Support unchecked pipelines

[Patryk27]

device.create_shader_module() has a (somewhat evil) cousin device.create_shader_module_unchecked() which allows to instantiate a shader without extra runtime checks (e.g. for array bounds).

This comes very handy for memory-intensive shaders, for instance yielding an extra 25% boost for my ray tracing thingie.

This commit introduces the concept of unchecked pipelines for Bevy abstractions as well.

[github-actions[bot]]

Welcome, new contributor!

Please make sure you've read our contributing guide and we look forward to reviewing your pull request shortly ✨

[JMS55]

Have you seen https://github.com/bevyengine/bevy/pull/9450?

Last time I tried this, I didn't get any real performance benefits. Curious to know why.

[Patryk27]

I haven't seen that pull request 👀 -- but disabling checks did yell ~25% performance boost for my ray-tracer on Metal.

Edit: in a hindsight, 25% sounds like too much, now that I think about it - realistically it must've been something like 10%; I benchmarked it about six months ago, trying to optimize my tracing shaders, and the extra bounds checks for accessing memory and managing the traversing stack quickly added up.

Edit 2: managed to find something - https://github.com/Patryk27/strolle/commit/684edd179207a3c177a9055950b0d0917ab0b907 (note that back then the create_shader_module_unchecked() function did not exist, so I just monkey-patched wgpu not to generate those checks) - according to me from the past, disabling bounds checks got me extra 10 FPS (so like 12%, considering the overall FPS=80 of my demo).

[valaphee]

Idk if it would make sense to tie this to release mode #[cfg(not(debug_assertions))] as a further release optimization

[Patryk27]

I think going with not(debug_assertions) might be too much - e.g. indexing out of bounds is safe in shaders (and with the bounds-check in place it defaults to returning zero), but it's undefined behavior ™️ when using create_shader_module_unchecked().

So it's sort of like compiling with --release made array[idx] UB when idx >= array.len() in typical Rust code - it would be pretty surprising 😅

[JMS55]

FYI I tried this recently on my meshlet shaders and none of them showed any speed improvements. They weren't memory bound for the most part, though. I don't have a memory bound shader to test it on ATM.

[Patryk27]

They weren't memory bound for the most part, though. I don't have a memory bound shader to test it on ATM.

The most important thing (for me) unchecked pipelines do is disabling array-bounds checking - by default, all array accesses like array[idx] get changed into idx < array.len() ? array[idx] : default() (which might get later optimized by the compiler downstream, but it's not always feasible); in BVH traversal, this adds noticeable overhead 👀