Iris - A Journey through OpenGL and beyond to learn Graphics

[primal shadow]

For reference here's the roadmap:

• Removing material depth writing from the raster pass
• Implicit tangents
• Moving to persistent culling over a BVH-per-object of cluster groups instead of the huge flat dispatch over clusters
• Software raster and making hardware raster less memory-hungry without mesh shaders (blocked on wgpu texture atomics)
• Better mesh conversion workflows and asset processing, and making it faster and higher quality
• Compute-based shading and software VRS (long term)
• Streaming (long term)

[left jacinth]

@wicked notch I copied the compute rasterization function you had, and modified it to render quads instead of triangles, and of course I know your code is directly ported from UE5 source, which I can see correlates very highly. I was wondering if you had thought at all about the licensing of UE5's source

the specific file itself just says "all rights reserved" at the top

but also they say this

you can't copy paste UE code without being subject to royalties, but you can "learn" from UE code...

At what point does it count as copying

especially if the code in question is a software triangle rasterizer, which was understood and modified to instead be a software quad rasterizer

[left jacinth]

https://www.unrealengine.com/en-US/faq

[wicked notch]

yeah, I was definitely thinking of removing that and replacing it, I guess I'll expedite that

[left jacinth]

what would you replace it with?

I don't understand copyright law man...

[wicked notch]

back in the day I was looking at various softrast methods

I dunno what to replace the subpixel calcs with tho

I gotta look at something public somehow, iirc AMD has some subpixel AA going on in their FFX repos

[faint crane]

Does "it was revealed to me in a dream" not work with Unreal's lawyers?

[wicked notch]

I mean my project is purely for personal use

so I don't really care

[left jacinth]

^

[left jacinth]

is there a place to find the logic for how hardware computes the scan-lines so that you make sure they match?

I didn't do too much searching but didn't find anything obvious

I don't really even know where to look

[wicked notch]

exam was so easy I almost forgot it existed (I was late thank god the prof was too )

[left jacinth]

dang

[rocky schooner]

Does the market have any demand for OpenGL?

[wispy spear]

the market has demand for people who understand graphics and compute pipelines, the api does not matter much

[buoyant summit]

no, I personally don't demand any opengl, tyvm (I'm the market)

[frank sail]

The market doesn't demand any API in particular

Customers (of any variety) dgaf about graphics APIs beyond what they may see when launching a game

Hmm, imagine launching cod 202x and seeing the choice to use fwog or vuk backends

add daxa to the list

[wicked notch]

how do you know cod doesn't use daxa already

potrick may be pulling some strings there

[velvet marsh]

the way I read this is that the market demand is for engineers who don't have problems learning a new graphics api quickly

[wicked thorn]

a

b

c

d

20000

[wispy spear]

damn, italians do talk a lot

🤏 🤌

[wicked notch]

real

[ebon ruin]

I know i lack context

but why ask here of all places?

[severe dome]

it is a graphics server

[ebon ruin]

but specifically in this thread

[severe dome]

oh lol i didn't even notice that lmao

[primal shadow]

@dull oyster something I thought of, given that we need to calculate the screen space AABB for occlusion culling anyways, why even bother with calculating the screen space diameter/radius of a cluster using a different formula? I guess it's slightly cheaper, which matters when you have millions of clusters that are likely to be skipped?

[primal shadow]

I'm the GPU now: https://github.com/JMS55/bevy/blob/meshlet-sw-raster/crates/bevy_pbr/src/meshlet/software_visibility_buffer_raster.wgsl

[wicked notch]

GPU lost GPU rights

only compute machine now

[pale horizon]

Time to do a software renderer on GPU

[loud crag]

the RTX 60 series will just be a big fat compute processor with nothing else

[buoyant summit]

we already have these it's called data center GPU

AMD doesn't put drawing hw into that at all, neither they put format conversion and sampling stuff

NV has some very tiny and weak drawing hw and not sure about sampling etc

[primal shadow]

Hmm, I can't figure out how to map the meshoptimizer meshlet triangles/indices into a meshlet-local index

let index_ids = meshlet.start_index_id + vec3(triangle_id * 3u) + vec3(0u, 1u, 2u);
let indices = meshlet.start_vertex_id + vec3(get_meshlet_index(index_ids.x), get_meshlet_index(index_ids.y), get_meshlet_index(index_ids.z));
let vertex_ids = vec3(meshlet_vertex_ids[indices.x], meshlet_vertex_ids[indices.y], meshlet_vertex_ids[indices.z]);
let vertex_1 = unpack_meshlet_vertex(meshlet_vertex_data[vertex_ids.x]);
let vertex_2 = unpack_meshlet_vertex(meshlet_vertex_data[vertex_ids.y]);
let vertex_3 = unpack_meshlet_vertex(meshlet_vertex_data[vertex_ids.z]);

I have 64 vertices and 64 triangles max per meshlet

You can take the triangle_id [0, 64) and map that to an index [0, 192)

Then adding that to the start_index_if of the meshlet gives you the position within the larger buffer

And then you can do the same thing with vertices

But now I have a 64 thread workgroup load 1 vertex per thread into shared memory

And then 1 thread per triangle can load those vertices and build the triangle

let index_ids = meshlet.start_index_id + (local_invocation_id.x * 3u) + vec3(0u, 1u, 2u);
let index_1 = get_meshlet_index(index_ids[0]) - meshlet.start_vertex_id;
let index_2 = get_meshlet_index(index_ids[1]) - meshlet.start_vertex_id;
let index_3 = get_meshlet_index(index_ids[2]) - meshlet.start_vertex_id;
let vertex_1 = screen_space_vertices[index_1];
let vertex_2 = screen_space_vertices[index_2];
let vertex_3 = screen_space_vertices[index_3];

But I'm not actually sure my index calculations are correct here. I don't think subtracting by start_vertex_id works to get the meshlet indices back into the [0, 64) range :/

[wide shadow]

Might be interesting. Now you can lock vertices when simplifying and new algorithm for optimizing mesh let's for Nvidia GPUs
https://github.com/zeux/meshoptimizer/releases/tag/v0.21

[primal shadow]

Tested on bevy's virtual geometry renderer 🙂

[wispy spear]

lustri got sucked in completely by the education programs

[primal shadow]

Zeux isin't in this discord, right? I have a question :/

I'll use github discussions

[wispy spear]

he used to

hes on the vk server iirc

[primal shadow]

There's a VK server? til

[wispy spear]

-,-

[primal shadow]

https://github.com/zeux/meshoptimizer/discussions/711

[wicked notch]

2/3 exams done

last one and I graduate let's goo

[wispy spear]

good luck with the last one my froggi

[primal shadow]

gl

Bunnies fight in hell (bugged)

[dull oyster]

won't the error depend on the cluster group's orientation relative to the camera? I am not sure how you could compare the parent group's error to the current group's error with AABBs.

[faint crane]

Incredible that every single statement is wrong.

[dull oyster]

"not compatible with raytracing", not with that attitude that's for sure

also "not that useful compared to traditional LOD", artists and coders have been struggling a lot with them, a Nanite-like system more or less makes it automatic

well obviously I'm preaching to the choir here
but still

[wicked notch]

the cooking's been on hold for a while now sadly, but I've been thinking a while about RT with nanite

[wicked notch]

do you make a BVH per cluster group in your thingy?

[dull oyster]

My tests are currently on hold for the GP Direct demo, but that's what I stopped on: a BLAS per cluster group, and an instance inside the TLAS per instance of the cluster group*

[wicked notch]

did you ever get around measuring tracing and building perf (for the TLAS)?

compared to full LOD base BLASes

[primal shadow]

Traverse Research has also done RT with hierchal LODs, but idr if they ever explained how

[wicked notch]

they promised a followup post but it's nowhere to be found :(

[dull oyster]

I did not attempt to measure the difference between LOD 0 and this BLAS-per-group attempt

Partly because it is not finished
Partly because I don't care, I want to stream the cluster in and out of memory, so I will need to be able to do it with raytracing enabled too

[primal shadow]

Contact them and ask them. I know @ Darius (they're on this server) worked on their GI stuff, so you could ask them, or ask if they can get you in contact with the person who did

[wicked notch]

ye the streaming is a big advantage, coupled with BLAS serialization too

[dull oyster]

though in my tests, tracing looked ok, did not notice that big of a difference
but building is a disaster with my current implementation

[dull oyster]

once my video for GP direct is done, I want to check if BLAS serialization can be a way to "fix" my perf

[glass sphinx]

such a dum opinion

everyone uses compressed clusters anyways lmao thats already not compatible with rt

[delicate rain]

I'll be doing it as my dp most probably

There already is a paper on it

[wicked notch]

damn

we don't deserve saky

[delicate rain]

https://graphics.cs.utah.edu/research/projects/ray-tracing-hw-adaptive-lod/

Not done by me

Lmao

Soon though perhaps 🧙‍♂️

I'm kinda cooking the topic while on vacation, exploring directions

[wicked notch]

huh this is actually not nanite

[delicate rain]

There are also two papers from adobe, talking about displacement without tessellation for rt

[delicate rain]

It's not?

