conservative max_storage_buffers_per_shader_stage

[ChristopherBiscardi]

Description

While upgrading Bevy to wgpu 29 (pr link), we're seeing new validation limits on macos m-series that lower the max_storage_buffers_per_shader_stage to 8.

Unfortunately, our early and late occlusion culling pipelines use 10 and 9 bindings respectively, which leads to occlusion culling breaking on macos specifically (other platforms seem fine).

Repro steps

1. check out Bevy's wgpu 29 PR: Wgpu 29 bevyengine/bevy#23277
2. Run an example like cargo run --example 3d_scene or the occlusion-culling specific cargo run --example occlusion_culling. They both show the validation issue, but the occlusion culling example will spam the output, making it harder to read.

Expected vs observed behavior

The limits seem to be too low (or perhaps we should set them differently now?), since this behavior was working in bevy v0.18.1, before the new limits.

Platform

macos m-series, wgpu 29

[cwfitzgerald]

This is on the agenda for next wgpu meeting.

[mate-h]

+1 , Running into this issue for porting Solari to mac as well.

[teoxoy]

Would using uniform buffers for some of the data work? There are 12 of those per shader stage too.

[teoxoy]

I think the only way to get more is via argument buffers.

[inner-daemons]

@teoxoy Even tier 1 argument buffers aren't supported on Mac1 (intel macs from as recently as 2020), so that's probably not the right approach for bevy. It seems that to fix this everywhere we'd either have to reopen the issues your PR closed by making limits higher than may be technically correct, or introduce new limits/validation which isn't really doable in a patch release as would be required for bevy.

Update: see Connor's comment

[wingertge]

The approach as of v29 also just doesn't make any sense for compute shaders, vertex shaders and acceleration structures aren't even relevant to them, and uniforms are far less common. So the partitioning of the limit into each buffer type seems extremely arbitrary and focused on just one specific use case. There has to be a better approach to the limits, but it would probably require additional properties beyond what's in the spec.

[inner-daemons]

@wingertge Yeah, the current approach is more focused on the spec I think. We would need more complex limits and validation in order to express these requirements fully. Which would of course require more changes than patch releases allow for. As said above this will be discussed in 2 days.

[cwfitzgerald]

@inner-daemons The feature set tables are lying to you, they don't acknowledge the existance of pre-mac2 devices. See the docs of the argument buffer tiers.

Tier 1 argument buffers are supported on all iOS, tvOS, and macOS GPUs.

Starting support at MacOS 10.13.

Tier 1 argument buffers can be implemented on any device, basically. See a copy of the feature set tables from 2019 https://web.archive.org/web/20190718194229/https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

[inner-daemons]

It looks like tier 1 wouldn't actually solve any problems. Tier 2 isn't even supported on iPhone X (which is what I and my family have), unless I'm misreading this.