gl_WorkGroupSize is {1,1,1} if it precedes workgroup size declaration in source

[martty]

Using the variable gl_WorkGroupSize before a workgroup size declaration in the source results in the builtin having the value {1,1,1}. I think this doesn't align with the specification, but even if it does, I think this behavior is counterintuitive and bug-prone, and should merit at least a warning or an error.

#version 460

layout(binding = 0) buffer d {
    uvec3 data1;
    uvec3 data2;
};

void fn(){
    data1 = gl_WorkGroupSize; // {1,1,1}
}

layout(local_size_x = 64) in;

void main(){
    fn();
    data2 = gl_WorkGroupSize; // {64,1,1}
}

http://shader-playground.timjones.io/a5c1b173a1fd1458c9059677b745b6c0

[greg-lunarg]

I agree that this is a bug, based on this passage from the GLSL spec:

It is a compile-time error to use gl_WorkGroupSize in a shader that does not declare a fixed
workgroup size, or before that shader has declared a fixed workgroup size, using local_size_x,
local_size_y, and local_size_z.

I interpret the word before here as referring to static position, since a dynamic interpretation is trivially true.

I consequently believe that the compiler should have given an error on the usage of gl_WorkgroupSize in fn() since it had not been completely initialized at this static position in the program.

I believe the GLSL spec is clear on this point and does not need to be updated.

I will ping the GLSL spec editor @gnl21 to confirm.

Would you like to submit a PR or have someone else fix it? I should be able to get to this in the near-to-medium term.

[gnl21]

I agree that this is a bug and that it should be an error. Perhaps the spec should say that no compilation unit (rather than shader) may use the workgroup size without declaring it.

[greg-lunarg]

Fixed with merge of #2549