I didn't read yet, but I got the impression of it being nanite

[wicked notch]

it doesn't rely on meshlets at the very least

[frank sail]

more like nanot

[wicked notch]

I'm reading it for the first time rn

[delicate rain]

I'll try to read it today or tomorrow

[wispy spear]

its from the utah pot people?

[wicked notch]

it actually looks like something Brian Karis talked about before going to meshlets

[delicate rain]

I think so

[delicate rain]

They get some nice results, I wonder if exploring meshlets with this might be viable

I have a lot of learning ahead of me

[wicked notch]

I was looking at this some time ago: https://www.intel.com/content/www/us/en/developer/articles/technical/real-time-ray-tracing-of-micro-poly-geometry.html

this is actually good ol' nanite

[delicate rain]

Nice nice, thank you!

[wicked notch]

Intel suggests creating new BVH formats inside Vulkan/DX12 to make nanite + RT easier

[wicked notch]

yeah ok this one is """"just"""" tessellation but RT

very nice

[delicate rain]

https://dl.acm.org/doi/pdf/10.1145/3478513.3480535 this one is also cool

Regardless of that, I wanted to say that I'll probably be hopping onto the nanite hype

Just in rt world

[wicked notch]

let's fucking gooo.meme

https://tenor.com/view/rage-emoji-rage-gif-2922587064166244739

[delicate rain]

Btw what are good resources for nanite, still the deep dive into nanite video, or something else?

[dull oyster]

The slides from the presentation have a lot of information: https://advances.realtimerendering.com/s2021/Karis_Nanite_SIGGRAPH_Advances_2021_final.pdf

If you want to look at actual implementations:
https://jms55.github.io/posts/2024-06-09-virtual-geometry-bevy-0-14/
https://jglrxavpok.github.io/ (shameless plug)

[wicked notch]

@ deccer could you pin this so I know who to blame when I dive deep into this rabbithole

[wispy spear]

(you guys can ping me any time, no need for the space :P)

[wicked notch]

I found the thing

https://youtu.be/NRnj_lnpORU?t=809

[wicked notch]

@delicate rain this is basically what that paper on RT tessellation is about (except they use a more complicated watertight type of tessellation scheme)

[primal shadow]

Can I get the TLDR on cluster compression. Rn I have a very basic scheme of just running the serialized bytes through lz4. I plan on making tangents implicit, and using oct-encoded normals, which brings me down to 32 bytes/vertex, but I don't have any further plans for compression, much less per-cluster.

[glass sphinx]

normal tangent bitangent fit in 32 bit and any other attribute compresses well with delta compression across the meshlet

[primal shadow]

Do you save data that way vs refrencing a single set of vertices for all meshlets?

I'll have to look into delta position/uv though

Iirc nanite makes position relative to the cluster bounding sphere center

Not sure if that's only disk, or runtime too, as extra fetches for the bounding sphere data seems bleh

[delicate rain]

Perfect, I will consume the knowledge

[wicked notch]

one more wrinkle for the brain

[glass sphinx]

stored per meshlet

[primal shadow]

@wicked notch you use the screen-space AABB size of the meshlet for determing SW/HW raster right? For the first pass, do you compute one AABB using last frame's transform for culling against last frame's depth pyramid, and a second using the current frame's transform for choosing SW/HW?

[wicked notch]

yes

[primal shadow]

I have most of software raster prototyped, just need to write the actual triangle raster code 😛

Spent some time removing the off-the-shelf serializer I was using in favor of writing the bytes out myself. Meshlet asset loading is ~9x faster now, and probably uses less temporary memory https://github.com/bevyengine/bevy/pull/14193

[buoyant summit]

tbf it's kinda annoying with current AS build APIs

but also lods aren't as necessary with rt as they are with rast

shorter build times, less memory and only somewhat faster trace

[delicate rain]

In the paper I posted earlier they did seem to provide a significant speedup

[primal shadow]

Started SW raster: https://github.com/JMS55/bevy/blob/3bd72429675286a05eaf1fa1254645c8799b44e2/crates/bevy_pbr/src/meshlet/visibility_buffer_software_raster.wgsl

Not working yet though

[wheat haven]

stealing from reading Retina code has me questioning some of my long-held C++ habits, how dare you try and make me improve

[pale horizon]

stole code from trained my brain LLM on 😎

[wispy spear]

lustri, i wish you good luck for the eksems

[wicked notch]

I've done 16 of them what's one more

[wispy spear]

still 🙂

what sleep deprivation does to a mf

[wicked notch]

real

[glass sphinx]

lvstri we need to absorb you into daxa

https://tenor.com/view/eatthatcookie-gif-10982703

[wispy spear]

😄

[glass sphinx]

grand unified vulkan-abstraction theory

also jaker and martty and everyone else

we can all join and make one crippled library instead of many

😳

[wispy spear]

https://tenor.com/view/rick-and-morty-kiss-the-gif-24878522

[glass sphinx]

at least i feel like there should be a lib to recommend to people other then using opengl thats more of a c api on vulkan

[wispy spear]

webgpu?

[glass sphinx]

yea that is ok

but nih

but we can nih together

[wispy spear]

i should be quiet, im totally unqualified for this

[glass sphinx]

webgpu also has some old-isms because of safety

i think we can be more ghetto

[glass sphinx]

you join too

we need an army

[pale horizon]

vma + vkb + vkguide = you’re all set

[glass sphinx]

vkguide does not have a good vulkan abstraction for regular use tho

its tutorial code

[wispy spear]

im also not a fan and abandoned it

[minor root]

One vulkan lib to rule them all

[wispy spear]

the problem is, that it feels like every other week there is a new way of how one should use vulkan

similar to every week a new js framework comes out which every web-dev needs to use

[frank sail]

the right way is my way

[wispy spear]

: )

and then there is the whole shader compiler bs in between which doesnt seem to allow certain things sometimes and you cant use vulkan as intended or something like that

[frank sail]

hmmm sounds like fuddery

I'm not particularly hindered by using glslang, as far as glsl compilation goes

[wispy spear]

good good

[pale horizon]

Only 1.3 changed quite a lot, and it’s much easier than before now
Sort of like GL changed with DSA

[frank sail]

except with vulkan it's like, actual hardware tiers rather than moronic API limitations being lifted

[wispy spear]

i remember when those tiers came around in dx12, and nobody used them

or dx11_2?

