Make scheduling benchmark more realistic

[joseph-gio]

Objective

• Current schedule bench is quite unrealistic.
• Redesign it based on the architecture of a typical bevy app.

Solution

Model the benchmarks around plugins, forming groups of systems.

• Systems within a group have orderings specified based on their TypeId.
• Inter-system dependencies are based on derived labels.
• Plugins have their own labels, which is passed down to all systems they contain. This allows plugin-wide dependencies sometimes.

[alice-i-cecile]

Love the constants, love the strategy for defining the graphs, love the increased realism. Eventually I'd be interested in a follow-up PR that adds a method to compute these key statistics on Schedules so we can ask users for usage data.

I think we can make this nicer with code generation macros, and strongly feel that this needs to be explicitly seeded.

Should be great!

[alice-i-cecile]

I would like the number of plugins and systems to be pulled out into constants here.

[joseph-gio]

The problem with this is that, afaik, every single const generic type needs to be explicitly named somewhere. You can't loop over const generics. I can make it prettier with macros, but I don't think it's possible to extract everything into constants.

I can think of a few potential solutions, but they will require re-architecting.

[alice-i-cecile]

Hmm, that's surprising 🤔 I'm not familiar enough with the details of macros to know if what I'm asking for is possible. We should see if we can get more expert opinions.

[alice-i-cecile]

Initial digging is promising:

1. https://stackoverflow.com/questions/67038789/writing-const-generic-enum-combinations-using-macro
2. https://stackoverflow.com/a/70233895
3. https://users.rust-lang.org/t/how-to-loop-in-const-generics/69042

And bingo: https://docs.rs/seq-macro/latest/seq_macro/

[alice-i-cecile]

This has my approval now :) I'd still like to be able to extract out the number of systems / plugins into constants, but I won't block on it in this PR if you can't get it working.

[joseph-gio]

I'd still like to be able to extract out the number of systems / plugins into constants, but I won't block on it in this PR if you can't get it working

So here's what I've gathered so far.

• seq_macro only works for literal consts, not const generics.
• Refactoring the code to not use const generics adds a ton of noise to the benchmarks (maybe because the compiler can't unroll the system-adding loops?). incorrect

How would you feel about moving the system-adding code out of benchmark, and only timing the first call to SystemStage::run? Originally I wanted to include add_system() calls in the benchmarks since that is a real part of bevy apps, but it might not be worth it.

[alice-i-cecile]

Sounds good. We can clean this up more later.

[joseph-gio]

Closing this since it's very out of date. If someone wants to update or replicate this, feel free.