1 ms for culling
#Iris - A Journey through OpenGL and beyond to learn Graphics
fiery bolt
ok holy fuck it's rendering 1.2 million meshlets that doesn't sound good 
same scene in unreal takes 1.92 ms
fiery bolt
is this lodding also?
yeah
primal shadow
@wicked notch what heuristic did you use for software raster vs hardware raster again? I'm having a hard time figuring out a good metric. Nanite briefly mentions that they compute some kind of longest triangle edge data per cluster and use that iirc.
fiery bolt
project that to screenspace and software raster if it's less than 16 pixels or something iirc?
primal shadow
For now I've just given up on tweaking it, left it as something silly, and have moved on to other bits
Apparently Brian Karis read my blog post though, awesome!
wicked notch
if we never hear from you again, UE has won
primal shadow
He said he enjoyed it ๐
fiery bolt
unlike that tencent talk
runic surge
He said he enjoyed it ๐
Jasmine let us know if Epic Games sends their hit agents against you
faint crane
I saw both of Tencent's talks between Advances and Moving Mobile Graphics but couldn't understand most of it.
Wish I got a picture, but Jasmine's article was linked in Advances. Slides are up, let me link it.
fiery bolt
ok so I just read the tencent slides and... it's just nanite but worse?
it's minimum effort nanite reimpl
instance cull, parallelize all meshlets, no software raster, and some wacky lod curve
faint crane
Could the lod curve somehow compensate for lack of software raster? They were targeting mid-high end mobile.
wicked notch
Could the lod curve somehow compensate for lack of software raster? They were ta...
if you don't target pixel sized triangles then yeah, sw raster becomes unnecessary
buuut the benefits of sw raster are elsewhere too
for example with VSM, sw raster is pretty much the greatest source of speedup
delicate rain
lies, culling is
wicked notch
culling is the second greatest source of speedup 
nah saky's right it's culling first and then sw rast
but the advantage of not going through the regular pipeline for triangles is massive
delicate rain
I was this close to getting into another insane argument
wicked notch
especially for small ones
delicate rain
no more
wicked notch
wink wink bistro trees
delicate rain
bistro trees 
wicked notch
I genuinely believe that sw rast for bistro trees would fix everything
but alas
I still don't have nanite shippable
wicked notch
same
solar sentinel
<@735360596228440066>'s too : )
Hey deccer, just jumping in... what's going on? ๐
I'm doing hardware raster as well. Software raster doesn't pay off for anything larger than a pixel sized triangle. In the UE source code, it's literally called MicroPoly raster (...for a reason). There's overshading to compute hardware gradients. Even Nanite bins and rasterizes wider triangles via HW.
wispy spear
Hey deccer, just jumping in... what's going on? ๐
your name was part of references on some siggraph presentation, two or three message up from when i booped you
solar sentinel
Holy shit
solar sentinel
your name was part of references on some siggraph presentation, two or three mes...
I didn't even realize that. People are referencing my reddit posts. Wow.
solar sentinel
nanite does triangles smaller than 16 pixels or so with SW
Granted, their limits are a bit fuzzy... but beyond a certain size the hardware is faster.
primal shadow
I'm doing hardware raster as well. Software raster doesn't pay off for anything ...
What kind of HW raster? Mesh shaders or?
I still haven't figured out a good heuristics for SW vs HW. NSight shows like +/- 0.5ms all the time for the raster pass making measuring perf impossible. Idk how to improve when the results are so variable.
frank sail
Make the test scene bigger
solar sentinel
What kind of HW raster? Mesh shaders or?
Nah my target minspec is 1050Ti. Just regular ole'. Though I do frustum/HiZ culling compute and write out MultiDrawIndirectCount params into a buf.
solar sentinel
I still haven't figured out a good heuristics for SW vs HW. NSight shows like +/...
There some great high poly megascans of giant buildings on Turbosquid... probably have to dish out a few bucks, but worth it. Take those and just replicate over and over side by side.
primal shadow
Make the test scene bigger
I tried with like 10 copies of the quixel megascan icelandic cliffs. Maybe that's not enough??? idk
wicked notch
more.meme.gif
primal shadow
Nah my target minspec is 1050Ti. Just regular ole'. Though I do frustum/HiZ cull...
1 draw indirect args per cluster? How do you not get super bottlenecked on the command processor? I had tried this before, it was fairly slow
primal shadow
There some great high poly megascans of giant buildings on Turbosquid... probabl...
Oh nice, thanks for the tip
solar sentinel
1 draw indirect args per cluster? How do you not get super bottlenecked on the c...
Not clustered, complete instances.
But total number of issued calls is once per material.
primal shadow
Not clustered, complete instances.
But you have clusters(?)
solar sentinel
But you have clusters(?)
whole-ass instances of geometry. Not limited to ~128 triangles.
I'm not doing Nanite.
I'm doing visibility buffer rendering only ๐
primal shadow
Ohhh ok
fiery bolt
I tried with like 10 copies of the quixel megascan icelandic cliffs. Maybe that'...
how many tris in one of them
my stress test scene has like 100 billion tris total
spends about 10 ms in raster (hw, no sw)
well, i impld sw inside my mesh shader and that bumped it up to 20 ms lol
primal shadow
Not home rn, will check later
fiery bolt
@wicked notch I sped up mesh generation significantly by generating meshlet AABBs correctly instead of going through every single vertex in the mesh
wicked notch
crazy how that works
fiery bolt
it even speeds up rendering because now meshlets are being culled!
velvet marsh
is your AABB structure something you wrote yourself?
fiery bolt
yeah
i wrote some goofy bvh too
it... works?
the nanite presentation glossed over how it works so idk how correct i am
primal shadow
@fiery bolt I forget who is who, did you do the WebGPU Nanite impl?
(Scthe on github)
primal shadow
@fiery bolt is your code open source anywhere? I'd like to check out your meshlet DAG building/simplification code
fiery bolt
<@488643966502436865> is your code open source anywhere? I'd like to check out y...
i think i'm using meshopt for simplification currently, not my own thing
changed to meshopt to debug something then forgor to switch back
primal shadow
Looks... very similiar to my own code ๐
Did you base it off of mine? (which is totally fine, just curious what your process was)
fiery bolt
Did you base it off of mine? (which is totally fine, just curious what your proc...
i did yeah
primal shadow
Nice, happy it helped
fiery bolt
primal shadow
If you end up learning anything, let me know please!
So far today's experiments have revealed:
- Setting target error = 1.0 helps a lot. No reason to limit the target error.