[wispy spear]

yes yes, but even then, when you scroll through our discord, there is not a single opinion on how to use it "properly" 😄

perhaps certain features of modern vulkan make sense for specific use cases only, but that was never clear or obvious to me, when reading/following those lose conversations about "what one should use/do" when it comes to bulkan

an actual guide would be nice (vkguide is not that)

also also dont listen to my random bs too much, i am not even focusing on vk right now, but i will from january onwards! mark my words lol

[pale horizon]

99% of what vkuide teaches is all that you need
You can disregard what anyone else says

[wispy spear]

i dont like vkguide

its half assed

and a similar thing like logl right now - it feels like

[frank sail]

so vulkan offers multiple ways to do everything, but each way has genuine pros and cons

[wispy spear]

yeah

a "guide" (perhaps the wrong word) around those pros and cons for way x y and z could be something to build upon

[frank sail]

a simple guide won't cut it if you want to be able to pick the right tool every time

you have to read lots of stuff

like there's a bunch of ways to upload data, and even though my requirements are pretty narrow it's still hard to choose

[wispy spear]

yeah thats the stuff im talking about,

those things are mentioned across the server somewhat frequently too

but its 20 different people all the time 🙂

[frank sail]

like in #1128020727380054046 right now

[wispy spear]

: )

[frank sail]

maybe you could make a rube goldberg uploading system like what devsh and co. cooked up

[wispy spear]

maybe i cant read, but thats exactly what the pros keep saying

but then its person a vs person b "yes you shouldnt to it this way, do it that way" 🙂

[frank sail]

at some point you have to decide for yourself based on the information presented

[wispy spear]

ofc

if you just want to display voxels for a minecraft clone its perhaps very different

compared to some open world terrain/planetrender streaming thingy

or maybe not, and you could use the same mekanism for both things

[frank sail]

it's also way too easy to overthink unless you have a concrete problem in front of you

[wispy spear]

yep

it would help if you know what you want to make exactly

[frank sail]

most of the time it's kinda obvious what you need to do to solve a problem in vulkan

like most cases of uploading data

don't tell the others I said this, but vkCmdUpdateBuffer works quite well in many cases

[wispy spear]

: )

[frank sail]

e.g. for a basic model viewer you could get away with vkCmdUpdateBuffer for per-frame data and then a simple duplicated buffer for per-object uniforms

or just vkCmdUpdateBuffer if you have few enough objects

[distant lodge]

rip vblanco

[loud crag]

vkguide is my lord and saviour

[wicked notch]

vblanco in shambles (it's joever)

I just used the khronos samples to learn vk honestly

I already knew GL so it was pretty easy

the GL -> Vk pipeline is real

[delicate rain]

Meh imo it doesn't really matter what you use to learn, even if you pick up some not so optimal practices, just listening to people on this server will eradicate most of them anyways

The biggest step imo is understanding the API, everything else you can just learn from other nih abstractions

[glass sphinx]

https://tenor.com/view/fish-sad-blob-fish-gif-15806450

[glass sphinx]

the dream

[distant lodge]

I'm glad vkguide/this server got me off of cached command buffers and on pipeline dynamic state

I think vulkan-tutorial presents this now too

[glass sphinx]

niiiiice

[delicate rain]

Okay this is cope

It basically all stands on the fact that someone will implement a part of the algorithm in hardware

Otherwise it reduces memory bandwidth but compute cost will go nuclear

It's also unfit for use with current hw rt acceleration

So they basically say "currently this will only work in software but it will be slow. If someone adds a new hardware unit and makes it a part of the API it will be fast"

Sadge

[buoyant summit]

this bit is fine it just needs AS build API adjustments

tbf

[delicate rain]

Interesting, I did not know that, but still I was kinda talking from the user perspective

[buoyant summit]

yes

[delicate rain]

As in as of right now it's basically useless for me 😅

[buoyant summit]

ok nvm

yes

that stuff in paper requires hw changes because it affects

but

I think global LOD is still useful

and global LOD doesn't affect traversal

with global LOD you'd just swap in differently detailed BLAS depending hit feedback or whatever

this is arguably more useful for something like a game as it lets you use less memory for the scene than needing full LOD in memory at all times

[delicate rain]

They mentioned that in the introduction - I thought you already could do that with the API (I know close to nothing about the API though)

[buoyant summit]

yes you could but

for very high tri meshes build cost is oof

and that's the bit that would be nice to address

[delicate rain]

I see, I need to experiment and build some intuition

[buoyant summit]

otherwise yes just do global LODs with rt

I think LOD decisions at traversal time are kinda whatever

[delicate rain]

There is also the paper from Intel about micro poly rt stuff, guess I'll read that next

[buoyant summit]

that one you also can't do in current apis

[delicate rain]

But I'm kinda worried

[buoyant summit]

but it only modifies build AFAIU

[delicate rain]

Ah right

[buoyant summit]

so the hw can do it

just the api kinda can't (just for now, hopefully)

well it can but not in a very useful capacity I mean

you'll succumb to AS build costs with the kind of meshes they're working with

[delicate rain]

I wonder how sophisticated the hw BVHs are

[buoyant summit]

@hallow cedar

[hallow cedar]

what

[buoyant summit]

this

[delicate rain]

It seems like that's 90% of all the rt research now - more efficient bvh (aabb-kdop-obb)

[hallow cedar]

ah

nvidia i think quite a bit more than amd, intel potentially too

amd is a pretty standard BVH4, nvidia was doing CWBVH-y stuff I heard somewhere but I'm really not sure

[delicate rain]

I imagine it's all secret, especially the details

[hallow cedar]

well amd's not really

[delicate rain]

But AMD's is slow no?

[hallow cedar]

https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/amd/vulkan/bvh/bvh.h here you go

[hallow cedar]

that's the catch

[buoyant summit]

all vendors likely use a simple bvh of some kind

but some vendors have more interesting hw bits for traversal

[hallow cedar]

speaking of i still have next to no idea about rdna4 rt and that's kinda weird to me

[buoyant summit]

rdna4 removes RT hw

[hallow cedar]

most accurate hw leak

[delicate rain]

It's sad, as I'd like to do something with hw rt but it seems all is either proposals or closed door stuff not really open to academia

[buoyant summit]

bro

get amd

[delicate rain]

So if you really wanna publish/research in that field, you're bound to software

[buoyant summit]

(or intel if you like extra pain)

hw rt is there

the open drivers are there

buy steam deck

you can hack the vk driver in whatever way you want

do whatever

with hw

directly

[delicate rain]

Uh

[buoyant summit]

nv rt in nvk does not yet exist and who knows when will

but the setup for rt on nv is a bit complex so it will be some time

I doubt you'd be up for REing nv prop blob

[delicate rain]

Yeah no and even if I were, there's a bunch of external limitations forced on me so it wouldn't be viable anyways

[buoyant summit]

ye

[buoyant summit]

this bit is very easy

very easy

[delicate rain]

Guess there is still the adobe displacement stuff that looked interesting though

[pale horizon]

good

[buoyant summit]

bro go back to your thread before I remove u from lyf

I swear on me gpu

[hallow cedar]

be careful with that

it might leave you hanging

[pale horizon]

They should remove TAA hw next

[buoyant summit]

what if there's no taa hw 😳

bro wants to remove the entire gpu

[hallow cedar]

CUs can do TAA with shaders -> remove CUs

only command processor pls thxbai

[buoyant summit]

computing on command processor be like

[pale horizon]

Retvrn to software rendering

[hallow cedar]

but what if software taa

maybe we will have to go outside

[buoyant summit]

we have that already it's called datacenter gpu

[pale horizon]

Remove TAA CPU instruction

[buoyant summit]

those don't have drawing hw

or format conversion, image load/store, sampling support

just software

[pale horizon]

Good

[buoyant summit]

just compute

only CUs

no drawing

[hallow cedar]

but what if our eyes do taa internally
your eyes do taa get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out get them out

[buoyant summit]

bro

https://tenor.com/view/spray-bottle-cat-spray-bottle-spray-bottle-meme-loop-gif-25594440

[pale horizon]

Me when I turn my head in game with TAA fr fr

[pale horizon]

My vision is kinda blurry when I turn my head fast
It’s joever 😦

[frank sail]

Same order of magnitude tbf

[delicate rain]

I was mostly meming

Although I've heard some hate from Patrick

[frank sail]

Oh, AMD's rt acceleration is indeed a lot slower, but it's still very usable

[delicate rain]

Last I've heard it's 4-5x faster than software, but maybe that is outdated

I want to do hw rt so much, but it's hard to find something to wrap it with (as in a topic for uni)

[finite quartz]

https://cdn.discordapp.com/emojis/1157634614857314394.gif?size=80&quality=lossless

[fiery bolt]

jaker you don't need to shill amd anymore

[primal shadow]

@wicked notch What's the theory behind the math for this? How did you get this equation? https://github.com/LVSTRI/IrisVk/blob/master/shaders/0.1/rasterizer.comp#L225

[wispy spear]

stolen from ue sources

[wicked notch]

you can see that in AMD's small triangle culling code

I don't feel qualified to explain that (also because it's been so long) but AMD should have everything you need

[primal shadow]

What do they call the project? Looking up small triangle culling code amd did not find anything

[twin musk]

geometryfx possibly

[primal shadow]

I didn't find anything there

Also https://github.com/GPUOpen-LibrariesAndSDKs/WorkGraphComputeRasterizer/tree/main seems empty, and has no shaders?? idk

