Meshlet continuous LOD by JMS55 · Pull Request #12755 · bevyengine/bevy

JMS55 · 2024-03-27T21:45:29Z

Adds a basic level of detail system to meshlets. An extremely brief summary is as follows:

In from_mesh.rs, once we've built the first level of clusters, we group clusters, simplify the new mega-clusters, and then split the simplified groups back into regular sized clusters. Repeat several times (ideally until you can't anymore). This forms a directed acyclic graph (DAG), where the children are the meshlets from the previous level, and the parents are the more simplified versions of their children. The leaf nodes are meshlets formed from the original mesh.
In cull_meshlets.wgsl, each cluster selects whether to render or not based on the LOD bounding sphere (different than the culling bounding sphere) of the current meshlet, the LOD bounding sphere of its parent (the meshlet group from simplification), and the simplification error relative to its children of both the current meshlet and its parent meshlet. This kind of breaks two pass occlusion culling, which will be fixed in a future PR by using an HZB from the previous frame to get the initial list of occluders.

Many, many improvements to be done in the future #11518, not least of which is code quality and speed. I don't even expect this to work on many types of input meshes. This is just a basic implementation/draft for collaboration.

Arguable how much we want to do in this PR, I'll leave that up to maintainers. I've erred on the side of "as basic as possible".

References:

Optimise extract_meshlet_meshes and prepare_meshlet_per_frame_resources

Co-authored-by: vero <email@atlasdostal.com>

…lods

pcwalton

I did a look over this, mostly for code style and so forth, not so much for the algorithm. In general this looks pretty straightforward. I definitely appreciate the clarity from the comments and splitting this up into a smaller patch!

Exciting stuff :)

crates/bevy_pbr/src/meshlet/asset.rs

pcwalton · 2024-04-10T17:18:59Z

crates/bevy_pbr/src/meshlet/cull_meshlets.wgsl

+fn lod_error_is_imperceptible(cp: vec3<f32>, r: f32) -> bool {
+    let d2 = dot(cp, cp);
+    let r2 = r * r;
+    let sphere_diameter_uv = view.projection[0][0] * r / sqrt(d2 - r2);


Suggested change

let sphere_diameter_uv = view.projection[0][0] * r / sqrt(d2 - r2);

let sphere_diameter_uv = view.projection[0][0] * r * inverseSqrt(d2 - r2);

The spec says inverseSqrt() does not work if e <= 0. Need to think if this is an issue.

i dont think sqrt is guaranteed to work for e <= 0 either. I'd just slab an abs or max 0 on it if its a worry, being explicit about edge case behavior intent is valuable and saturating/clamping is often free in gpus

crates/bevy_pbr/src/meshlet/from_mesh.rs

pcwalton · 2024-04-10T17:27:37Z

crates/bevy_pbr/src/meshlet/from_mesh.rs

+        unsafe {
+            meshopt_optimizeMeshlet(
+                &mut meshlets.vertices[meshlet.vertex_offset as usize],
+                &mut meshlets.triangles[meshlet.triangle_offset as usize],
+                meshlet.triangle_count as usize,
+                meshlet.vertex_count as usize,
+            );
+        }


Can we add a safe wrapper for this to the meshopt crate instead? It looks like the safe wrapper would be trivial, and that would let us avoid messing with the unsafe code settings in this crate.

Yeah it's not published yet though, so for now, unsafe it is.

pcwalton · 2024-04-10T17:35:22Z

crates/bevy_pbr/src/meshlet/asset.rs

-    /// The total amount of triangles summed across all meshlets in the mesh.
-    pub total_meshlet_triangles: u64,
+    /// The total amount of triangles summed across all LOD 0 meshlets in the mesh.
+    pub worst_case_meshlet_triangles: u64,


Maybe max_meshlet_triangles?

I personally feel that that's less clear. Up to maintainers.

i prefer keeping it as is

pcwalton · 2024-04-10T17:37:25Z

crates/bevy_pbr/src/meshlet/from_mesh.rs

+    let t = (lod_level - 1) as f32 / 19.0;
+    let target_error = 0.1 * t + 0.01 * (1.0 - t);


I personally like factoring tunable parameters like this out into consts at the top of the file so people can play with them if they need to. Up to you.

