parameter in meshopt affects perf, if at all

#

Because you will get different meshlets if you have a stricter cone weight

prisma folio Dec 13, 2023, 2:35 AM

#

I don't think it affect build perf at all, looking at how it's used

cloud osprey Dec 13, 2023, 2:36 AM

#

Ye it looks like it only affects runtime perf

prisma folio Dec 13, 2023, 2:51 AM

#

Do shaders get the benefit of short-circuit? e.g. if I did bool isVisible = CullCone() && CullFrustum();, if it failed the cone test would it skip frustum entirely?

cloud osprey Dec 13, 2023, 2:58 AM

#

it's a language feature

#

I mean if one thread succeeds the test then you're executing both sides regardless

lone moon Dec 13, 2023, 4:09 PM

#

Cone culling is typically not worth it, especially in the software version of meshlets because you can very easily discard backface primitives

#

just compute the det

#

https://zeux.io/2023/04/28/triangle-backface-culling/ here's why it's not worth it

prisma folio Dec 13, 2023, 4:17 PM

#

"software version"?

lone moon Dec 13, 2023, 4:17 PM

#

the one where you use compute shaders instead of mesh shaders

prisma folio Dec 13, 2023, 6:04 PM

#

it's beautiful

west hamlet Dec 13, 2023, 6:07 PM

#

: D

prisma folio Dec 13, 2023, 6:51 PM

#

turned fully away from the mesh...4311 meshlets enter, 3762 survive. sooomething's fucky

west hamlet Dec 13, 2023, 7:07 PM

#

they snuck away when you werent looking

prisma folio Dec 13, 2023, 7:38 PM

#

that could be a problem

#

yep, my index buffer only allowed up to 200k tris, rip

#

#

vkCmdDispatch(): groupCountX (73459) exceeds device limit maxComputeWorkGroupCount[0] (65536).
oh dear

lone moon Dec 13, 2023, 8:00 PM

#

was it always that low?

#

maybe it's an intel skill issue here

prisma folio Dec 13, 2023, 8:01 PM

#

probably

cloud osprey Dec 13, 2023, 8:01 PM

#

I remember my whole PC locking up when I was doing compute experiments with vulkan and accidentally using a too-large dispatch or group size

prisma folio Dec 13, 2023, 9:03 PM

#

concerned

prisma folio Dec 13, 2023, 9:35 PM

#

[13:31:29] Luna-E: [Vulkan] Vulkan ERROR: Validation Error: [ VUID-vkCmdDrawIndexedIndirect-None-08613 ] Object 0: handle = 0x18dd50911a8, type = VK_OBJECT_TYPE_QUEUE; | MessageID = 0x1d58dc14 | vkQueueSubmit():  (set = 1, binding = 2) Descriptor index 0 access out of bounds. Descriptor size is 46827900 and highest byte accessed was 84637343 Command buffer (0x18dd550edf8). Draw Index 0x1. Pipeline (0xb3c7bc000000007f). Shader Module (0x53e60f000000006b). Shader Instruction Index = 310.  Stage = Vertex. Vertex Index = 1 Instance Index = 0.  Shader validation error occurred in file res://Shaders/StaticMesh.vert.glsl at line 60.
60:   gl_Position = Scene.ViewProjection * transform * vec4(position, 1.0);. The Vulkan spec states: If the robustBufferAccess feature is not enabled, and any VkShaderEXT bound to a stage corresponding to the pipeline bind point used by this command accesses a storage buffer, it must not access values outside of the range of the buffer as specified in the descriptor set bound to the same pipeline bind point (https://vulkan.lunarg.com/doc/view/1.3.268.0/windows/1.3-extensions/vkspec.html#VUID-vkCmdDrawIndexedIndirect-None-08613)

#

GPU assisted validation is my new best friend

lone moon Dec 13, 2023, 9:45 PM

#

Do note that OOB validation is very funky sometimes (I know this isn't OOB related, just be careful bleakekw )

prisma folio Dec 13, 2023, 9:51 PM

#

but it's exactly OOB related

lone moon Dec 13, 2023, 9:52 PM

#

oh yeah I misread KEKW

#

but ye, be careful with OOB val

prisma folio Dec 13, 2023, 10:01 PM

#

triangleOffset = 11531648

#

well that...can't be right

#

I've obviously got some weird sync error or undefined behavior because I've got flickering meshlets

#

or my two compute dispatches could be fighting over the same buffer memory, that'd do it

#

okay now this is downright bizarre... I shouldn't have to do any sync between 2 vkCmdDrawIndexedIndirect calls, right?

lone moon Dec 13, 2023, 10:30 PM

#

if there is nothing potentially catastrophic in between, no

prisma folio Dec 13, 2023, 10:31 PM

#

I am...so confused. Somehow, despite the fact that I'm giving 4 indirect draws, only the first one ends up being displayed. Renderdoc can see the others just fine.