- 255 v / 128 t is better than 64/64 (meshopt won't let me do 256 vertices :P)
fiery bolt
yeah i did the first one
and nanite does the latter
i didn't do the latter exactly because meshopt doesn't like 256 lol
i do 128/128
well, 128/124 because mesh shaders
primal shadow
I got better triangle fill rate with 255/128 vs 128/128
fiery bolt
what's your test model?
primal shadow
Stanford dragon, currently
fiery bolt
that is shade flat by default so you have zero vertex reuse
i had that issue with all the three stanford models
primal shadow
I have:
- Stanford dragon
- Stanford bunny
- Jinx from arcane form sketchfab merged down to 1 mesh (I don't really support multiple materials per meshlet mesh)
- Icelandic cliffs form quixel megascans
primal shadow
that is shade flat by default so you have zero vertex reuse
wdym?
fiery bolt
every tri has unique vertices
the bunny too
oh yeah i also use the max of error instead of adding the max child
fiery bolt
the bunny too
(and lucy)
primal shadow
oh yeah i also use the max of error instead of adding the max child
I was doing that originally, but then your error is not cummulative between LODs
So if LOD 0 has 0 error, LOD 1 has 10 error, and LOD 2 has 20 error
LOD 2's total error should be 30 relative to the bash mesh (LOD 0)
fiery bolt
not really
because it's a world place displacement
higher LODs will naturally have a higher error
primal shadow
Why? And that's tangential, no?
fiery bolt
lemme try to ms paint something
primal shadow
You don't want to know error relative to the previous LOD, you want to know error relative to the original mesh right?
fiery bolt
actually no idk how to draw it lmao
basically, when you simplify, you're collapsing edges into vertices
and the simplification error is the maximum displacement a vertex moved (sort of)
as you simplify more, edges get longer
so when you collapse, you get a higher error naturally
primal shadow
Sure, and that's a consistent metric for DAG cut purposes, yes
fiery bolt
i think this would lead to double-counting if you add
making your error higher than it should be
primal shadow
But that means your error projection is no longer saying "is this LOD imperceptible from the base mesh at this distance"
fiery bolt
the nanite presentation says they max, not add iirc
so in this case, if you're collapsing the red edge
you get this as the output
then you collapse this
leading to this
primal shadow
Compariosn of 255/128 vs 128/128 on icelandic cliffs btw: https://paste.rs/AFqkf.txt. Meshlet occupancy is a map of triangles_per_meshlet:count_of_meshlets key:value
fiery bolt
the total displacement in the two simplification steps is equal to half the length
which is equivalent to the second collapse error
primal shadow
can you rephrase that last bit?
I understand the images, but trying to understand the implications still
fiery bolt
so if you compare the first and last image, the vertex in the center has moved half the total length, right?
that's your final error of the simplification
fiery bolt
yeah, that's equivalent to the displacement done by the second collapse
primal shadow
Also I reread the nanite slides rq
They say they do the max of child's parent error for the BVH
TGhey don't mention how they handle choosing the parent error in the first place though
primal shadow
?
wicked notch
it's a leq comparison
so it doesn't matter
if parent >= threshold && current <= threshold
primal shadow
yeah at least as large implies max imo
So if that's true, then it's this?:
- LOD 0 clusters have error = 0.0, parent_error = INFINITY
- Group and simplify LOD 0 to form a group
- Group error = max(group_error_from_simplify, all_child_errors)
- LOD 0 parent errors = group_error
- LOD 1 error (not parent) = group_error
fiery bolt
yeah
fiery bolt
wicked notch
it is indeed add
fiery bolt
they add child errors to the parent error?
wicked notch
they add child errors to the parent error?
they add the bounds but yes same thing in UE's case
the bounds of the group directly affect the error calculation
which is shrimply copied
fiery bolt
me no understand
wicked notch
UE formatting
fiery bolt
primal shadow
Where the heck does LODError come from?
fiery bolt
i assume that's set to 0
wicked notch
blender -> subdivision modifier -> simple
be careful because you can turn those 2 million triangles into 2 quintillion
speaking from experience 
fiery bolt
also for runtime error calculation, you should store the group bounding sphere and place your error test sphere on that, instead of at the center of the group
@primal shadow ^
fiery bolt
blender -> subdivision modifier -> simple
but that's booooring
i'll just use lucy then
primal shadow
also for runtime error calculation, you should store the group bounding sphere a...
Wdym?
fiery bolt
currently you project a sphere with center = group center and radius = error, right?
primal shadow
Yeah
fiery bolt
that undercounts error for tris closer than the center
primal shadow
?
fiery bolt
you have triangles closer to the camera than the group center right
so if you test the error at the group center, those triangles might have an error higher than what you calculate
page 11
primal shadow
Wait so what's the solution?
Use the meshlet's center, and don't have any center for the group?
fiery bolt
store the entire bounding sphere of the group
and place the error test sphere at the closest point to the camera
fiery bolt
yeah
primal shadow
Calculate closest point on the sphere to the camera
Then place the sphere at center = closet point, radius = error
fiery bolt
yep!
the only issue is that idk what to do when the camera is inside the sphere
bounding or error sphere
because then your sqrt(d2 - r2) gets a negative value in it
primal shadow
Bleh that's going to be a lot more involved change, I'l add this to the TODO list
fiery bolt
because then your `sqrt(d2 - r2)` gets a negative value in it
if you clamp the value inside sqrt to 0, you get infinite projected error
which is sus
fiery bolt
i'm not sure
primal shadow
I mean yeah we want the LOD calculation to be as accurate as possible, but I feel like it's such a handwave to begin with...
fiery bolt
idk what the source of my overculling is
it's probably occlusion culling
but it might be this
primal shadow
My occlusion culling is just broken
Using SPD to build the depth pyramid is not correct unfortunately
fiery bolt
yeah
primal shadow
Say your depth texture is a non-power-of-2
What size do you make mip 0 of the depth pyramid?
primal shadow
So 1800: 1800/2 = 900, rounded down to 512?
primal shadow
So you don't enforce that the new depth pyramid is a power of 2? Hmm
fiery bolt
nope
primal shadow
I think that may be wrong, 1s
(depth_view.get_view_width() + 63u) & ~63u,
(depth_view.get_view_height() + 63u) & ~63u,
I think that rounds up to the nearest multiple of 64?
fiery bolt
yeah
primal shadow
Are you doing that? That may be you issue if not
fiery bolt
yeah i'm not
but why is he doing that
is it because each workgroup does a 64x64 block
and that he doesn't bound check in the shader 
primal shadow
Probably? Idk
Not sure if it's bounds checks though
Could be correctness
Like, extra bounds checks probably matters a lot less than extra vram usage from this
fiery bolt
i'll add that and check
primal shadow
Hopefully you can figure it out, I don't really want to go stare at hiz code ๐ญ
fiery bolt
i'll let you know
primal shadow
Wait before you start I have another question
When simplifying meshlet groups, which edges do you need to lock?
fiery bolt
ones that are only a part of one tri
primal shadow
eloberate?
primal shadow
Which border though?
fiery bolt
of the group
primal shadow
The border of the group, or border between meshlets?
fiery bolt
you don't want to lock between meshlets no
only the group
that's the secret sauce to nanite
oh yeah i also use larger groups than 4
i think it helps simplification perf
primal shadow
What do you use?
primal shadow
you might wanna test it out though
Assertion failed: index_count / 3 <= kMeshletMaxTriangles, file vendor/src/clusterizer.cpp, line 717
The groups are too big for meshopt to split ๐
primal shadow
Oh btw, are you using METIS or meshopt for clusterizing?
