Forward+/Mobile: Shadow filtering is broken with High or Ultra quality

[SaracenOne]

Tested versions

e2c6daf

System information

Godot v4.4.dev (65f3df5) - Windows 10.0.22631 - Single-window, 2 monitors - Vulkan (Forward+) - dedicated NVIDIA GeForce RTX 4090 (NVIDIA; 32.0.15.6094) - AMD Ryzen 9 7950X 16-Core Processor (32 threads)

Issue description

Found another regression introduced in e2c6daf. It looks like all filtering for shadow maps no-longer works. Once again, only affects Forward+ and mobile renderers.
Post e2c6daf

Pre e2c6daf

Steps to reproduce

• Create a directional light.
• Enable a shadow.

Minimal reproduction project (MRP)

N/A

[DarioSamo]

Was the filtering improved recently? Perhaps it was a conflict from when I rebased it. CC @clayjohn

Commenting to participate and look into it later.

[clayjohn]

There were some changes recently, but they only take effect when using TAA/FSR2.2. I don't think that is what we are seeing here

[Calinou]

I can't reproduce this in 4.4.dev 04692d8 (Linux, GeForce RTX 4090 with NVIDIA 560.35.03):

Lavapipe is also able to render it correctly:

TAA and FSR2.2 are disabled.

This can be tested in Truck Town which uses Forward+, but should render correctly when switched to Mobile too.

[DarioSamo]

I'm unable to reproduce this as well. The only way I can get the shadow to show up like in OP's picture is to turn the Light's blur parameter to 0.

Could you upload an MRP with the issue? Perhaps there's a particular configuration setting we're missing.

[Calinou]

I've finally found out the cause: it only breaks when the shadow filter quality is set to High or Ultra. This affects both Forward+ and Mobile.

This affects all light types.

Minimal reproduction project: test_shadow_filter.zip

[DarioSamo]

I've finally found out the cause: it only breaks when the shadow filter quality is set to High or Ultra.

Do you happen to know off the top of your head if these qualities correspond to a sample count >= 16? Perhaps it just needs a little expansion of the bits available to the sample count in the new specialization constants structure.

[DarioSamo]

Should be fixed by #98614, please let me know if it works.