#

actual window output

#

indirect 0

#

indirect 1

#

I even merged them into one MDI

#

how even

lone moon Dec 13, 2023, 10:34 PM

#

mayhaps it's the kompute

#

you do need sync between dispatches if you access the same memory

prisma folio Dec 13, 2023, 10:35 PM

#

the dispatches do use the same buffer, but at separate offsets, no overlap at all; does that still need sync?

lone moon Dec 13, 2023, 10:36 PM

#

I don't think so, put a megabarrier just in case

prisma folio Dec 13, 2023, 10:36 PM

#

    VisibleMeshlets.Indices[index + (BatchID * MeshletsPerBatch)] = meshletId;

#

const auto b = vk::MemoryBarrier2(vk::PipelineStageFlagBits2::eAllCommands,
                                  vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite,
                                  vk::PipelineStageFlagBits2::eAllCommands,
                                  vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite);

Like so? doesn't seem to change anything

regal elk Dec 13, 2023, 10:42 PM

#

you should change your event browser to $action() Barrier so you can see where they're at

#

makes it a bit easier to investigate sync issues

prisma folio Dec 13, 2023, 10:44 PM

#

These are the barriers I have already between compute and draw, I would think it covers everything...

regal elk Dec 13, 2023, 10:45 PM

#

do you barrier before the compute?

prisma folio Dec 13, 2023, 10:47 PM

#

Before meshlet cull (barrier to sync a vkCmdUpdateBuffer)

#

between meshlet and triangle cull

#

these also to sync against previous frame reads

regal elk Dec 13, 2023, 10:54 PM

#

that looks right

lone moon Dec 13, 2023, 10:54 PM

#

prisma folio Before meshlet cull (barrier to sync a `vkCmdUpdateBuffer`)

vkCmdUpdateBuffer is sus

prisma folio Dec 13, 2023, 10:55 PM

#

it's what I did to initialize instanceCount to 1

lone moon Dec 13, 2023, 10:55 PM

#

Hmm

#

I'll allow it

#

vkCmdFillBuffer makes me feel at ease tho

prisma folio Dec 13, 2023, 10:56 PM

#

I'd need to do both if I did that

#

or have one of the compute invocations do it

lone moon Dec 13, 2023, 10:57 PM

#

nah it's ok

#

does it still flicker if you do only one dispatch and one draw?

prisma folio Dec 13, 2023, 10:57 PM

#

It's not flickering at all now, it's just not showing the second indirect draw at all

cloud osprey Dec 13, 2023, 10:58 PM

#

prisma folio Before meshlet cull (barrier to sync a `vkCmdUpdateBuffer`)

where is this vkCmdUpdateBuffer

lone moon Dec 13, 2023, 10:58 PM

#

what does the mesh viewer show in rdoc for the second draw

regal elk Dec 13, 2023, 11:00 PM

#

if you open one of the buffers in the data viewer, sometimes you can see the values change as you progress through the timeline when there's a sync bug

prisma folio Dec 13, 2023, 11:00 PM

#

#

the values do change, but I assumed that was because each compute invocation can complete in any order

#

since it writes to buffers using an atomicAdd as an index

regal elk Dec 13, 2023, 11:02 PM

#

they shouldn't go wacky within the confines of one captured frame though

prisma folio Dec 13, 2023, 11:02 PM

#

cloud osprey where is this vkCmdUpdateBuffer

prisma folio Dec 13, 2023, 11:04 PM

#

regal elk they shouldn't go wacky within the confines of one captured frame though

if I just go back-and-forth between this and the barrier below the values do change, yeah

#

isn't rdc actually executing the dispatch anew every time you click on it though?

lone moon Dec 13, 2023, 11:05 PM

#

yes

#

non deterministic workloads unfortunately do that

prisma folio Dec 13, 2023, 11:05 PM

#

then it makes sense to me why the values would shuffle

cloud osprey Dec 13, 2023, 11:05 PM

#

it can be fine though

regal elk Dec 13, 2023, 11:06 PM

#

oh, maybe I only noticed them shuffling when I was investigating sync issues to begin with, making me think it was due to them

cloud osprey Dec 13, 2023, 11:06 PM

#

yeah if you use atomics to append data to a buffer then the order will be nondeterministic, but that's not necessarily an issue

prisma folio Dec 13, 2023, 11:06 PM

#

the vertexCount of the draw command stays the same always though, which tells me it's all the same just in a different order

#

#

the rendered output also stays stable in rdc

#

am so confused

cloud osprey Dec 13, 2023, 11:09 PM

#

what card

prisma folio Dec 13, 2023, 11:09 PM

#

intel uhd 630

cloud osprey Dec 13, 2023, 11:09 PM

#

o hek

#

can I clone your repo

prisma folio Dec 13, 2023, 11:12 PM

#

I hope so, never really tested it on other people's setups https://github.com/eearslya/luna/

#

right now it's configured to load Resources/Models/Bistro.glb (not part of the repo)

#

first compile is a bitch (thank you glslang)