Interested to know if you compared the two at all
fiery bolt
The groups are too big for meshopt to split ๐
how are you splitting?
primal shadow
how are you splitting?
build_meshlets(indices, vertices, 255, 128, 0.0)
fiery bolt
in each group?
i do the same
it shouldn't be causing issues
meshopt clusterizes the source mesh after all
and it does lucy (28 mil tris) in 2-3 min
fiery bolt
oh i also made this shitty python script that automatically tiles any gltf
primal shadow
oh i also made this shitty python script that automatically tiles any gltf
Hah, you think I can load gltfs ๐
fiery bolt
Hah, you think I can load gltfs ๐
shouldn't bevy be doing that for you
primal shadow
Yes but not for meshlet meshes
Our asset processing APIs are sadly fairly poor atm, so I don't have a good way to convert GLTF -> bunch of meshlet meshes + scene file
(yet)
fiery bolt
ah
my issue is that my engine is not an engine
the only thing it can do is load gltfs
primal shadow
Join Bevy ๐ฆ
fiery bolt
does bevy do vulkan 
primal shadow
No, wgpu.
There is an alternative vulkan backend in a non-official crate, but it dosen't work with any existing stuff ofc.
faint crane
it's minimum effort nanite reimpl
Thinking this is the best description of my own effort TBH.
faint crane
Now I just need an excuse to skip on my externally broken occlusion culling.
I swear it was broken when I found it. No way to put it back together.
primal shadow
Based on recent discussion, bumped to 255v/128t, and some other changes https://github.com/bevyengine/bevy/pull/15023
loud crag
Based on recent discussion, bumped to 255v/128t, and some other changes https://...
that code looks so fragile because you arenโt using any constants for these values
primal shadow
that code looks so fragile because you arenโt using any constants for these valu...
Code quality comes later :p
loud crag
itโs something iโd put under โcode correctnessโ which would be the first thing to work on
primal shadow
There's too much else to work on in the mean time, small stuff like this is not a priority.
fiery bolt
did some optimization and now it culls 1 million stanford dragons in 2 ms 
raster takes 23 ms though (only hw, no sw)
800 billion tris at 30 fps
ebon ruin
But can it support foliage
fiery bolt
i have acheived scene independence
7.2 trillion triangles at 60 fps
@primal shadow the edge detection really does help simplification massively
thanks for the insight
fiery bolt
not really no
the other stanford dragon with 7.2 million tris imported in about 1.5 min
primal shadow
ยฏ_(ใ)_/ยฏ
Well, I'll take a closer look at your code soon
I need a break from DAG building ๐
fiery bolt
good question lemme check
primal shadow
Also do you have a link to your github? I lost it
fiery bolt
it is... whatever this is
primal shadow
it is... whatever this is
uhh
I mean it's kinda dragon shaped
sure
Managed not to turn into a sphere
fiery bolt
probably because meshopt doesn't generate vertices
primal shadow
You went back to your own simplifier?
Also hey, at some point I'd appreciate an explination on how to implement subpixel SW raster
The stuff I found online never made sense to me
And it seems like you implemented it
fiery bolt
And it seems like you implemented it
nah i haven't yet
but i will soon
i'll let you know when i do
fiery bolt
You went back to your own simplifier?
nah i'm using meshopt
primal shadow
probably because meshopt doesn't generate vertices
Oh nvm you're saying it didn't turn into a sphere because of this
fiery bolt
yeah
primal shadow
nah i haven't yet
Ah gotcha, nvm yeah looking at your sw shader I see it was something else. Thanks!
fiery bolt
my sw rasterizer is actually really bad
doing everything in hardware is very slightly faster lmao
primal shadow
Really? I saw way faster speeds with SW raster
Did you have mesh shaders already though?
fiery bolt
Did you have mesh shaders already though?
yeah it's all mesh shader based
and i do backface culling in the mesh shader
primal shadow
That explains it. Nanite was started before mesh shaders, and mesh shaders give a lot of similiar speedup.
fiery bolt
i think with a bit of optimization and async compute overlap i can eek out a fair bit of perf with sw tbh
mainly the async compute overlap
the mesh shaders are completely bottleneck on hw so they have terrible util
primal shadow
Yeah I don't have mesh shaders or async overlap ๐ญ
fiery bolt
impl mesh shaders into wgpu + naga and use them on native 
fiery bolt
naga isn't terrible
i rewrote/restructured it for module-level scoping long ago
wasn't too bad
no idea how the code is today though
wicked notch
7.2 trillion triangles at 60 fps
insane
ebon ruin
Question about Nanite/what you guys do
Meshlets are merged as the LOD gets lower to prevent edge cruft
Doesnโt this rely on the LOD of both meshlets decreasing? So what would happen if you need to lower one meshletโs LOD, but another meshlet must stay higher?
fiery bolt
LOD isn't decide at meshlet level, but at meshlet group level
and you only lock the boundary of the group
and after simplification, you split the group into meshlets again
primal shadow
and you only lock the boundary of the group
Lock the boundary of the group, *except for where it intersects the mesh border
fiery bolt
compute analytic derivatives instead of using ddx/ddy perhaps
@primal shadow oh yeah i also found out that my border vertex detection was completely wrong, fixed it and pushed to my branch (PRed to your bevy repo)
might wanna test it out again lol
primal shadow
compute analytic derivatives instead of using ddx/ddy perhaps
I am, my ddx/ddy aren't from the ddx/ddy() functions, I compute them as part of the visbuffer resolve
primal shadow
<@145540119141679105> oh yeah i also found out that my border vertex detection w...
Oh, thank you, I appreciate it. I am extremely short on time this week, but please do remind me if I don't get to it by saturday,
fiery bolt
will do if i remember
lmao i fixed edge classification which led to fixing cracks which led to significant occlusion culling improvements which now means i can do 7.2 trillion triangles in under 3 ms
4.3 +- 0.3 if i lock clocks to base
wispy spear
!remindme 3d remind jasmine about the thing
vivid boughBOT
fiery bolt
mobile drivers coming in clutch (words you'd never thought you'd ever see)
wispy spear
do you mean crutches?
fiery bolt
@languid vector methinks i have found a solution
the original sphere projection algo
that takes near clipping into account
project the full bounds to the screen, and scale the error using that
languid vector
thanks!
fiery bolt
idk if it works yet
languid vector
lmao i fixed edge classification which led to fixing cracks which led to signifi...
what was wrong?
let me guess - you classified edges without generating shadow index buffer with geometry-only data?
the only thing I wonder is how the hell you manage to do 7.2 trillion triangles in 4ms
Is hierarchy this major optimization?
fiery bolt
let me guess - you classified edges without generating shadow index buffer with ...
nah I was unlocking edges that were mesh borders but not group borders, but did it completely wrong
so it was unlocking group borders
leading to cracks
fiery bolt
Is hierarchy **this** major optimization?
yeah it's really useful
you can get rid of a bunch of instances early
and never even consider more than 1 or 2 meshlets for them
you also do frustum and occlusion culling inside hierarchy traversal
so you can save a bunch of bandwidth too
fiery bolt
you also do frustum and occlusion culling inside hierarchy traversal
atleast, I do because it's literally just a conventional bvh with error stuck on top
languid vector
I've only done the frustum culling so far
I can't really understand what am I bound to since not a single metric is nsight is loaded more than 80%
fiery bolt
memory
languid vector
though "warp can't launch" is insanely high
fiery bolt
or raster
fiery bolt
though "warp can't launch" is insanely high
which shader type
languid vector
which shader type
task/mesh pipeline
fiery bolt
what's ISBE alloc stalled at
languid vector
I am not sure I understand what ISBE is ๐
languid vector
fiery bolt
I think it's memory allocation for mesh shader outputs
if your rasterizer is cooked you'll stall at that because it's still processing other tris
languid vector
I also do nanite-style quantization so it should have saved a bit of bandwidth
I mean, it is nearly x4 compression
fiery bolt
I have zero compression lol
it's not about mesh shader read bandwidth, it's about how fast the rasterizer can chug triangles
you're raster bound
languid vector
I have never thought I will be raster bound 
and I guess the only way to solve it occlusion culling + sw raster?
