I need to index into a texture array using indices which are not dynamically uniform. This works fine on NVIDIA chips but you can see the artifacts on AMD due to the wavefront problem. This means, a lot of pixel invocations get the wrong index value. I know you fix this by using NonUniformResourceIndex in hlsl. Is there an equivalent for Vulkan glsl?

This is the shader code for reference. As you can see, index is an arbitrary value for each pixel and is not dynamically uniform. I fix this for hlsl by using NonUniformResourceIndex(index)

layout(set = 0, binding = 0) uniform sampler textureSampler;
layout(set = 0, binding = 1) uniform texture2D albedoMaps[256];

layout(location = 0) out vec4 oColor;

void main()
{
	uint index = calculate_arbitrary_texture_index();
	vec2 texCoord = calculate_texcoord();
	vec4 albedo = texture(sampler2D(albedoMaps[index], textureSampler), texCoord);

	oColor = albedo;
}

Thank you

Hey Matias,

Thanks for the code. But I am still getting the wavefront tiling artifacts. Is there no way to get over this problem? Since hlsl can do it, I assume it is not a hardware limitation.

This makes me curious - How do cross compilers handle this when translating hlsl to glsl or they just don support it

7 hours ago, mark_braga said:

Hey Matias,

Thanks for the code. But I am still getting the wavefront tiling artifacts. Is there no way to get over this problem? Since hlsl can do it, I assume it is not a hardware limitation.

This makes me curious - How do cross compilers handle this when translating hlsl to glsl or they just don support it

This is indeed a hardware limitation. Texture descriptor are scalar register on GCN. If your index is not uniform across the lanes. It creates the artefacts you see. What the NonUniform intrinsic does is to generate a loop, masking threads per index and do the texture fetch until all lanes have done it. It does not do it at the DXC/DXIL bytecode but just tag thing in case the driver has to do it.

If you have access to ballot and readfirstlane in glsl, it is easy to write yourself two version of the shader for GCN and nVidia, doing it or not depending on the GPU.