SSAO only samples mip 0 leading to worse than expected performance.

[Elabajaba]

Bevy version

0.12/main/probably every version since ssao was added

Relevant system information

AdapterInfo { name: "AMD Radeon RX 6800 XT", vendor: 4098, device: 29631, device_type: DiscreteGpu, driver: "AMD proprietary driver", driver_info: "22.11.2", backend: Vulkan }

Also happens on dx12.

What you did

Add SSAO to a scene or run the ssao example, profile it with radeon gpu profiler to see how long the ssao passes take, compare performance with the XeGTAO sample it's based on (https://github.com/GameTechDev/XeGTAO), then spend a few days trying to figure out why bevy's is so much slower.

What went wrong

This line always gives a negative mip value because the samples are all between 0..1 (and thus the length of the samples has a negative log2 value), which causes it to be clamped to 0.0 so we're always sampling mip 0 which has a huge performance impact.

XeGTAO uses a "working depth buffer" that seems to have a range of 0-10000, with the closest thing in the demo scene being ~1.4.

Additional information

Simply multiplying the sample on that line by 10000 doesn't really appear to have a visual effect in static scenes (but improves performance from 3.2ms to 1ms for the main SSAO pass on my gpu at 4k), but in the ssao example it artifacts pretty badly.

I believe the fix for the artifacting is to improve the denoiser (eg. by just copying the xegtao one).

Turns out it's using pixel space to calculate the mip level, so we need to multiply by the resolution instead of a magical 10000 value.

Ssao.mp4