afaik occlusion culling is a massive optimization
really massive
languid vector
how large is your test scene
25x25 grid of happy buddha meshes :>
fiery bolt
how many tris is that
you should use the xyzrgb dragon
it's got 7.2 million tris in it
or Lucy
languid vector
how many tris is that
1.1m
fiery bolt
27 mil iirc
languid vector
or Lucy <:froge_love:1105211408255295624>
yeah lucy is fat ass mesh
languid vector
1660 Ti
fiery bolt
yeah you need occlusion cull
languid vector
I guess so
fiery bolt
yeah lucy is fat ass mesh
it takes 17 min to build for me lmao
languid vector
it takes 17 min to build for me lmao
holy cow
languid vector
fiery bolt
still better than unreal
languid vector
unreal takes longer?
fiery bolt
I let it run for an hour once and it froze
so I killed it
if this shit works I can finally start streaming
languid vector
if this shit works I can finally start streaming <:froge_love:110521140825529562...
"this shit" is a new sphere projection method?
fiery bolt
yeah
languid vector
I was just about to ask you if you already have implementation
my dumb ass can't handle the maths now
fiery bolt
I'm making coffee so I can consume copious amounts of caffeine to understand wtf the paper does
languid vector
I personally can't understand the meaning of "conservativeness" in all these papers
fiery bolt
I personally can't understand the meaning of "conservativeness" in all these pap...
every choice you make must overestimate error, not underestimate
languid vector
every choice you make must overestimate error, not underestimate
Ahh, it makes sense now
fiery bolt
your bounding box must be larger than the object, never smaller
because otherwise you will cull too much
languid vector
yess, it makes sense now. thanks!
alr, it has 3 out variables:
out vec2 perpendicularDirection, out vec2 U, out vec2 L
now need to figure out what tf to do with it
fiery bolt
uhhhhhhhhhhh error is decreasing as i get closer
languid vector
uhhhhhhhhhhh error is decreasing as i get closer
so it aint working
fiery bolt
i've done something wrong
languid vector
so as I understand, it takes view and sphere and outputs a polygon with N sides with coordinates in screen-space? I guess I only need AABB from the sphere
fiery bolt
the core algorithm take a sphere in view space and tells you the min, max along a specific axis
languid vector
so I simply build an AABB and then compute its area that I use as error estimation?
"simply build" using this algorithm I mean
fiery bolt
i'm doing y axis so it's always pi / 2
languid vector
yup, makes sense
does it output size in screen-space?
or I need to project it further
languid vector
I see
I guess you are better at maths than me, so if you don't mind, can I come back with questions if have some after reading of the paper?
languid vector
thanks!
fiery bolt
same
languid vector
no, it indeed returns nans and I can't understand why
languid vector
@fiery bolt so I figured out the issue. I have inverse Z camera and was passing inverted clip planes to the algorithm ๐คก
tho haven't figured out how to use U and L for projection yet
languid vector
I mean, near plane was further than far plane
which produced a tons of NaNs due to sqrt of negative value
fiery bolt
ah lmao
languid vector
did u figure out how to use U and L?
fiery bolt
that's the screenspace bounds
languid vector
projected to the far plane?
fiery bolt
yeah you need to project
languid vector
how do I project projected stuff
languid vector
Alr I see, I will have a look at sample code
so yeah, it works relatively awful without reprojection
fiery bolt
i think i have something that sorta works?
languid vector
i think i have something that sorta works?
how did u fix the issue when error gets smaller if you get closer to the sphere?
languid vector
vec2 parent_U, parent_L;
GetBoundsForPhiLengyel(0.0f, parent_projected_bounding_sphere.center, parent_projected_bounding_sphere.radius, camera_data.near_clip_distance, camera_data.far_clip_distance, parent_U, parent_L) ;
vec4 parent_projected_points[2];
parent_projected_points[0] = camera_data.proj * vec4(parent_U.x, 0.0f, camera_data.near_clip_distance, 1.0f);
parent_projected_points[1] = camera_data.proj * vec4(parent_L.x, 0.0f, camera_data.near_clip_distance, 1.0f);
const float parent_result_error = (parent_projected_points[1].x / parent_projected_points[1].w) - (parent_projected_points[0].x / parent_projected_points[0].w);
this is how I do it
fiery bolt
public f32 project_error(f32x4 bounds, f32 error) {
let center = mul(this.mv, f32x4(bounds.xyz, 1.f)).xyz;
let radius = bounds.w * this.scale;
let err_frac = error / bounds.w;
if ((center.z + radius) <= this.near) return 0.f;
let dist2 = dot(center, center);
let a = sqrt(dist2 - center.z * center.z);
let t2 = dist2 - radius * radius;
let t = sqrt(max(t2, 0.f));
let in_sphere = t2 <= 0.f;
f32x2 bounds[2];
// cos(theta) = t / dist
// sin(theta) = r / dist
// T = (rotate(theta) * (a, z) / dist) * t,
// removing the dist divide in cos, sin
// ncos(theta) = t
// nsin(theta) = r
// rotate(theta) == rotate(ntheta) / dist
// therefore, T = (rotate(ntheta) * (a, z) / dist2) * t
// saving us two divides and a sqrt!
var v = in_sphere ? f32x2(0.f) : f32x2(t, radius);
let clip_sphere = (center.z + radius) >= this.near;
let off = this.near - center.z;
var k = sqrt(radius * radius - off * off);
[unroll]
for (int i = 0; i < 2; i++) {
if (!in_sphere)
bounds[i] = mul(f32x2x2(v.x, v.y, -v.y, v.x), f32x2(a, center.z)) * v.x / dist2;
let clip_bound = in_sphere || (bounds[i].y < this.near);
if (clip_sphere && clip_bound)
bounds[i] = f32x2(a + k, this.near);
v.y = -v.y;
k = -k;
}
let ndc_size = abs(bounds[0].x / bounds[0].y - bounds[1].x / bounds[1].y) * this.h;
// NDC size has a range of [0, 2] mapping to [0, height],
// but don't divide by 2 because the error is divided by 2 at build-time.
return ndc_size * err_frac * this.screen.y;
}
slang
languid vector
got it
languid vector
Alr, I will get back to the code tomorrow, gotta sleep. it is 3 AM in my country
fiery bolt
you still have 3 hours till bedtime!
wicked notch
make that 6
primal shadow
Time to debug tangents again
primal shadow
```java
public f32 project_error(f32x4 bounds, f32 error) {
let cent...