Somewhere here, maybe https://github.com/GPUOpen-Effects/GeometryFX/blob/02c4139eff3ee9a1c180a25fa4a6020a477f9fee/amd_geometryfx/src/Shaders/AMD_GeometryFX_Filtering.hlsl#L165

[wispy spear]

finish it first

[faint ruin]

oh wait I didn’t even realize Elias was the one that wrote that vulkan article about how he learned wowza

sorry my mind goes a million miles a second sometimes

[wicked notch]

elias made it

insane

[faint ruin]

I wasn’t paying attention to the names until after reading the messages :(((

I’m a little dense sometimes

[wicked notch]

it's ok

don't worry about it chief

[faint ruin]

I’ll finish Vkguide up first
I’ve been super happy with it this far, probably the best thing to do before asking further questions

[wispy spear]

doesnt hurt to take notes about things which make no sense yet

and there is a post (something something vkguide) in #1019722539116802068 from vblanco, the author of that guide

[faint ruin]

thank ya thank ya

[pale horizon]

What was the deleted post about? 😅

[faint ruin]

Knee jerk reaction to delete it buuuut
My initial question was essentially asking once one finishes vkguide, compared to say Vulkan-tutorial, it’s much more applied in that you’re making an engine. The question was essentially in regards to making a smaller renderer inspired by vkguide to better fit a different use case, or if basing a new project off what I did in vkguide makes sense where things do not see

Like, if I wanted to make a simulation application or a tool that lets me test different rendering techniques or run different demos of implemented graphics papers. Or I want to make a nice terrain generator. How can I take what I learn in vkguide into these tutorials in a way that tailors it to my use cases?

[pale horizon]

The renderer in vkguide is already quite minimal, and you’ll likely need much more for a “small” renderer, tbh
You can drop “drawing transparent objects” part if you don’t need it.

Also you don’t need “GPU driven” part most likely, unless you plan to render 1000s of objects

[wispy spear]

i would read the whole thing anyway

[faint ruin]

That’s my plan to at least read up to and through gpu driven rendering (maybe not before actually iterating on ideas), cause I have time to learn and wanna improve my understanding and application of the API

[pale horizon]

Yeah, you can read the last chapter, but just READ don’t implement, haha

[faint ruin]

I have ideas on projects I wanna take on, it’s just a matter of reasonably converting what I have here to my needs.

I also feel a certain way of “cheating” using sample code and tutorials like this without actually understanding the underlying API (which happens through reading and actually getting your hands dirty)

But that’s a personal disposition I’m working on discarding

it makes no sense to have that especially if I am new to Vulkan.

[wispy spear]

you also better create a post in #1019722539116802068 about the projects you have in mind, like the other frogs did/do too

https://tenor.com/view/or-else-wednesday-addams-family-cut-neck-gif-5060381

[faint ruin]

oh no, I will not be subject to the threats of those who came before me
post my projects I shall

[wispy spear]

hehe

[faint ruin]

Vulkan has been fun, learning it has been exhausting but I like the challenge
it definitely helps to have done stuff with other APIs too
I’m hyped, thanks guys

[pale horizon]

You should copy it once
Then you can re-read it again, but start organizing things more to your liking.
Also, the best understanding happens when you debug something, or implement some technique not described in the tutorial.

E.g. shadow mapping, post-processing effects, PBR stuff and so on

I can also highly recommend getting into bindless textures ASAP
Descriptor sets are not that good and bindless + buffer direct access allows you to mostly get rid of them

[faint ruin]

Hard hard agree. That’s where I’m most excited cause I know I’m gonna learn the most through the struggles.
But are descriptor sets really not that good to you?

[pale horizon]

It’s not like they’re “bad”. It’s just managing/allocating/defining them is PITA when with bindless you just pass your textures with push constants and that’s it

[faint ruin]

I see

Push constants seems to be something I might wanna use later on

But yeah, the process for setting everything up with the sets and layouts does seem rather tedious

[pale horizon]

No, you need them right at the start
And vkguide uses them extensively

[faint ruin]

you know this guy made a good point to me to finish it lol

seems like I’ll keep digging through then

[faint ruin]

I was considering stopping cause I felt I was escaping the fundamentals of VULKAN and dealing with engine architecture rather than Vulkan itself, hence why I supplement with the specs and the Wien lectures

[pale horizon]

Push constants are fundamentals of vk 😅
But yeah, it’s good to reference multiple resources. Also TU Wien lectures are quite good indeed.

Don’t worry about “not getting” it - you’ll understand more with practice and when you solve real problems and add code on your own.

I think I know like 10% of vk and still I managed to make something presentable

[faint ruin]

That makes me feel better about going forward on this route of learning

Not necessarily the 10%, but the knowledge that leaning into the fundamentals and just… working with it. You don’t need to learn everything at once

It’s like a fighting game. You don’t need to learn every combo at the beginning

[pale horizon]

Yeah. You’ll rewrite it 10 times anyways, lmao

[faint ruin]

LOL probably true

[faint ruin]

Your article gave me the kick to just go in and get started so I appreciate you talking to me about this. Major appreciate and a follow from me :)

[wicked notch]

man

you ever try to sit down and study

but while you read your eyes go out of focus

I think I'm going to die soon

[hallow cedar]

while you read
wish I'd get that far

i know the "going out of focus" stuff stuff though

it usually happens when I'm tired

[wicked notch]

it's barely 5pm tho

the 32°C of hellfire in my room doesn't help

[hallow cedar]

ah you too

i had 35°C earlier this week

death

[delicate rain]

Did you just start?

Usually happens for me the first hour or so that I start

Once I get into it properly it's usually fine

And then once I get tired it comes back yeah

[buoyant summit]

no AC?

today was my first moderately productive day in a while actually

did some CTS work that I can't talk about details of

god VK-GL-CTS is a chonker

[hallow cedar]

it's insane

on my old laptop I had to basically compile until I run OOM, then compile with lower thread-count (singlethreaded at time) for a bit, then abort and continue with higher thread count

because -j1 is incredibly dog slow but -j6 (yes it's a hexcore 🐸) makes me run oom

[buoyant summit]

buy chungus ram

why are there no laptops with option for 64G

or 128G

pls

ayymd how can I survive off 32G with a 12c SMT2 part

my current work laptop is actually a

4c SMT2 16G thing

which is...

well..

it works

[hallow cedar]

i was considering retrofitting 64g into my laptop

[buoyant summit]

with swap it's actually fine, so browser etc can get swapped out while important stuff gets to use ram

[hallow cedar]

went with 32g for now and it's ok

[buoyant summit]

you should

32G felt a bit uncomfortable on my 5900X PC

[hallow cedar]

yes on pc i'm 64G

on laptop i'm 8c16t32gb

but mediumterm I kinda want to look at alternatives

[buoyant summit]

I just want a nice reasonable laptop

[hallow cedar]

when I get a non-shit internet connection I'm considering just having a remote power switch for my pc and doing remote-y dev

[buoyant summit]

where 5p69e cpu pls

[hallow cedar]

I suppose my case is sorta special though

[buoyant summit]

ye that sounds nice

[hallow cedar]

I need chonker gpu for rt, but I've come to hate gamer bricks

[hallow cedar]

(the idea would be like a raspi that operates some digital power/reset button in order to do power cycling and so on)

though I'd like to do boot/kernel selection perhaps

hmmmm

anyway I should probably ramble about this in my own thread instead 🐸 sorry

[glass sphinx]

slam stimulants

[wicked notch]

'tis but a temporary relief

btw it's two hours later and the situation has barely improved

here's a conversation between me & my friends about the situation at hand (the exam is tomorrow)

first message is "how is it going"

[loud crag]

L

[wicked notch]

real

[velvet marsh]

it's going to be ok, unless it won't

[buoyant summit]

wym temporary

tbh heat is also temporary

[frank sail]

life is also temporary

[velvet marsh]

do you all need a hug today

[wicked notch]

I need to pass this godforsaken exam

[wheat haven]

what subject

[wicked notch]

image processing and neural networks

it's 6 credits but it should be 1000

there's like 5 books we had as material

my notes are 600k words and a 400 pages

it's insane

[velvet marsh]

sounds over the top and unnecessary

[wicked notch]

lads

it is

finally

(for real this time)

over

[faint crane]

Is it over, or are we so back?

