Vulkano does not detect incorrect work group sizes

[HadrienG2]

My little validation layer has just told me that if I set an invalid workgroup size (e.g. 1025) using a specialization constant, that invalid setting will propagate all the way through the compute pipeline build process without any vulkano function erroring out in any place, final pipeline build command included.

Given that vulkano's high level API aims to detect and report invalid Vulkan API usage, and specialization constants are the officially sanctioned way to set a compute shader's workgroup size from CPU code, this behavior does not look right.

[Rua]

Does Vulkano give you a validation error if you hardcode the size in the shader?

[HadrienG2]

With a hardcoded size that is too large, shaderc will refuse to compile the GLSL to SPIR-V (because it has access to gl_MaxComputeWorkGroupSize), so for GLSL shaders at least that is not an issue.

[Rua]

That's strange. How can it know the value of gl_MaxComputeWorkGroupSize on the target machine that the shader will run on?

[HadrienG2]

Now that you mention it, that is strange indeed. I should check the GLSL spec again, maybe that's actually always 1024 on all hardware even if that feels strange because usually the gl_MaxStuff limits are more of the "at least X" kind?

In any case, the vulkano_shaders build error message, which likely comes from shaderc, looks like this:

error: compilation error:
       src/square.comp:9: error: 'local_size' : too large; see gl_MaxComputeWorkGroupSize
  --> src/square.rs:32:15
   |
32 |         path: "src/square.comp"
   |               ^^^^^^^^^^^^^^^^^

error[E0425]: cannot find function `load` in module `shader`
  --> src/square.rs:62:26
   |
62 |     let module = shader::load(device)?;
   | 

[Rua]

In any case, if the workgroup size is too high for the limits of the device, and it's not erroring out when building the pipeline, then that is a Vulkano bug. Can you confirm that that is the case?

[HadrienG2]

Indeed that is the case!