Btw after further thought, projecting the LOD sphere based on the culling sphere makes no sense to me
fiery bolt
why not?
primal shadow
you'd have different projections for different parts of the same group, which makes no sense
fiery bolt
how would you?
primal shadow
and when you do a BVH, it's based on the group, not individual meshlets
fiery bolt
you're using the group LOD sphere
primal shadow
wait what culling sphere do you use?
fiery bolt
a merged sphere of all lower lods' group lod sphere
to do BVH you need a bounding sphere or else traversal won't be monotonic
languid vector
you still have 3 hours till bedtime!
@fiery bolt so we have an answer for "how to project the sphere" question, but do we have an answer for "where tf to place the sphere" question?
I mean if camera is inside the sphere
Just snap to camera position?
dull oyster
I still don't understand why the sphere should not just be at the center of the group
languid vector
I still don't understand why the sphere should not just be at the center of the ...
So test is conservative
You should never underestimate error, you can only overestimate
here is 2 groups that have the same error. Obviously the left one will have perceptually bigger error because it is bigger by itself, so it will be closer to camera
fiery bolt
<@488643966502436865> so we have an answer for "how to project the sphere" quest...
I ended up doing something different that seems to work
but it's rendering way too much
like, 2 million meshlets
my culling queues are filling up
languid vector
like, 2 million meshlets
ideally u want 128 times less for 1080p monitor
dull oyster
So test is conservative
The test is made so that the error is not (in theory) perceptible, wouldn't making it more conservative just use more VRAM when you could use a lower fidelity lod for the same visual result?
fiery bolt
the test only guarantees the error isn't perceptible when you use bounds
dull oyster
I understand that it keeps higher fidelity lods more, but I'm not sure it's really necessary
fiery bolt
if you just use the center of the group it can't guarantee that
dull oyster
I may be missing the point completely but it bothers me that I don't see the issue ^^'
languid vector
I understand that it keeps higher fidelity lods more, but I'm not sure it's real...
it is not required but it is kinda more valid
fiery bolt
I may be missing the point completely but it bothers me that I don't see the iss...
- assuming error at the center is 0.999 px, the error for all triangles in the group closer than the center (which should be about half) will be more than a pixel
- when you do BVH, you need the outermost node's projected error to always bound that of it's children, so if you just use the center of the node, groups closer than that (again, about half), might have a higher projected error than the BVH node itself
languid vector
1. assuming error at the center is 0.999 px, the error for all triangles in the ...
languid vector
1. assuming error at the center is 0.999 px, the error for all triangles in the ...
btw, do u clamp the error sphere position to camera position if camera is inside the group sphere?
fiery bolt
i was sending my code but discord seems to be blocking it
languid vector
๐
fiery bolt
i assume it's being blocked for spam
// 2D Polyhedral Bounds of a Clipped, Perspective-Projected 3D Sphere (Michael Mara, Morgan McGuire).
// We get the projected bounds on the axis that is the longest upon projection (need to be conservative!),
// which is the one from (0, 0) to the sphere's center.
public f32 perceptible_error_distance(f32x4 bounds) {
let center = mul(this.mv, f32x4(bounds.xyz, 1.f)).xyz;
let radius = bounds.w * this.scale;
if (center.z + radius <= this.near)
return 0.f;
let dist2 = dot(center, center);
let a = sqrt(dist2 - center.z * center.z);
let proj_center = f32x2(a, center.z);
let t2 = dist2 - radius * radius;
var t = sqrt(max(t2, 0.f));
let in_sphere = t2 < 0.f;
// cos(theta) = t / dist
// sin(theta) = r / dist
// T = t * rotate(theta) * proj_center / dist,
// removing the dist divide in cos, sin
// ncos(theta) = t
// nsin(theta) = r
// rotate(theta) == rotate(ntheta) / dist
// therefore, T = t * rotate(ntheta) * proj_center / dist2
// saving us two divides and a sqrt!
let ncos = t;
let nsin = radius;
let wt_z = dot(f32x2(-nsin, ncos), proj_center) / dist2;
var t_z = t * wt_z;
if (in_sphere || t_z < this.near) {
// let off = this.near - center.z;
// let k = sqrt(radius * radius - off * off);
// let t = f32x2(a + k, this.near);
t_z = this.near;
}
return t_z;
}
public f32 error_perceptible_at(f32 error) {
// Don't divide by 2 because the error is already divided by 2 during build.
return this.screen.y * this.h * this.min_scale * error;
}
public bool should_visit_bvh(f32x4 lod_bounds, f32 parent_error) {
return this.perceptible_error_distance(lod_bounds) <= this.error_perceptible_at(parent_error);
}
public bool should_render(f32x4 lod_bounds, f32 error) {
return this.perceptible_error_distance(lod_bounds) > this.error_perceptible_at(error);
}
there
idk why i'm returning t_z
fiery bolt
yeah it does that
languid vector
we can always increase the threshold tho
fiery bolt
idk how correct it is
jagged valve
my cpu code is in rust tho
How do you handle your parameter binding business with a rust/slang setup?
jagged valve
bindless everything
ngl I had no idea that was even possible
fiery bolt
why is removing frustum culling reducing the number of meshlets drawn
but increasing the number of bvh nodes traversed (as it should)
languid vector
why is removing frustum culling reducing the number of meshlets drawn
task failed successfully
primal shadow
@fiery bolt ok I'm taking a look at your PR today
If i pass ptr::null() for vertex_locks, and add back SimplifyOptions::LockBorder, it should be equivilant to the old code right? For comparison purposes
So uhh it's a bit... aggressive
231 -> 4 meshlets is... a choice ๐
let me try this on the cliff instead of bunnies
Cliffs:
Something's a bit off
Anyways back to compression
fiery bolt
If i pass ptr::null() for vertex_locks, and add back SimplifyOptions::LockBorder...
probably? but not sure
are these numbers with the edge detection or without?
primal shadow
are these numbers with the edge detection or without?
Wdym? Expand the results, I showed on main/your PR
fiery bolt
Wdym? Expand the results, I showed on main/your PR
like, is this after replacing vertex locks with a nullptr or without
fiery bolt
hmmmmm
fiery bolt
231 -> 4 meshlets is... a choice ๐
i'm assuming this is the simplification queue length?
so it doesn't mean that the whole mesh was 231 meshlets and it's now 4
what's the threshold at which you reject a simplified group?
fiery bolt
my frustum culling has been wrong all along...
lmfao
@languid vector 7.2 trillion tris in 1.46 +- 0.29 ms
languid vector
<@323156168614871040> 7.2 trillion tris in 1.46 +- 0.29 ms
holy crap
I am rendering 500m tris at 30ms
:c
fiery bolt
occlusion culling
languid vector
but without occlusion culling
yeah
and without hierarchy
and without sw raster
fiery bolt
occlusion cull is a bigger win than hierarchy
i'm gonna tune my sw raster thresholds rn
it's not that much of a difference
around 20% boost
languid vector
what was the best boost for your virtual geom renderer?
I mean runtime performance
fiery bolt
wdym
primal shadow
i'm assuming this is the simplification queue length?
Yeah, amount of meshlets at each level
primal shadow
so it doesn't mean that the whole mesh was 231 meshlets and it's now 4
It does though
primal shadow
what's the threshold at which you reject a simplified group?
