Cascaded shadow map artifact

[Josh]

I'm seeing these artifacts at the border region of CSM shadows.

[reepblue]

Yep, first brought it up here.

 

[Josh]

If I isolate each stage in the shader, each one displays fine. I think this is an error in the driver.

Added this to PBR/Lighting.glsl at line 269:

if (index != 1) return lighttype;

 

[Josh]

I did some more testing and it looks like it might be a memory barrier issue, maybe. The pattern that appears looks similar to another problem I saw in post-effects, but I don't think the directional light shadow stages are all rendering each frame.

You can see here the divisions between stages are clean.

 

 

[Josh]

It's got to be a driver bug, or some obscure GLSL rule I don't know about. If I isolate one stage the shadowmap looks fine.

It looks to me like there is a problem when the texture that is used in a lookup is not the same across near neighboring pixels...

[Josh]

Perhaps another manifestation of the same problem here:

 

[Josh]

Reported to AMD here:
https://community.amd.com/t5/opengl-vulkan/cascaded-shadow-map-artifacts-in-glsl/m-p/637524#M5003

Also asking on Khronos forum:
https://community.khronos.org/t/glsl-shader-logic-on-amd-cards/110143

[Josh]

Probably related, you can see a similar error in the terrain shader.

[Josh]

I got an interesting response that sounds right:

Quote

 

This might be an artifact resulting from indexing and sampling textures using a non Dynamically uniform expression. AMD hardware needs the sample index to be the same for all active invocations inside the subgroup. 

One fix would be to do:

#extension GL_EXT_nonuniform_qualifier : require textures[nonuniformEXT(index)]

So wrapping the texture index inside of the nonuniformEXT function which exist exactly for cases where you need non dynamically uniform ressource indexing.

 

https://community.amd.com/t5/opengl-vulkan/cascaded-shadow-map-artifacts-in-glsl/m-p/637976#M5009

[Josh]

Maybe an explanation:
https://community.khronos.org/t/cascaded-shadow-map-bug-between-splits/66929/11

[Josh]

I updated the example with the Vulkan 1.2 nonuniform features and sent a new download link to AMD.

[Josh]

The core problem of this issue is now solved and will be included in an upcoming update.

I never did get the nonuniform feature working. I am not sure if it even works, and I am not sure if it has a negative impact on performance, and I just don't care,

 

[Josh]

On 10/11/2023 at 6:49 PM, Josh said:

Probably related, you can see a similar error in the terrain shader.

This seems to be fixed on AMD. I am not sure why because I did not merge the terrain textures like I did shadow maps.