[Performance + VRAM] Re-implement 16-byte Vertex Format + use 4 Byte alignment

[thr3343]

The 16-Byte vertex format for terrain was planned in earlier versions of VulkanMod, but was scrapped due to causing performance regressions on AMD's GCN architecture

This patch exploits a workaround by doubling the vertex byte alignment to 4 bytes instead of 2, which fixes the regression and allows the 16-byte format to be used on GCN at full performance.

The 16-Byte vertex format provides the following advantages over the current 20-byte format

• VRAM usage reduced by 20%
• Reduced bandwidth usage when loading/updating Chunks
• FPS + Performance improvements on some hardware (e.g. Nvidia Turing or later)

[thr3343]

This doesn't seem to improve performance on Nvidia RTX 2000+, while older tests with alternate versions of the 16-Byte format did

I suspect this is due to alignment again
where perhaps Nvidia requires a smaller 2 byte alignment for full performance, unlike GCN which requires 4 bytes

//2 Byte aligned: At least 10%+ FPS on Nvidia Turing+, but has Performance regressions on GCN
layout (location = 0) in ivec4 Position;
layout (location = 1) in vec4 Color;
layout (location = 2) in uvec2 UV0;

//4 Byte aligned: Slower due to not hitting the fast FP16 path on Nvidia RTX 2000+
layout (location = 0) in ivec4 Position;
layout (location = 1) in vec4 Color;
layout (location = 2) in uint UV0;

This is problematic as this PR should improve performance on Nvidia not decrease it

So will mark as draft until this missing performance uplift on Nvidia is fixed

[smiRaphi]

Hey, could this be updated? as it fixes all the vram issues I have with the mod (with the newest version I can't even get to a rd of 96 but with this I can get to 164)

[smiRaphi]

thanks but everything looks a bit... off

[thr3343]

MB made a typo in the terrain shader (Incorrect multiplication)
The issue should be fixed now

(i.e. Caused the GPU to use the wrong UVs for the block atlas)

[thr3343]

Has been added into upstream as of b35e7a3: PR has served its purpose, so will close

If i'm not mistaken Nvidia Turing or later needs 2 byte alignment (i.e. not using 4 byte attributes in vertex Format) for optimal performance, but never confirmed this for sure

(Lack the free time currently to do community test builds)