I allow simplifying at least 5% (i.e. 95% the same). Maybe I should change that and the test your changes again
primal shadow
it's not that much of a difference
(if you have mesh shaders :P)
fiery bolt
It does though
no, because some groups would've been rejected and will not be simplified ever again
primal shadow
also true
fiery bolt
I allow simplifying at least 5% (i.e. 95% the same). Maybe I should change that ...
yeah
i do 60% iirc
primal shadow
Tried 65%
LOD: 0, meshlet count: 15616, meshlet occupancy counts: {128: 15615, 88: 1, }
LOD: 1, meshlet count: 8066, meshlet occupancy counts: {128: 6406, 127: 1114, 64: 352, 63: 156, 126: 29, 62: 7, 43: 1, 125: 1, }
LOD: 2, meshlet count: 4719, meshlet occupancy counts: {128: 2925, 64: 507, 63: 465, 127: 222, 32: 157, 31: 133, 62: 98, 96: 62, 95: 59, 126: 27, 30: 20, 94: 17, 61: 13, 29: 5, 125: 3, 93: 2, 46: 1, 28: 1, 66: 1, 85: 1, }
LOD: 3, meshlet count: 1358, meshlet occupancy counts: {128: 811, 32: 11, 111: 9, 64: 9, 16: 9, 48: 9, 11: 9, 113: 8, 109: 8, 9: 8, 118: 8, 10: 8, 79: 8, 95: 8, 23: 7, 47: 7, 120: 7, 122: 7, 29: 7, 63: 7, 73: 7, 94: 7, 6: 7, 127: 7, 93: 7, 106: 7, 99: 6, 34: 6, 45: 6, 117: 6, 70: 6, 12: 6, 75: 6, 97: 6, 33: 6, 55: 6, 5: 5, 98: 5, 19: 5, 53: 5, 7: 5, 112: 5, 74: 5, 100: 5, 14: 5, 22: 5, 65: 5, 101: 5, 110: 5, 25: 5, 28: 5, 49: 5, 27: 5, 68: 5, 61: 4, 51: 4, 56: 4, 119: 4, 126: 4, 44: 4, 13: 4, 24: 4, 60: 4, 102: 4, 71: 4, 15: 4, 123: 4, 18: 4, 76: 4, 39: 4, 50: 4, 124: 4, 26: 4, 115: 4, 116: 4, 17: 4, 30: 4, 91: 4, 31: 4, 80: 4, 2: 4, 96: 4, 37: 4, 52: 3, 46: 3, 43: 3, 62: 3, 121: 3, 125: 3, 4: 3, 90: 3, 89: 3, 72: 3, 20: 3, 85: 3, 84: 3, 114: 3, 35: 3, 40: 2, 83: 2, 8: 2, 1: 2, 41: 2, 3: 2, 103: 2, 21: 2, 108: 2, 81: 2, 82: 2, 57: 2, 42: 2, 36: 2, 78: 2, 88: 2, 38: 2, 104: 1, 58: 1, 77: 1, 107: 1, 69: 1, 86: 1, }
is that the cliff?
huh i get
561, 284, 152, 80, 40, 19, 10, 6, 3, 2
meshlets per lod
for the bunny
primal shadow
is that the cliff?
Yes
primal shadow
Idk. I'm doen with trying to improve DAG building atm.
Next todos are compress meshlet data and then BVH-based persistent culling
fiery bolt
Next todos are compress meshlet data and then BVH-based persistent culling
you can't do persistent queues with wgpu atm
wgsl doesn't support coherency
primal shadow
Hmm, do you need it?
fiery bolt
yeah coherency is required for any non-atomic writes to be visible to other workgroups in the same dispatch
same reason it's needed for SPD
primal shadow
Which part needs those?
fiery bolt
Which part needs those?
the whole persistent queue shtick?
primal shadow
Right I know what it does, I'm curious what needs it though for persistent queues
Oh wait
fiery bolt
you're writing to a queue lol
fiery bolt
yep
primal shadow
yeah I see
primal shadow
Isin't it litterly just adding coherent (glsl) / globallycoherent (hlsl) to the buffer decleration though?
fiery bolt
and so does nanite on PC
fiery bolt
Isin't it litterly just adding coherent (glsl) / globallycoherent (hlsl) to the ...
yep
primal shadow
Ok I'll patch naga, easy enough
fiery bolt
oh yeah you also need forward progress guarantees
primal shadow
That, or spirv passthrough in bevy
fiery bolt
yeah that works
primal shadow
oh yeah you also need forward progress guarantees
Yeah ik avbout that part
let me find the spirv thing
fiery bolt
metal on M series fails that
primal shadow
https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#Decoration coherent decoration
fiery bolt
so you have to fall back to dependent dispatches for apple silicon
loud crag
metal on M series fails that
fails whag
primal shadow
metal on M series fails that
Naga is already broken on metal when it comes to atomic<u64? anyways so idc
fiery bolt
fails whag
you don't get forward progress on metal
loud crag
whatโs that
fiery bolt
it'll just keep spinning
fiery bolt
whatโs that
guarantees that workgroups will get switched out eventually even if they're not waiting for a mem access
loud crag
Naga is already broken on metal when it comes to atomic<u64? anyways so idc
pretty sure 64-bit atomics exist
fiery bolt
no API actually gives that to you
primal shadow
pretty sure 64-bit atomics exist
They do, but naga's MSL backend is bugged on them
loud crag
guarantees that workgroups will get switched out eventually even if they're not ...
why the hell would you want that
fiery bolt
but it mostly works on nvidia and amd
fiery bolt
why the hell would you want that
spinlocks
loud crag
They do, but naga's MSL backend is bugged on them
skill issue use a better api
fiery bolt
no API actually gives that to you
so nanite doesn't use them on PC
dependent dispatches work well enoughโข๏ธ
I might do persistent queues after streaming
maybe
wicked notch
I'm pretty sure cuda does define forward progress guarantees
so that's why it mostly works on NV
loud crag
https://arxiv.org/pdf/2109.06132
ahhh i understand
fiery bolt
intel who?
those CPUs that crash a lot
should I do the whole fixed size page thing or half-ass dynamic allocation
wicked notch
fixed page all the way
why would you even want dynamic alloc
just do tlsf on gpu
it's just a couple of fls and ffs
you might need a lock tho
acutally yeah no binning is bad
do standard gpu malloc
fiery bolt
why would you even want dynamic alloc
because I don't have to do the goofy group part nonsense
primal shadow
@languid vector I have some more questions on the global mesh compression when you get a chance
- I don't get the idea behind the function you sent me a bit ago to calculate the bitrate/step_size for the global/mesh quantization grid. Should this not be a fixed value used for all of your meshes? Also, what's even the point of the global quantinization since you store your meshlet centers in a full vec3<f32> anyways no?
- Given that each meshlet stores a bitstream of vertex positions, for meshlet X with triangle index Y, how do you read the vertex position data? With fixed-size positions, i.e. one vec4<f32> per vertex, I can just store one u32 per meshlet pointing to the start of the meshlet's vertices in the large array of vertices, and each triangle index can be a single u8 pointing to an offset off of that starting position. But I'm not sure how to structure things with a bitstream.