[wicked notch]

we are truly back

exam time is finally over

[delicate rain]

Ez clap

Time to shill some gp

[wicked notch]

ngl I was so lucky

I knew nothing about morphological processing (had to skip because it was impossible to understand all that garbage in a few days)

but they didn't ask me about it so

[frank sail]

you did it laddie

[frank sail]

damn son

[delicate rain]

Wtf

I think I took like 500 words worth of notes accumulated across all my years studying lol

[frank sail]

I was so stupid I would take "notes" in math lectures just copying what the prof wrote, hoping it would make me magically understand the topic better

normally I'd barely write down anything though

[delicate rain]

I did that a little the first year

But yeah just listen and try to understand

Much better spent time than copying and not catching what he says

[frank sail]

my math prof was foreign and had an accent that I could hardly understand (my bad hearing didn't help either)

[delicate rain]

Oh god, yeah that does not sound good

[frank sail]

anyway I somehow pulled through

[delicate rain]

Having pretty much all lectures on YouTube kinda allowed me to get away with almost no notes too btw

[pale horizon]

LVSTRI gonna come up with novel waifu generators next

[wispy spear]

wb lustri, time to bikeshed UE6 clone

[loud crag]

@wicked notch where did you get this piece of code from?
https://github.com/LVSTRI/Retina/blob/1152267563683bb6a8a6731ce31b8bc07ffe88c4/src/Retina/Sandbox/Shaders/Visbuffer.mesh.glsl#L99

    const vec3 v0 = sh_ClipVertices[indices.x];
    const vec3 v1 = sh_ClipVertices[indices.y];
    const vec3 v2 = sh_ClipVertices[indices.z];
    const float det = determinant(mat3(v0, v1, v2));

    gl_MeshPrimitivesEXT[currentIndex].gl_CullPrimitiveEXT = det > 0.0;

since idk why this works, I'm a little confused

because that does backface culling, right?

since it does det > 0.0

however in my Metal renderer that does frontface culling, and I need to do det < 0.0 instead

is this related to Y+ being up instead of down?

[wide shadow]

the code is from here I guess and if you do proj_mat[1][1] *= -1.0f to flip Y axis then you do frontface culling instead of backface culling
https://zeux.io/2023/04/28/triangle-backface-culling/

[wicked notch]

yes this is related to your coordinate system

[loud crag]

ok thank

[velvet marsh]

a negative determinant indicates a reflection

so it's not just a up thing, but a handedness thing

you could have a different up axis and still have a positive determinant

and you could have the same up axis and a negative determinant

[primal shadow]

Help me brainstorm names that are better than meshlets, but less verbose than virtual geometry 😅
I was thinking maybe MicroMesh, but that's kinda a thing already (RT technique)

[wicked notch]

meshlet = cluster

if that's what you want

[wispy spear]

micromesh is a thing already

[primal shadow]

No a name for the user-facing feature in bevy. Something catchy like nanite.

[wispy spear]

i linked the papers somewhere in frogfood a week ago or so

[primal shadow]

In my code meshlet = from asset, cluster = instance of meshlet

[wicked notch]

that's weird

could've just done cluster and clusterInstance innit?

or is that reserved somewhere else

[primal shadow]

Ehh that's a bit verbose with how often I use the word cluster 😅

[wicked notch]

this is a matter of preference ofc

personally I'd rather a little bit of verbosity if that means not using two synonymous words in different contexts

[primal shadow]

It's subject to change, I need to clean all my code up, it's on the TODO. I'll think about it.

[primal shadow]

I want to try and come up with a better name for the feature though since it's confusing that it's called meshlet, which is also used everywhere in the code 😛

[wispy spear]

who cares whether rust goes into nodejs? (im sorry if thats you cody)

[wicked notch]

you could reuse what unreal calls HLOD 😈

[faint crane]

Nah, alcohol as a naming scheme.

[wispy spear]

ah

[wicked notch]

(in UE HLOD is a very different thing from virtual geometry)

[primal shadow]

We have HLODs already as a seperate thing, although I think we call them VisibilityRange

[wispy spear]

just keep the same naming scheme jasmine, everyone doing it knows what x and y means, and you cooking up some new terminology will be conchfusing

unless you want to coin those new shit for some weird reason in the bevy world, to stand out or something idk

[primal shadow]

Ok forget the variable naming

I need a new name for the rendering feature itself, like Nanite

[wispy spear]

Bevvite 😉

[primal shadow]

Currently it's just "MeshletPlugin" and "MeshletMesh"

Which are terrible

[wispy spear]

Meshlette

MeshlettifiedMesh

[primal shadow]

Those are so much worse 😅

[wispy spear]

ey 😄

why not just Mesh

and the engine decides whether it needs to meshlettify it or not, via some switch

[primal shadow]

We have a Mesh, but it's an entirely seperate thing, and would be very confusing to users who don't use the feature

[wispy spear]

remove it then

[wicked notch]

do you fancy acronyms?

[primal shadow]

Rip transparency and animated meshes I guess 😛

Sure, acrynyoms are fine

[wispy spear]

and make Mesh 2.0 the new Mesh

and add animations back in heh

[wicked notch]

CBVHLOD (cluster based virtualized hierarchical level of detail )

[primal shadow]

Also remove half of bevy's userbase... Even my brand new laptop dosen't support Int64 atomics (thanks intel)

[wicked notch]

intel skill issue

[wispy spear]

thats what you get for targeting peasant hardware

[primal shadow]

Oh, I'm like I don't actually do the BVH part yet, but it dosen't stand for that lmao

[wicked notch]

yes yes, the confusion is part of a good acronym

[primal shadow]

Arc is annoyingly stuck at SM 6.5 and not 6.6

[wicked notch]

the more confusing the better the acronym

[primal shadow]

I think I'll just go with VirtualMeshPlugin and VirtualMesh 😛

[wicked notch]

but yeah naming is hard

[primal shadow]

What makes a mesh "virtual"? Idk it's a meaningless term, but it sounds good

[wicked notch]

I think the virtual is mostly referring to virtual texturing, as nanite author explains it the logical transition was "virtual texturing" (decoupling memory budget and texture sizes from rendering) to "virtual geometry" (again, removing concerns about poly budgets and automating LODs)

[primal shadow]

No I know, but for like non-rendering people, virtual is such a meaningless term lol

[wicked notch]

o

yeah idk how to explain it to non rendering people

[primal shadow]

The other annoying thing with VirtualMesh is that it has mesh in the name

So now I'm going to have dumb variables like virtual_mesh_meshlets

Which I mean, ig it's fine

Best I can come up with

[wicked notch]

tell your users "you better pray I don't call my next variables x and y"

make them appreciate

[velvet marsh]

LLM's are pretty good at coming up with names imo, at least sparking ideas

vmeshes is my contribution

[primal shadow]

VMesh is not bad, thanks

[frank sail]

nangons

nanogons

minimesh

[wicked notch]

nanogoon

[frank sail]

geographs

[faint crane]

nanot

[buoyant summit]

megageometry

megametry

ur welcome

[buoyant summit]

bro I swear one more and your walls will collapse

[primal shadow]

@wicked notch do you understand nanite's scanline code? I don't get how it's solving for the passing X-interval

It's also not working for me 😦

[primal shadow]

Nvm think I got it

Idk why I swear I've tried this code before, but now it's suddenly working...

Mostly working, some pixels are slightly off still, but close!

[primal shadow]

Haven't finished the heuristics for SW/HW render switching, but only getting maybe a ~10% speedup vs HW only... Nowhere near what Nanite is quoting.

On the plus side I save a lot of memory having to only allocate data per-cluster instead of per-triangle, as I don't have access to mesh shaders to make HW raster fast

[wispy spear]

time to extend webgpu to support mesh shaders

[primal shadow]

The wgpu side stuff wouldn't be so bad, I don't look forward to the naga changes needed though :/

I suppose I could use spirv passthrough and write it in glsl instead ig..

[wispy spear]

what is naga? the shader middleware?

[primal shadow]

Shader transpiler. Wgsl -> msl/hlsl/spirv/glsl

[wispy spear]

ah

[fiery bolt]

possibly has to do with the heuristics, but also shader micro-optimization and triangle density you're testing on

[primal shadow]

I'm guessing it's the triangle density yeah. There's not really much left to optimize in the shader.