#

oh and make sure to run it from the root dir; ala Build\Bin\Luna.exe

cloud osprey Dec 13, 2023, 11:28 PM

#

compilin

prisma folio Dec 13, 2023, 11:28 PM

#

prayin

cloud osprey Dec 13, 2023, 11:29 PM

#

frog_thinkk
Bitmask.hpp(155,115): error C3539: a template-argument cannot be a type that contains 'auto'

prisma folio Dec 13, 2023, 11:29 PM

#

u wot

cloud osprey Dec 13, 2023, 11:30 PM

#

[[nodiscard]] constexpr auto operator~(IsBitmaskType auto a) noexcept -> Bitmask<decltype(a)>

#

I guess it doesn't like the Bitmask<decltype(a)>

#

I'm using msvc btw

lone moon Dec 13, 2023, 11:31 PM

#

need C++20 for this to work

prisma folio Dec 13, 2023, 11:31 PM

#

cmake should be configuring c++20

cloud osprey Dec 13, 2023, 11:31 PM

#

I checked the cmake and it's using cpeepee20

lone moon Dec 13, 2023, 11:31 PM

#

msvc skill issue?

cloud osprey Dec 13, 2023, 11:34 PM

#

msvc doesn't have an issue here
https://godbolt.org/z/MfexqcYTT

prisma folio Dec 13, 2023, 11:34 PM

#

https://developercommunity.visualstudio.com/t/Get-C3539-with-abbreviated-function-temp/10119570?space=21&q=Git+LFS this was back in august, claims to have released a fix?

cloud osprey Dec 13, 2023, 11:34 PM

#

ah lemme update vs

#

yeah if I go to an old enough version in godbolt, msvc dies

prisma folio Dec 13, 2023, 11:38 PM

#

fun

#

I've been on clang for a while

cloud osprey Dec 13, 2023, 11:49 PM

#

even more errors now bleakekw

#

just linker errors, so I'm nuking the build folder and trying again

prisma folio Dec 13, 2023, 11:50 PM

#

oh boy

cloud osprey Dec 13, 2023, 11:50 PM

#

I think I saw that msvc didn't recognize some compile flags btw

#

I didn't look very hard though

#

but I'dn't be surprised if you put some clang-only flags

prisma folio Dec 13, 2023, 11:51 PM

#

I don't think I set any compile flags manually 🤔

cloud osprey Dec 13, 2023, 11:51 PM

#

o

prisma folio Dec 13, 2023, 11:51 PM

#

oh wait

#

-march=native

cloud osprey Dec 13, 2023, 11:51 PM

#

ye that's the one

#

doesn't matter though

prisma folio Dec 13, 2023, 11:52 PM

#

ye that one's just for glm intrinsics really

cloud osprey Dec 13, 2023, 11:54 PM

#

it's building but I'll have to leave in a few mins

#

damn I'm getting a billion linker errors again

prisma folio Dec 13, 2023, 11:55 PM

#

odd

cloud osprey Dec 13, 2023, 11:55 PM

#

they're all in glslc and shaderc

#

it first has a bunch of errors complaining about runtime library mismatch

#

so you may need to set a build option for shaderc

prisma folio Dec 13, 2023, 11:56 PM

#

even weirder, huh

#

I'll try and swap to msvc myself

cloud osprey Dec 13, 2023, 11:56 PM

#

also getting a bunch of errors like this

prisma folio Dec 13, 2023, 11:56 PM

#

I'm willing to bet this whole vulkan issue is an intel L tho

cloud osprey Dec 13, 2023, 11:57 PM

#

lol I could've shrunk that error window

prisma folio Dec 13, 2023, 11:57 PM

#

tfw msvc doesn't have constexpr floor and max, wat

#

or maybe log2

cloud osprey Dec 13, 2023, 11:57 PM

#

it's only constexpr in c++23

prisma folio Dec 13, 2023, 11:57 PM

#

well that's a clang oddity then

cloud osprey Dec 13, 2023, 11:58 PM

#

all of that is only constexpr in c++23

#

I ran into it myself kekkedsadge

#

anyways gtg, will be bach in 1-2 hours

prisma folio Dec 13, 2023, 11:58 PM

#

o/

cloud osprey Dec 13, 2023, 11:58 PM

#

opengl with your msvc journey

prisma folio Dec 14, 2023, 12:16 AM

#

turns out the linker errors are because vs is stupid and is trying to compile a shared AND static version of the same lib, despite me disabling it in cmake

#

set_target_properties(shaderc_shared PROPERTIES EXCLUDE_FROM_ALL ON)

guess vs doesn't care

#

works if you specifically build Luna-Launcher instead of the entire solution

#

still working out the constexpr stuff

#

and now the rest of it is hating on my bitmasks

prisma folio Dec 14, 2023, 6:21 PM

#

#

clean build \o/

#

also figured out how to make vs ignore certain projects so building the whole solution works

prisma folio Dec 14, 2023, 7:00 PM

#

also fixed the draw issue

#

apparently meshlet cull having layout(local_size_x = 128) in; is important

#

I'm not quite sure why though since I don't think the meshlet cull shader does any local sharing?

#

unless atomicAdd is a local thing too

lone moon Dec 14, 2023, 7:29 PM

#

it highly depends on how the shader was architected

#

does one gl_LocalInvocationID map to a single primitive? meshlet? maybe vertex?

#Luna Engine - C++ and Vulkan