- When quantizing, I'm not entirely sure how to handle sign-ness (i.e. negative or positive). For meshlets, I guess I could map -radius..radius to 0..diameter, and then do ceil2(log2(diameter)) to determine the bitrate. For meshes/global grid, I guess I just store the meshlet centers as a full vec3<f32> still? (but quantized to the grid instead of absolute coordinates)
wicked notch
because I don't have to do the goofy group part nonsense
but goofy groupy thingies are cool
fiery bolt
but goofy groupy thingies are cool
how do you check if all parts of a group are loaded
without Yet More Indirection
wicked notch
literally who cares
you're already bw limited
one more buffer
and maybe another one after that
fiery bolt
no but then I have to deal with those buffers
fiery bolt
dynamic gpu malloc
wicked notch
me when I commit theft
ebon ruin
and maybe another one after that
The Khronos group when creating the Vulkan specification thought about you specifically when creating descriptor sets
fiery bolt
oh yeah it's also more work when building
wicked notch
fiery bolt
The Khronos group when creating the Vulkan specification thought about you speci...
what the fuck is a descriptor set
wicked notch
The Khronos group when creating the Vulkan specification thought about you speci...
what's a descriptor set
ebon ruin
what the fuck is a descriptor set
watch your language
wicked notch
never heard of er
fiery bolt
me neither
wicked notch
do you mean pointers maybe
ebon ruin
I thought you were a vulkan man
wicked notch
I am yes
fiery bolt
I just put a Tex2D<f32> in my push constants
wicked notch
I am a vulkan 1.3 man
ebon ruin
are you pulling my leg
fiery bolt
and it just works
wicked notch
are you pulling my leg
yes we all use bindless and pointers here
wicked notch
descriptor sets are a bad dream that don't exist anymore
fiery bolt
we have vulkaned someone else
wicked notch
we have vulkaned someone else <:froge_love:1105211408255295624>
an honest day's work
we need more fresh blood for john khronos ๐
ebon ruin
not yet
ebon ruin
i am a GL guy
wicked notch
not for long
fiery bolt
pointers
wicked notch
pointers
fiery bolt
lol
ebon ruin
useless
fiery bolt
no
ebon ruin
meaningless
fiery bolt
nanite makes you a real clown
ebon ruin
its all meaningless
wicked notch
nanite makes you a real clown
true and real
fiery bolt
its all meaningless
ok time for you to impl nanite
wicked notch
watch out for edge cases
ebon ruin
the only thing desirable is hw rt but everytime i mention it L says โNo.โ
ebon ruin
ok time for you to impl nanite
I did on Scratch
fiery bolt
ok now make it fast
ebon ruin
no
fiery bolt
28 trillion trongles in 1.5ms
ebon ruin
iโm not scared of vulkan its just useless to me
fiery bolt
hit that and we'll let you not use vulkan
ebon ruin
pointing to what?
wicked notch
memory
fiery bolt
if you can't do 25 trillion triangles in 1.5ms you have to use vulkan
ebon ruin
neither have you
wicked notch
quadric error metrics maybe perhaps
fiery bolt
my error projection makes no sense
wicked notch
my error projection makes no sense
real
fiery bolt
I just did the screen space projection thing and then multiplied two things
fiery bolt
because just using t led to holes
wicked notch
best error function ever
ebon ruin
this is like the dark wizards trying to convince me to use dark magic
wicked notch
show him
wicked notch
the slang push const
fiery bolt
wicked notch
give him a taste of slang's ultimate power when combined with vk13
wicked notch
(don't tell him about the driver bugs and invalid spirv and out of spec optimizations)
fiery bolt
fuck nvidia's driver
wicked notch
all my homies hate nv
ebon ruin
me too
ebon ruin
i shouldve gone with rx
fiery bolt
fucking useless thing crashes with a misaligned addr if i use groupshared mem in my software rasterizer
wicked notch
I have zero idea how they say slang is "production ready"
fiery bolt
i've looked through the entire spirv slang shits out
wicked notch
"release" "driver"
fiery bolt
it's perfectly fine
fiery bolt
fucking useless thing crashes with a misaligned addr if i use groupshared mem in...
so i wrote it in hlsl
works now
now i have three definitions of my gpu scene structs
wicked notch
bruh
fiery bolt
this was in a minimized diff that just wrote zeroes to the arr
still crashed with slang so i assume that's the issue
fiery bolt
and using groupshared arrays in my mesh shader also works
wicked notch
the whole thing is ub
fiery bolt
and using groupshared arrays in my mesh shader also works
but atomics don't!
slang default InterlockedAdd uses device scope for everything so i wrote my own spirv asm thingy for VMM (and all barriers etc)
it still died
do you think novideo will hire me if i tell them i'll fix their shit shader compiler
oh yeah VMM make available | make visible doesn't work either for some reason so i had to split my dispatches
fiery bolt
actually seek grass
what is that
faint crane
Haven't implemented that yet.
primal shadow
<@323156168614871040> I have some more questions on the global mesh compression ...
I guess to answer my own question for 2., each meshlet can store the first bit of it's vertex positions within the bitstream, as well as it's quantization factor, which can then be used to calculate how many bits each vertex position uses for the meshlet which is fixed per-meshlet, giving you random access within the meshlet.
buoyant summit
@wicked notch in cuda and the likes you use a special dispatch for forward progress
and the api requires that your dispatch size is under some limit
for forward progress guarantees to hold
and the limit varies device to device
and yes that would be useful in vk no less, I'm just saying it's a bit tricky to use
it's not super straightforward
wicked notch
ye it's flimsy
primal shadow
<@145540119141679105> did you do the thing?
Yeah yesterday
primal shadow
pub struct MeshletMesh {
/// Bitstream-packed vertex positions.
pub vertex_positions: Arc<[u8]>,
/// Octahedral compressed normals and uncompressed texture coordinates for vertices.
pub vertex_attributes: Arc<[u8]>,
/// Triangle indices for meshlets.
pub indices: Arc<[u8]>,
/// The list of meshlets making up this mesh.
pub meshlets: Arc<[Meshlet]>,
/// Spherical bounding volumes.
pub bounding_spheres: Arc<[MeshletBoundingSpheres]>,
}
/// A single meshlet within a [`MeshletMesh`].
#[repr(C)]
pub struct Meshlet {
/// The bit offset within the parent mesh's [`MeshletMesh::vertex_positions`] buffer where the vertex positions for this meshlet begin.
pub start_vertex_position_bit: u32,
/// The offset within the parent mesh's [`MeshletMesh::vertex_attributes`] buffer where the vertex attributes for this meshlet begin.
pub start_vertex_attribute_id: u32,
/// The offset within the parent mesh's [`MeshletMesh::indices`] buffer where the indices for this meshlet begin.
pub start_index_id: u32,
/// The amount of vertices in this meshlet.
pub vertex_count: u8,
/// The amount of triangles in this meshlet.
pub triangle_count: u8,
/// Number of bits used to quantize vertex positions within this meshlet.
pub bits_per_vertex_position: u8,
/// Unused. (TODO: Get rid of this in the disk representation?)
pub padding: u8,
}
Ok, got this so far.
wicked notch
why is everything Arc 
frank sail
I thought it was atomic ref counted
Idk what the ref countedness means in the context of a single uint
wicked notch
it's just shared_ptr
primal shadow
why is everything Arc <:nervous:836935166437883934>
Reasons internal to how bevy's renderer works to avoid copying the data across threads
primal shadow
can't you just make MeshletMesh arc?