[fiery bolt]

try out some raw megascans assets perhaps

[wispy spear]

@primal shadow https://youtu.be/xT_-oSo-w-E?t=601 that reminded me of your shtuff

[primal shadow]

Thanks, will take a look in a bit

Trying to figure out why tf triangles wrap around to the other side of the screen when doing scanline...

[primal shadow]

Not much that applies to me idt, but early-depth test (do non-atomic load and depth test, before atomic write) seems interesting, I can't actually test it though, as wgsl dosen't allow mixing atomic/non-atomic ops on the same buffer.

Also so many interesting RTGI papers from HPG this years, mainly https://www.highperformancegraphics.org/slides24/hpg24_oscgi.pdf, https://www.highperformancegraphics.org/posters24/Salmi - Fast Local Neural Regression For Low-Cost Global Illumination Denoising.pdf, and https://www.highperformancegraphics.org/posters24/Kawala - Work Graphs based Denoising for Real-Time Ray Tracing.pdf

I don't have any time to experiment though, as I need to focus on meshlets :((

[wispy spear]

ah oki

[primal shadow]

And my sutpid fucking shader won't work and renderdoc is giving me different results when viewing different threads

Any idea why renderdoc says thread 2 writes a "good" value to LDS, but then threads 0 reads garbage at index 2??

[wispy spear]

schrodingers triangle

[frank sail]

I don't think RenderDoc simulates shared memory

Well, interactions between multiple threads

[primal shadow]

Really? Ahh

Well then idk how to even debug this tbh

[frank sail]

Ok, it doesn't run other threads at all

To debug a compute thread simply go to the compute shader section of the pipeline state viewer and enter the group and thread ID of the thread you would like to debug. This thread will be debugged in isolation with no other threads in the group running.

This means there can be no synchronisation with any other compute thread running and the debugging will run from start to finish as if no other thread had run.

[primal shadow]

Yeahh idk how to even go about trying to figure out to figure out why my triangle is wrapping around the screen when it should be clipped

[frank sail]

Make a version of it that doesn't use shared memory

[primal shadow]

I guess do all the math on the CPU?

Oh, good idea lol

[faint crane]

Didn’t sign up but I’m at SIGGRAPH if you want to show your frogs.

Or bunnies apparently.

[wispy spear]

unfortunately im not in the gaming/graphics industry and therefore neither at siggraph nor hpg nor any other giraffics related conference : (

[faint crane]

I can spread the gospel of VSM or whatever acronym we want to give it.

[wispy spear]

please do 🙂

[primal shadow]

Ok so I've learned so far that for the X-axis, the triangle's bounding box min > max

Which is obviously backwards...

How even though?? ahh

// Compute triangle bounding box
let min_x = u32(min3(vertex_0.x, vertex_1.x, vertex_2.x));
let min_y = u32(min3(vertex_0.y, vertex_1.y, vertex_2.y));
var max_x = u32(ceil(max3(vertex_0.x, vertex_1.x, vertex_2.x)));
var max_y = u32(ceil(max3(vertex_0.y, vertex_1.y, vertex_2.y)));
max_x = min(max_x, u32(view.viewport.z) - 1u);
max_y = min(max_y, u32(view.viewport.w) - 1u);

I think it's because I clamp the max values to be within screen bounds, but not min...

https://github.com/LVSTRI/IrisVk/blob/master/shaders/0.1/rasterizer.comp#L94C1-L96C6 stealing @wicked notch 's code should fix it 😛

[wicked notch]

it's UE's so beware

[primal shadow]

Ah ok

Still, I figured out the issue myself, and I just wanted to check if it's ok to early-out here or not

Which yeah, it's fine

Ok so that fixes the weird wrap-around issue

Things are still a bit pixelated with the scanline variant though, not sure why

might be an off-by-1 pixel somewhere

Like something's weird here

[primal shadow]

Idk how Nanite quotes 1.1ms for raster time in lumen in the land of nanite

That's insane

[fiery bolt]

the persistent threads bvh traversal probably saves a bunch of time, as well as whatever micro-opt they did in the compute rasterizer

still, wasn't it 2-3 ms?

or was that the second demo

[primal shadow]

No it was 1.1ms just for the raster, nothing else

The only thing I can think of is I use 64 triangle clusters, and tend to have low fill rate. They use 128 triangle clusters (better warp occupancy), and probably do better on fill rate.

[fiery bolt]

meshopt asserts at more than 124 tris per meshlet doesn't it

and the simplifier doesn't work all too well (it went from simplifying 14000 -> 144 meshlets once)

I need to make my simplifier work

[wicked notch]

I'm cookin lads

I'm cooking something special

[primal shadow]

What of?

[wicked notch]

a special sauce

the sauce that clusterizes and simplifies stuff

that I promised jaker 10 decades ago

[faint crane]

There was a presentation on dynamic LoD and “nanomesh” for mobile at the part 1 of “advancements in real-time graphics in games” course. Slides expected to be out next week.

I recall them having a 32 bit visibility buffer which stored cluster and triangle id at 25 and 7 bits. Also an exponential distance based heuristic for selecting a cut and culling.

[dull oyster]

an exponential distance based heuristic for selecting a cut
Exactly what I need for selecting clusters for raytracing 👀

[faint crane]

They have another presentation Thursday. I’ll have to try my hand at it when slides are out since I couldn’t follow verbally.

Results were impressive for high and low end mobile though. Very encouraging.

[frank sail]

let's friggin go

[primal shadow]

Slides are already up somewhere, they gave the presentation earlier

At gdc iirc

[fiery bolt]

yes pls my metis is giving me completely unconnected meshlets in my partitions

[faint crane]

Your Bevy article was referenced on screen.

[fiery bolt]

i'm getting trolled by metis

[primal shadow]

What screen?

[faint crane]

Visibility Buffer Rendering during part 2

[primal shadow]

Oh, cool! Are the slides posted anywhere?

Website doesn't have them yet 😦

[faint crane]

Sometime next week. Photography isn’t allowed so I didn’t try.

You can have this though.

[wide shadow]

might be interesting for benchmarks
https://github.com/activision/caldera?tab=readme-ov-file

[frank sail]

loading it is a pain in the butt I'm finding

it's split into a million tiny usd files

[fiery bolt]

dumb them all into blender, pray it doesn't crash, export as gltf, pray it doesn't crash, repeat until it stops crashing

[glass sphinx]

no way blender can handle that

prob faster to write a cpp program to do that

[fiery bolt]

it finally works

[wispy spear]

is that the atomium sculpture in brussels

[fiery bolt]

that's a random table i stole obtained legally from megascans at max quality

[fiery bolt]

(it is repeated 1536 times in that scene)

(for a total of 3 billion tris)

[wicked notch]

it is a truly painful lib to work with

plus it was never meant for this shit

I'll roll my own

and then give it to you frogs

[buoyant summit]

make sure it's written in rust

or I will oxidize the pins on your cpu

or pads, depending on which ones your cpu has

[wicked notch]

this channel is anti rust

I like my iron clean and with no oxygen

[fiery bolt]

all my code is rust

[faint crane]

This was part of the talk BTW. https://x.com/SebAaltonen/status/1818429250908610782

Sebastian Aaltonen (@SebAaltonen) on X

If you spam enough in Twitter, your Tweet might eventually transform into a SIGGRAPH presentation slide :)

Thanks for @FilmicWorlds for a great presentation!

Twitter · Jul 30, 2024, 11:31 PM

"Variable Rate Shading with Visibility Buffer"

[primal shadow]

Ok so I peeked at nanite's code and was able to fix my scanline SW raster

Adding the "is point in triangle" test to each pixel of the scanline instead of assuming it's covered fixed it

Not sure why that's neccesary, I guess some kind of partial coverage/subpixel thing

[wicked notch]

man

https://tenor.com/view/jeonghan-horse-on-beach-jeonghan-man-jeonghan-jeonghan-silly-horse-on-beach-man-gif-13991646989968971533

the kitchen is overloaded

why are there so many edge cases

I've been doing this for hours and I'm nowhere near done

and it's 3am

guess I'll keep going tomorrow, my eyes are not staying open

this reminds me of self balancing rbtrees but worse

"if the node has no black successors, at least one ancestor is red or if it has no left children and at least one right child or if ..."

