add a feature for memory tracing with tracy

[mockersf]

Objective

• Expose a feature for tracing with Tracy to profile memory (https://docs.rs/tracy-client/0.15.2/tracy_client/struct.ProfiledAllocator.html)
• This is a separate feature than just tracing as it can have an additional cost

[hymm]

does this fix #6927?

[james7132]

I tried this a while back. For whatever reason, it broke logging due to registering a logger/subscriber before logging initialization in bevy_log happens. Not sure if that's a problem here.

@hymm I think this only partially helps with monitoring memory usage and allocator usage (which we should be minimizing), but it doesn't let have detailed snapshots of memory to find what is using too much memory.

[mockersf]

Another view showing all allocation events happening during a span, here clear_entities. We can see the span it was allocated in, and the size

does this fix #6927?

At least a very good first step. @james7132 ?

[mockersf]

I tried this a while back. For whatever reason, it broke logging due to registering a logger/subscriber before logging initialization in bevy_log happens. Not sure if that's a problem here.

Maybe that has been fixed now, at least logging doesn't seem broken for me.

[mockersf]

and if you zoom on the graph, you can see each events, the span the allocation started, for how much, and the span it ended

[james7132]

At least a very good first step. @james7132 ?

Oh definitely, this is still great.. Though I would really like a Chrome-like support for labeled allocations to more readily dive into the exact split of memory usage at a given time.

Maybe that has been fixed now, at least logging doesn't seem broken for me.

Great. LGTM then.

[Elabajaba]

On Linux I think there are probably better tools (bytehound for profiling and heaptrack for the visualization is excellent in my experience), but this will probably be the best we can do for Windows (existing Windows tools either don't work with Rust, or they don't give any real insight into the application's allocations other than the total amount allocated).

If anyone has an Intel CPU then it might be worth comparing this to VTune.

As a side note, I don't think we can use Tracy's GPU profiling unless someone can take over and finish nagisa/rust_tracy_client#41, then integrate it with wgpu-profiler (and then it would be limited to backends that support wgpu::Features::TIMESTAMP_QUERY).

[Davidster]

out of curiosity, did you see a huge slowdown from enabling the memory tracing? On my project I have to keep it off because it's too slow

[mockersf]

out of curiosity, did you see a huge slowdown from enabling the memory tracing? On my project I have to keep it off because it's too slow

Yup!

The lighting example went from 500fps to 90

[hymm]

Not going to block on it, but we should have something for how to use the feature in profiling.md.