Yeah, but I also need some sort of smart pointer inside anyways to store unbounded arrays [u8]
It was either Arc (shared_ptr) or Box (unique_ptr)
fiery bolt
or you can do a custom DST
primal shadow
?
fiery bolt
you can make your own DST with a header and an unsized tail of bytes
primal shadow
Ehhh maybe some other time if it becomes an issue lol
primal shadow
Compression continues to frustrate me greatly
I still have not seen good explinations for how half of it works
Mostly the purpose of the global grid
wide shadow
Mostly the purpose of the global grid
The global grid is there to avoid cracks in the model between different meshes. If every mesh would have its own grid then there would be cracks because each mesh aligned differently
faint crane
Twitter
fiery bolt
The global grid is there to avoid cracks in the model between different meshes. ...
yeah this is required for kitbashing or any sort of modular design
primal shadow
The global grid is there to avoid cracks in the model between different meshes. ...
I don't understand why
Nor do I think that's right
I think nanite was saying you need the same grid to avoid cracks between objects you would get from different grids
But unclear why they have one in the first place
glass sphinx
I think nanite was saying you need the same grid to avoid cracks between objects...
but isnt that what lukasino said?
primal shadow
Right, but it's missing some context
Yes if you do different grid sizes per mesh there would be issues
But what's the purpose of the grid in the first place?
Like you could also solve the same problem by... Not quantizing anything to a grid
glass sphinx
how would you quantize it
fiery bolt
without quantizing you need the full 12 bytes for position
and it won't byte-compress very well I think
primal shadow
But you can compress with the per-cluster encoding, which afaik(?) is lossless(?)
So it'll compress to a more compact bitstream regardless
I think the global quantizing step beforehand is just to reduce the precision, in order to need less bits for the second step? Idk for sure
And then it makes a bunch of other things more complicated
There's two steps. The quantization with a fixed step size for all meshes, and then encoding vertices per-cluster
I think the purpose of the first is to reduce excess precision (so less bits are needed), and the second is just a more compact encoding of the same data
glass sphinx
if your compression doesnt yield the exact same vertex positions ror border vertices you get cracks
i guess the world grid helps keep the qualtization consistent
i dont think the quantization is lossless
wispy spear
looks like caldera hotel
wicked notch
it isnt going to be the easiest thing to do
I'm so cooked
why do I recognize that this is bistro
wispy spear
oh hehe
wheat haven
why do I recognize that this is bistro
I did too, don't feel bad ๐
primal shadow
i guess the world grid helps keep the qualtization consistent
Why though? How does that work? I haven't figured out why it helps
fiery bolt
Why though? How does that work? I haven't figured out why it helps
because within a cluster you may just happen to have vertices at 0, 2, 4, 6, 8.0000000901, 10
wispy spear
@wicked notch i might have posted it somewhere before, im not sure, but it mentions mesh decimation/remeshing etc not sure if its useful or was useful as an alternative or better working meshlet thingy https://github.com/pmp-library/pmp-library but ill link it anyway
wicked notch
very nice, I'll check it out in a bit
primal shadow
Interesting issue I've found, shadows look too bad atm in bevy for meshlets
It's choosing too low of a lod I'm guessing, and the error is very visible when viewed by the main camera
I probably need to add a lod bias to shadow views
primal shadow
Ironically nanite does the same, but they bias towards a less accurate lod, as VSM is so high res anyways it's never a problem
Mhmm also true
wicked notch
VSM is too powerful
primal shadow
Rasterization is a hack and does not work for non-primary views
And I won't pretend otherwise
frank sail
Banned
wicked notch
banned
finite yacht
Promoted to Admin
frank sail
not you too
loud crag
Rasterization is a hack and does not work for non-primary views
writing a raster renderer just feels like strapping hacks upon hacks together
fiery bolt
don't waste your 0.25 rays per pixel on shadows, use them for GI and real specular smh
primal shadow
Ok after much research, I finally understand nanite's vertex quantization now
primal shadow
/// A single meshlet within a [`MeshletMesh`].
#[derive(Copy, Clone, Pod, Zeroable)]
#[repr(C)]
pub struct Meshlet {
/// The bit offset within the parent mesh's [`MeshletMesh::vertex_positions`] buffer where the vertex positions for this meshlet begin.
pub start_vertex_position_bit: u32,
/// The offset within the parent mesh's [`MeshletMesh::vertex_normals`] and [`MeshletMesh::vertex_uvs`] buffers
/// where non-position vertex attributes for this meshlet begin.
pub start_vertex_attribute_id: u32,
/// The offset within the parent mesh's [`MeshletMesh::indices`] buffer where the indices for this meshlet begin.
pub start_index_id: u32,
/// The amount of vertices in this meshlet.
pub vertex_count: u8,
/// The amount of triangles in this meshlet.
pub triangle_count: u8,
/// Number of bits used to quantize vertex positions within this meshlet.
pub quantization_bits: u8,
/// Number of bits used to to store the X channel of vertex positions within this meshlet.
pub bits_per_vertex_position_channel_x: u8,
/// Number of bits used to to store the Y channel of vertex positions within this meshlet.
pub bits_per_vertex_position_channel_y: u8,
/// Number of bits used to to store the Z channel of vertex positions within this meshlet.
pub bits_per_vertex_position_channel_z: u8,
/// Unused. (TODO: Get rid of this in the disk representation?)
pub padding: u16,
/// Minimum quantized X channel value of vertex positions within this meshlet.
pub min_vertex_position_channel_x: f32,
/// Minimum quantized Y channel value of vertex positions within this meshlet.
pub min_vertex_position_channel_y: f32,
/// Minimum quantized Z channel value of vertex positions within this meshlet.
pub min_vertex_position_channel_z: f32,
}
The perfect 256 bits of metadata
wide shadow
```rust
/// A single meshlet within a [`MeshletMesh`].
#[derive(Copy, Clone, Pod...
Do you have duplicated vertex positions for each meshlet? Meaning each vertex had its own "vertices"
fiery bolt
good
primal shadow
Do you have duplicated vertex positions for each meshlet? Meaning each vertex ha...
I assume you meant each meshlet has it's own set of vertices, and yes I do. I was skeptical at first, but it allows streaming and much better compression, and vertex data memory usage is a large bottleneck.
wide shadow
I assume you meant each meshlet has it's own set of vertices, and yes I do. I wa...
Yes, thats what I meant. I am just checking ๐
primal shadow
Update: It's been taking a while due to learning and then being sick, but my fever finally broke and I finished all the CPU-side changes for compressed per-meshlet vertices.
Just need to figure out how to do the GPU bitstream reader
I am also using a fixed quantization factor per mesh rn, nanite has an "auto" mode that chooses the best one, I'll have to figure out how they did that.
fiery bolt
don't all meshes need to have the same quantization factor
because it would lead to cracks otherwise
primal shadow
Maybe, idk. The original nanite presentation said it's user-selectable and has to be the same for different meshes, but unreal has an "auto" option.
Choose the precision this mesh should use when generating the Nanite mesh. Auto determines the appropriate precision based on the size of the mesh. The precision can be overridden to improve precision or optimize disk footprint.