[ebon ruin]

i had a feeling you dont sleep

sleep bro

or you’ll become munted and die

[fiery bolt]

edges cases for what

[primal shadow]

Did a massive refactor of the >1000 lines MeshletGpuScene into separate parts. Took ages, but much cleaner now, minus one of the new subsystems that's basically a copy paste. I still need to clean that up, and then optimize things so that we're not spending ~1ms/frame of CPU time on extract + preparing resources. https://github.com/JMS55/bevy/commit/b8ab371a2566c5c0f2fe743224854f8cad452b12

Overall CPU timings

[faint crane]

Some stuff on meshlets from HPG: https://gpuopen.com/download/publications/DGF.pdf

https://www.youtube.com/watch?v=lx1v-BczaBM&t=19005s

[wicked notch]

pog

gotta read this

[faint crane]

Also some good stuff for BVH fans. https://gpuopen.com/download/publications/HPLOC.pdf

https://www.youtube.com/live/qPmwbp7A3BQ?t=2151s

I only attended SIGGRAPH, missed this was going on until they held a concluding panel at SIGGRAPH.

Just need a METIS killer.

[wicked notch]

I'm cooking

I only need to figure out testing because spamming node allocations is fun and all but eh

I wonder if someone made some test suite for this algo

[faint crane]

Test in prod. PR to Unreal.

[wicked notch]

you know what

you're actually a genius

I forgot I had ue cloned and built

I'll just drop my shit in there and see what happens

[faint crane]

Cat is in lap. Maybe I gained a brain cell.

METIS but with good parameter names would be the most exciting thing out of this last week TBH.

I tried porting it once, but saw a bunch of GOTOs and cursed control flow.

[primal shadow]

Iirc it's proposed for hardware implementation, not (neccesairly) software, although don't quote me on it

[faint crane]

Froge computing when?

[primal shadow]

This paper seeks to close the gap by defining a block-compressed geometry format that is designed for arbitrary geometry topologies and can be directly consumed by future fixed-function hardware.

It's a bit of both, they say you can use it in mesh shaders and stuff, but also suggest it as a future hardware-level representation

[primal shadow]

Capped off the weekend by opening up the next meshlet PR and writing up the description for it https://github.com/bevyengine/bevy/pull/14623. Once this is merged, I plan to improve the CPU performance, and then either look into fixing occlusion culling bugs, or do persistent threads style culling and save a large chunk of time + eliminate a large amount of memory allocations.

[fiery bolt]

occlusion culling bugs

that's what I'm trying to fix (or was anyways until I decided to rewrite all my shaders in slang)

[primal shadow]

I'm using SPD for my depth pyramid generation, which is defeinitly not conservative 😛

I might just have to give up and write a more complex, slower, multi-dispatch downsampling pass

[fiery bolt]

you can set your own reduction tho

[primal shadow]

Only the reduction op (e.g. average, max, min, whatever)

But the problem is sometimes you need to read more than a 2x2 when downsampling

Something like that, I haven't looked into it too much

[fiery bolt]

yeah a max (for inverse z) should be conservative

[primal shadow]

Read https://miketuritzin.com/post/hierarchical-depth-buffers/

[fiery bolt]

I was also fixing my SPD before rewriting everything in slang lol

dxc infinite looped when I made my 6th mip globallycoherent

[primal shadow]

Fun. I don't even have access to that, I just split it into 2 dispatches (the second is only needed if you have a large enough initial texture, which is usually no)

[fiery bolt]

isn't the limit 512x512

[primal shadow]

idr off the top of my head

[fiery bolt]

it'd be 2^7 or 8 depending on how you generate mips

unless you go to quarter res directly

[fiery bolt]

@primal shadow i noticed you use the current camera view in the early pass - shouldn't that be the previous frame view?

[primal shadow]

So I do, crap. previous_view is imported, but it somehow isin't used anymore... I could've sworn I had fixed this. Thanks for pointing it out.

Oh you know what, maybe I bind previous view as the current view? Hmm

[fiery bolt]

then you'd be frustum culling and making lod decisions against the previous view

[primal shadow]

Yeah it's not that

I really don't know how this got messed up, because I explicitly fixed it in a PR dedicated to this kind of thing 😅

Maybe a lost commit? idk

[fiery bolt]

lol

[primal shadow]

I'm so confused what my code is doing 🤔

Oh ok I think I've fixed it

This should be it https://github.com/JMS55/bevy/commit/d49854ab521c09b997161a66605a31ab6c55d6d2?diff=unified&w=0

[fiery bolt]

hmmm 90ms for 3 billion tris, I should probably get started with the bvh and persistent threads

I'm gonna TDR so much

[wispy spear]

increase the TDR timeout

[dull oyster]

render 1 billion tris and temporaly accumulate them

[wicked notch]

are you actually bound by culling tho

or just raster

[fiery bolt]

it shouldn't be raster, my lods and occlusion culling sorta work

[wicked notch]

did you profile 🐸

[fiery bolt]

and nosight shows me that my task shader has the shittiest occupancy known to man

like, 8 warps per SM

30% instruction issue rate

[wicked notch]

show TS

[fiery bolt]

TS?

[wicked notch]

task shader

[fiery bolt]

ah

the worst code you're ever gonna read

[wicked notch]

I mean you could've done the meshlet emit count thingy with subgroup ops instead of atomicAdd'ing one every time

but it's fine otherwise

[fiery bolt]

as in subgroup add, then index 0 does the atomic?

[wicked notch]

doesn't need the atomic anymore innit

because your group size will be equal to subgroup size

[fiery bolt]

my workgroup size isn't subgroup size

[wicked notch]

yes

you will set it to that though if you want it to work

with subgroup ops that is

[fiery bolt]

well I guess

[wicked notch]

but it's fine either way

[fiery bolt]

yeah

[wicked notch]

profile your shader and look which inst is taking up the most cycles

[wicked notch]

me when LGSB

[fiery bolt]

actually, can nosight tell me the number of mesh workgroups dispatched

[wicked notch]

update to the beta nosight

[fiery bolt]

wait where do you get beta versions

am i stupid

[wicked notch]

it's actually just the 2024.2 version

[fiery bolt]

oh

lmao

[wicked notch]

nsight is in release only the other stuff is in public beta

I'm dumb

[fiery bolt]

[wicked notch]

I just updated too

when september comes around I need to see if I can ask my supervisor to sign an NDA with NVIDIA

so I can get nsight pro

[fiery bolt]

where do i get the mesh shader stats, i've been using 2024.2 for a bit an never seen them

[wicked notch]

it's been a long time since I last used nsight

(literally only 2 months)

[fiery bolt]

i need to get myself a supervisor, i'm just a lowly first year

well technically second soon

top stall 1: LGSB 49%

i wonder if i can (ab)use mesh shaders to switch between hardware and software raster in the same shader

SetMeshOutputCounts(0, 0) and software raster or something

[ebon ruin]

What did gp do before virtual geometry lol

[fiery bolt]

actually release games

[wicked notch]

too real

[fiery bolt]

i reduced my mesh shader payload size and went from 60 to 27 ms

what the fuck

[wicked notch]

you should keep it to less than 104 bytes

as per nv's suggestions

[fiery bolt]

i was at... 768

[wicked notch]

incredible

[fiery bolt]

now it's 256

one index per meshlet

how do i make it smaller

[wicked notch]

like this https://github.com/LVSTRI/Retina/blob/1152267563683bb6a8a6731ce31b8bc07ffe88c4/src/Retina/Sandbox/Shaders/Visbuffer.task.glsl

[fiery bolt]

good idea

but my late culling pass has fragmented indices

i should get rid of the whole meshlet pointer shit tbh

[loud crag]

my apple gpu hung when i used subgroup ops instead of an atomic

[wicked notch]

gpu issue ig

[faint crane]

There are frogs in your computer eating up all the threads.

[frank sail]

frogs in my computer

[ebon ruin]

i wonder if apple had a good reason to use their own graphics api

it wouldnt surprise me if they did it just to make programmers slightly more suicidal

[buoyant summit]

apple compiler miscompiles subgroup ops

[loud crag]

in what way

more details pls

[buoyant summit]

cc @craggy shale

[buoyant summit]

in doing lots of invalid transforms to these ops like moving across control flow, and not using correct tangles (sets of threads that participate in a simd_ op)

[loud crag]

so the asahi compiler wouldnt have had the same hang issues?

how tf can they not compile correctly for their own hardware

also still not eure what you mean

too eeepy for this

[faint crane]

Frogs in your brain.

