Viewer freezes on M2 Pro

[AsherJingkongChen]

Settings

• Apple M2 Pro
• Google Chrome 123.0.6312.87 (arm64)
• Metal 3

Descriptions

Freezes at or after unpacking some sample models. I don't know why and can't record this (My whole screen freezes). The memory usage looked normal. My cursor can move. However, my computer just shutdown unexpectedly.

Expectations

It should work on Google Chrome (arm64) and on Apple's Metal 3. Since it is based on WebGPU.

[AsherJingkongChen]

By the way, the other viewer works well on both Safari and Chrome (based on THREE.js). To test without having a new MacBook, GitHub's runner macos-14 perhaps helps.

[KeKsBoTer]

Thanks for reporting this issue!
Unfortunately, I think this is an issue related to webgpu/wgpu on M2 as my code should not be able to freeze the whole screen.

Did you try to build/run the native viewer?

[AsherJingkongChen]

Thanks for reporting this issue!
Unfortunately, I think this is an issue related to webgpu/wgpu on M2 as my code should not be able to freeze the whole screen.

Did you try to build/run the native viewer?

Build and run a native viewer executable

Build settings

Rust 1.77 (stable)

Build commands

cargo build --release --bin viewer

Run commands

./target/release/viewer ~/Downloads/point_cloud.ply cameras.json

Running

1. Launched slowly
2. The window appeared and froze my computer. I could move my cursor.
3. Froze, and everything stopped.
4. My computer crashed and restarted itself.
5. Sent a crash report to Apple.

Miscellaneous

I can run wgpu demo triangle or hello.
I think this issue relates to Apple's kext smbfs.

[AsherJingkongChen]

I will also figure out the solution.

[arcman7]

I will also figure out the solution.

Any luck with this?

[AsherJingkongChen]

Any luck with this?

@arcman7 I did not find the cause.

However, a similar problem can be dispatch size mismatch, improper subgroup size, or a long loop in GPU on the sorter module.

You can test the sorter (wgpu_sort). My screen got locked then.

I have written a sorting module in Rust for my trainer. I planned to write my own viewer in Rust in a few months.

[arcman7]

improper subgroup size

Can you elaborate on this? What makes you think there could be a mismatch occurring?

[AsherJingkongChen]

Can you elaborate on this? What makes you think there could be a mismatch occurring?

When manually adjusting the subgroup size in the wgpu_sort benchmark:

• Setting it to 32 caused the execution to block, while a size of 16 did not.
• The benchmark for a subgroup size of 16 reported 0ns, though the GPU continued running after the program exited.

The issue may be due to tasks taking too long, possibly caused by a loop not exiting as expected — something I’ve encountered before.

[arcman7]

That's interesting, I'll have to try tweaking those values. Thanks for the tip!

[qwoprocks]

This is probably too late, but I’ve made a patch to the radix sort algorithm that allows this to run on M-series GPUs at https://github.com/qwoprocks/web-splat-apple-silicon

Tested on my M1 Macbook Pro, achieves about 50 fps consistently without crashing