I used to have it in a previous commit. I can't decide why I removed it. Probably because it's just a heuristic that mostly works for the meshes I tested. Might need to be tunable by users in the future, idk. It'll need further testing when we try it out on a large array of meshes, but this PR was just to get something usable.

pcwalton · 2024-04-10T17:37:30Z

crates/bevy_pbr/src/meshlet/from_mesh.rs

+    );
+
+    // Check if we were able to simplify to at least 65% triangle count
+    if simplified_group_indices.len() as f32 / group_indices.len() as f32 > 0.65 {


pcwalton · 2024-04-10T17:40:03Z

crates/bevy_pbr/src/meshlet/from_mesh.rs

+    xadj.push(adjncy.len() as i32);
+
+    let mut group_per_meshlet = vec![0; simplification_queue.len()];
+    let partition_count = (simplification_queue.len().div_ceil(4)) as i32;


4 meshlets per group, ideally. 4 meshlets (256 triangles), simplify group to 50% (128 triangles), split back into meshlets = 2 meshlets (64 x 2 triangles).

I should probably rename this to target_partition_count, there are no hard limits.

Earlier code mentions that the meshlets are grouped into roughly 4 in a group. How come 4 is always used here if there aren't guaranteed to be 4?

It's the partition count. Given N nodes (meshlets), Metis divides them into P (partition_count) partitions (groups).

So given 1000 meshlets, we tell Metis to divide the 1000 meshlets into (1000/25)=250 groups. It's not guaranteed that we have exactly 4 per group though. Some groups might have slightly more or less than 4 meshlets, as it tries to spread the cost out as defined by the edge weights between nodes. We just have to hope for an even split between all partitions, which would give 4 meshlets per group if that were to happen.

crates/bevy_pbr/src/meshlet/from_mesh.rs

Co-authored-by: Patrick Walton <pcwalton@mimiga.net>

…lods

superdump · 2024-04-18T13:10:29Z

assets/models/bunny.meshlet_mesh

Do you think you will need to change this more? Storing binary files in git and having history of them isn't the nicest thing so if we could avoid it, that would be good. :)

Yep, it will definitely change a couple more times... Storing it externally also sucks though as then you can't run the example easily... Idk what we should do.

It will almost certainly change again, the format is not bit packed yet.

If I get Gltf->MeshletMesh asset preprocessing in, we can ship the repo with the GLTF file, and just have users convert it as part of the example. That way we won't have to worry about the MeshletMesh asset changing. But that's a good bit off, I don't want to block on that.

@bevyengine/maintainer-team any thoughts on this?

Just to note as it happened on Discord - it was decided that hosting the file out of tree would be better, and prompting the user to download it. The reason is that binary files that are modified and committed to the repository will have all versions in git history and the only way to remove them is to rewrite the git history.”, which isn’t nice. But also, rewriting git history breaks all open PRs and branches.

…lods

crates/bevy_pbr/src/meshlet/from_mesh.rs

superdump

Just a few nits/clarifications, otherwise this looks good to me.

…lods

superdump

I approve, but I still want to hear other maintainers' thoughts about the meshlet mesh file. Every change will go into the history and everyone not doing a shallow clone, which will be basically everyone, will download the whole history of all files. That isn't good.

# Objective - #12755 introduced the need to download a file to run an example - This means the example fails to run in CI without downloading that file ## Solution - Add a new metadata to examples "setup" that provides setup instructions - Replace the URL in the meshlet example to one that can actually be downloaded - example-showcase execute the setup before running an example

alice-i-cecile · 2024-06-03T20:36:15Z

Thank you to everyone involved with the authoring or reviewing of this PR! This work is relatively important and needs release notes! Head over to bevyengine/bevy-website#1303 if you'd like to help out.