[buoyant summit]

they use llvm

[primal shadow]

I never tested to see if that was faster, but I'm also switching to persistent threads soon anyways.

[fiery bolt]

I don't understand what exactly they use to bound their BVH

probably need to put some thought into it

but first I need to fix my cursed task shader abuse

[fiery bolt]

https://github.com/Themaister/Granite/blob/master/assets/shaders/post/hiz.comp

[loud crag]

is that the reason for the uniform control flow issues

i remember a presentation from apple about that in llvm and how they managed that

[buoyant summit]

idk

you can make llvm actually work

NV has done that somehow

@craggy shale pls msl simd_ miscompilation examples

[loud crag]

the guy from apple i spoke to about this said he had no idea why my shader was causing it by looking at it briefly

just blamed it on mesh shaders

[craggy shale]

just what i had on mastodon

[buoyant summit]

https://tenor.com/view/honestly-quite-incredible-drinking-water-underwater-cool-gif-22285702

[primal shadow]

Wait, someone adapted SPD to do conservative downscaling??? This is litterly what I need

[loud crag]

havent we all done this now

min reduction sampler or with just min() when reading the 4 samples

[loud crag]

i cant figure out mastodon search so i can‘t find shit

[primal shadow]

In a single pass in a compute shader? You can do it with more than 1 pass of a fragment shader, but that's slooow

[loud crag]

ye lvstri does this and ive done this (semi-successfully)

with the nvidia spd

[buoyant summit]

https://mastodon.gamedev.place/@gob/109548806532038622

it's old tho so maybe they have fixed things since

gob also had a screenshot somewhere where they cse'd simd_add(1) across control flow

[primal shadow]

As-is, or did you need to modify the algorithm at all?

[wicked notch]

no changes required

it was very straightforward actually

[loud crag]

i was only using simd_prefix_exclusive_sum and simd_sum

oh wait i get it

if i was using 32 meshlets per invocation, i had no loop (optimised away) and then it worked fine

anything else and it hung, probably from using the simd stuff in the loop

might still be related tbf

though again im no expert on these compiler shenanigans so no idea what this “loop reconvergence” means

mayhaps i could inspect the generated IR and see whats going on

[fiery bolt]

yup

[fiery bolt]

https://themaister.net/blog/2024/01/

Building the mip-chain in one pass is great for performance, but causes some problems. With NPOT textures and single pass, there is no obvious way to create a functional HiZ, and the go-to shader for this, FidelityFX SPD, doesn’t support that use case.

The problem is that the size of mip-maps round down, so if we have a 7×7 texture, LOD 1 is 3×3 and LOD 2 is 1×1. In LOD2, we will be able to query a 4×4 depth region, but the edge pixels are forgotten.

The “obvious” workaround is to pad the texture to POT, but that is a horrible waste of VRAM. The solution I went with instead was to fold in the neighbors as the mips are reduced. This makes it so that the edge pixels in each LOD also remembers depth information for pixels which were truncated away due to NPOT rounding.

I rolled a custom HiZ shader similar to SPD with some extra subgroup shenanigans because why not (SubgroupShuffleXor with 4 and 8).

[primal shadow]

Huh, I've read that before, but must have forgot that part

[wicked notch]

this refactor is going very well

I say to myself as I wonder why I can't keep working on the same thing

[fiery bolt]

9.6 ms to lod and cull 210384384 meshlets

85% SM utilization

i doubt i can get eek out more perf out of this

which means it's BVH time

[wicked notch]

single ownership rules

[glass sphinx]

this is now the meshlet rendering channel

[wicked notch]

always has been

[fiery bolt]

@primal shadow https://github.com/bevyengine/bevy/blob/6183b56b5d6fd7e2e8cf1f3c85da7fd3dab25ea4/crates/bevy_pbr/src/meshlet/from_mesh.rs#L90
another thing i noticed, shouldn't the fold acc start with 0, not the parent error?

[fiery bolt]

nanite channel

wish.com nanite channel

[wicked notch]

our nanite is better

by virtue of being open source

and that we're all learning from each other

[fiery bolt]

my occlusion culling still decides to cull shit randomly

for no reason whatsoever

it's so random i think the bounding sphere generation is fucking up somewhere

[primal shadow]

Yeah good catch, I think it should :P. Thanks for pointing this out.

[fiery bolt]

your performance is now gonna half lol

[primal shadow]

Lol we'll see

[fiery bolt]

actually no it might improve

[primal shadow]

I have a ton of stuff to implement still, there's so much I can improve

[fiery bolt]

because it'll switch to the lower LOD earlier

[fiery bolt]

yeah same

if only this occlusion culling started to work properly...

[primal shadow]

Mine is broken, I need to fix it

[fiery bolt]

who's isn't

[primal shadow]

The fold might be better written as a map and then a max, hmmm

[fiery bolt]

i do map(..).reduce(f32::max).unwrap()

[faint crane]

Stochastic occlusion culling. Eventually it will converge into something which works.

[fiery bolt]

unfortunately it doesn't

maybe i should let it run for longer

[wicked notch]

I now have the the greatest timeline abstraction thanks to Dolkar

[glass sphinx]

dolkar is indeed the timeline god

[wispy spear]

[loud crag]

@wicked notch in your idea of sparse vsm, when would you free pages that are not visible? In the normal VSM we have those two buffers for (1) free pages, and (2) cached/invisible pages, and we allocate from free pages before re-using cached pages.
Now the problem I am having is this line in the Metal docs:

If the heap runs out of memory, Metal skips any remaining tiles in the request.
This means I need some way of figuring out when to free cached pages, so that I don't run out of memory... What were you imagining to solve this kind of problem?

[wispy spear]

count up the space you consume as you request pages, and if above a certain threshold you flush old pages?

[loud crag]

hmm sure i guess i could query the amount of sparse memory I am currently using, and then when I get close to OOM I could just evict a few unused but cached pages

[wicked notch]

"my sparse" is the same as everyone else's, so you just free all pages that are not visible

[loud crag]

what about caching though

[wicked notch]

we only cache in view pages

[loud crag]

oh bruh

ok ig that simplifies a lot

[wicked notch]

to keep out of view pages cached you can just implement an LRU cache

[loud crag]

what's that

[wicked notch]

and pop pages as needed before allocating new ones

[wicked notch]

a cache policy

one of many

least recently used (LRU)

[loud crag]

well that could be something for later then

[buoyant summit]

details?

[wicked notch]

it's actually the easiest solution but my dumb ass brain couldn't reason about it

so you just have one timeline semaphore per queue as usual

but the timeline value is shared between them and it's always monotonically increasing

to ensure that stuff gets deleted properly, you just collect every queue's last reached value (from the semaphore if needed) and take the min of that

then you use that as the value to compare when deleting stuff

also if a queue isn't used for a while, you just fast forward its last reached value to the current maximum value (otherwise the deletion queue would get stuck)

the source is here https://github.com/Dolkar/Tephra/blob/f488fae39bf97819bb7f3df7007fceaa1bb504c3/src/tephra/device/timeline_manager.cpp

I took heavy inspiration (to avoid saying straight up copied) from Dolkar's abstraction

[buoyant summit]

I see

cool

[fiery bolt]

i just realized, instead of taking LOD error as a sphere with center = centroid of the group, shouldn't it be a sphere with center = closest point of the group bounding sphere

[primal shadow]

Wdym closet point?

[fiery bolt]

closest point to the camera

so you'd have to store the group bounding sphere, project that, and then place your 'lod sphere' at center - (0, 0, radius)

[delicate rain]

I don't 🥸

LRU cache is cope, what I'll do is sort based on the cascade index, so that I free the cached pages in the lowest cascade first

Since those will have the shortest lifespan anyways

[wicked notch]

smart

[fiery bolt]

why is meshopt generating meshlets with an average of 42 tris

[frank sail]

meshopt knows the meaning of life

[wicked notch]

real

I'm almost back btw

I just need this one last feature I promise

just one last feature

[fiery bolt]

100 billion trogne