All WGPU demos randomly crash on Apple Silicon

[Umenokin]

Hi

Incredible repository which I found just yesterday.

Issue: The app will force Mac to freeze for a while completely blocking all IO until the system will force itself to reboot.

zig build bullet_physics_test_wgpu-run will cause this behavior 100% of the time.
zig build triangle_wgpu-run will cause pretty randomly but still 70-80% of the time.

I haven't been working with low-lever computer graphics for years and it reminds me how it will crash the graphics driver when you try to access the GPU with the wrong address.

[kamidev]

I am unable to reproduce this on my machine. What are you running?

➜  zig-gamedev git:(main) sw_vers; clang --version; zig version
ProductName:		macOS
ProductVersion:		13.6
BuildVersion:		22G120
Homebrew clang version 17.0.1
Target: arm64-apple-darwin22.6.0
Thread model: posix
InstalledDir: /opt/homebrew/opt/llvm/bin
0.12.0-dev.594+8fab4f98c

[hazeycode]

Sounds like it could be out-of-memory. Might be worth monitoring memory usage to see. Can you post your system specs also?

[Umenokin]

clang --version; zig version
Apple clang version 14.0.3 (clang-1403.0.22.14.1)
Target: arm64-apple-darwin22.5.0, M2 MacBook Pro 64gigs
Thread model: posix
zig version: 0.12.0-dev.467+0345d7866

Not sure how to provide more details since it completely freezes my system until it decides to reboot itself.

[michal-z]

@Umenokin Try release config: zig build -Doptimize=ReleaseFast

[foxnne]

I'm not fully convinced that this isn't related to that old Dawn bug on macOS. I thought it was fixed but I've since had very strange crashes similar to this, only on macOS. General purpose allocator reports no memory leaks. Really hard for me to predictably recreate.

Also, after several crashes, I checked storage and I had no memory left, which is interesting. Does macOS use the SSD when the system runs out of ram? It took a while for "System Data" to drop back down. I'm not sure if this is what causes the system to freeze/reboot.

[hazeycode]

IIRC macOS does use a swap file on disk when memory is exhausted.

The GPA will only pick up memory leaks from allocations that went through it. Dawn may well be allocating memory for CPU or GPU that wouldn't be tracked by the GPA.

[kamidev]

IIRC macOS does use a swap file on disk when memory is exhausted.

The GPA will only pick up memory leaks from allocations that went through it. Dawn may well be allocating memory for CPU or GPU that wouldn't be tracked by the GPA.

Yes, macOS has a swap file. And it may be used even when memory is NOT exhausted. Here is an article that talks a bit about that: https://www.digitaltrends.com/computing/what-is-swap-used-in-mac-activity-monitor/. You can see the current size of the swap file with: "sysctl vm.swapusage". Even on a Mac with ridiculous amounts of RAM, it will probably be there.

[kamidev]

Bullet_physics_test_wgpu-run now seems to freeze completely for me, too. I just started the app and checked memory with Activity Monitor, noticed a small but noticeable memory leak... but it also said 361 GB of private memory was used. I tried to take a screenshot but my Mac froze and had to be rebooted before I could.

[foxnne]

I'm fairly certain this has to do with Dawn as it affects both my mach-core and zig-gamedev projects. Similar to what you said I can sometimes observe a small memory leak but the swap file seems to eat all available storage.

[hazeycode]

Newer version of Dawn (#463) gives more validation error reporting. We see important validation errors in our samples. I'm optimistic that resolving these validation errors will solve these ill-effects that we have on Apple Silicon.