ricky26 and others added 30 commits January 12, 2024 01:44

Parallelise meshlet buffer writes

20083fa

Skip some copies during meshlet upload

2283864

Merge pull request #17 from ricky26/meshlet

3f403bb

Optimise extract_meshlet_meshes and prepare_meshlet_per_frame_resources

Format

a9f1d9f

Clippy

3bfa80b

Add TODO

c16c3f2

Move variable out of loop

fe3dacb

WIP: Prepass/deferred support

3c9b361

Skip prepass draw prepare if only the depth prepass is used

92639e3

WIP: Prepass support

8a4b9aa

Implement deferred prepass node

0bd667a

Revert example changes comitted by accident

cda73dc

Add docs

58391d1

Bugfix when using multiple materials

1d4d89b

Improve example

ee2278c

Fix prepass draws

737d42d

Fix material depth being cleared

b04244a

WIP: Prepass/deferred support

ebc68e1

Fix imports

bb1490b

Finish prepass support

25d2943

Revert unintended change

5487522

TODO: Remove asset data from buffers on unload

dec0dda

Allow for gaps in PersistentGpuBuffer

60f2c91

Remove outdated doc

e1049d7

Reuse meshlet buffer data where possible

0f446fa

Update crates/bevy_pbr/src/meshlet/material_draw_prepare.rs

b48daeb

Co-authored-by: vero <email@atlasdostal.com>

Update crates/bevy_pbr/src/meshlet/gpu_scene.rs

109bb4f

Co-authored-by: vero <email@atlasdostal.com>

Update crates/bevy_pbr/src/meshlet/visibility_buffer_raster_node.rs

11b96e3

Co-authored-by: vero <email@atlasdostal.com>

Improve docs

1637b6a

Merge commit '4778fbeb65d840eb39bda017fd428ebfc17a1153' into meshlet

f9a8032

Merge commit 'ab7cbfa8fca5c17481dac5ab6b4921bd00910e89' into meshlet-…

e57d12c

…lods

JMS55 requested a review from superdump April 9, 2024 14:36

Merge branch 'main' into meshlet-lods

e456412

pcwalton reviewed Apr 10, 2024

View reviewed changes

JMS55 and others added 4 commits April 10, 2024 14:55

Update crates/bevy_pbr/src/meshlet/from_mesh.rs

7f2fd53

Co-authored-by: Patrick Walton <pcwalton@mimiga.net>

Fix error tense

cf7d3a9

Co-authored-by: Patrick Walton <pcwalton@mimiga.net>

Merge branch 'main' into meshlet-lods

48026f2

Merge commit '0256dacba47d697c7ae391a22f121c74de60c2b5' into meshlet-…

73fe48e

…lods

atlv24 approved these changes Apr 15, 2024

View reviewed changes

JMS55 added 2 commits April 16, 2024 16:38

Merge commit 'cae07ef56a932ce598d6aa1656f72734c9e7ac27' into meshlet-…

865c9d3

…lods

Small doc note

f3dde54

superdump reviewed Apr 18, 2024

View reviewed changes

Merge commit '831616662297b1586506a43fa3eefc75602bc31f' into meshlet-…

0e5ee9f

…lods

superdump reviewed Apr 20, 2024

View reviewed changes

crates/bevy_pbr/src/meshlet/from_mesh.rs Show resolved Hide resolved

superdump reviewed Apr 20, 2024

View reviewed changes

crates/bevy_pbr/src/meshlet/from_mesh.rs Show resolved Hide resolved

superdump reviewed Apr 20, 2024

View reviewed changes

JMS55 added the S-Ready-For-Final-Review This PR has been approved by the community. It's ready for a maintainer to consider merging it label Apr 20, 2024

Merge commit '13cac2eeff7e3c4da7ae636ba3766f59b4e343d0' into meshlet-…

fb0a669

…lods

superdump approved these changes Apr 21, 2024

View reviewed changes

Remove bunny asset in favor of downloading

81bda63

alice-i-cecile added this pull request to the merge queue Apr 23, 2024

github-merge-queue bot removed this pull request from the merge queue due to failed status checks Apr 23, 2024

alice-i-cecile added this pull request to the merge queue Apr 23, 2024

Merged via the queue into bevyengine:main with commit 6d6810c Apr 23, 2024

mockersf mentioned this pull request Apr 24, 2024

Example setup for tooling #13088

Merged

alice-i-cecile mentioned this pull request Jun 3, 2024

Write release notes for PR #12755: Meshlet continuous LOD bevyengine/bevy-website#1303

Closed

JMS55 mentioned this pull request Apr 5, 2025

Write release notes for occlusion culling bevyengine/bevy-website#2072

Merged

	let sphere_diameter_uv = view.projection[0][0] * r / sqrt(d2 - r2);
	let sphere_diameter_uv = view.projection[0][0] * r * inverseSqrt(d2 - r2);

		let t = (lod_level - 1) as f32 / 19.0;
		let target_error = 0.1 * t + 0.01 * (1.0 - t);

Uh oh!

Conversation

JMS55 commented Mar 27, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pcwalton left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

atlv24 Apr 18, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

superdump left a comment

Choose a reason for hiding this comment

Uh oh!

superdump left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

alice-i-cecile commented Jun 3, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

JMS55 commented Mar 27, 2024 •

edited

Loading

atlv24 Apr 18, 2024 •

edited

Loading