Using validation layer and mesh shaders with a uniform causes performance drop overtime

[Sharlock93]

Environment:

• OS: Windows 11 24H2 (OS Build 26100.7462)
• GPU and driver version: NVIDIA GeForce RTX 3080 Laptop GPU - VkVersion: 1.4.325 - Driver: 591.236.0
• SDK or header version if building from repo: Latest from the repo
• Options enabled (synchronization, best practices, etc.):
  VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_EXT |
  VK_VALIDATION_FEATURE_ENABLE_BEST_PRACTICES_EXT |
  VALIDATION_CHECK_ENABLE_VENDOR_SPECIFIC_NVIDIA |
  VK_VALIDATION_FEATURE_ENABLE_SYNCHRONIZATION_VALIDATION_EXT.

Describe the Issue

Performance of the application decreases overtime when validation layer is used with a simple mesh shader that has a uniform binding.
This is a very simple shader:

#version 460
#extension GL_EXT_mesh_shader : require

layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
layout(triangles, max_vertices = 3, max_primitives = 1) out;

struct Vout {
    vec2 uv;
};

layout(binding = 0) uniform camera {
    mat4 proj;
    mat4 cam;
};
layout(location = 0) out Vout v_out[];

void main() {
    SetMeshOutputsEXT(3, 1);
    gl_PrimitiveTriangleIndicesEXT[0] = uvec3(0, 1, 2);
    gl_MeshVerticesEXT[0].gl_Position = transpose(proj)*transpose(cam)*vec4(0,0,0,1);
    gl_MeshVerticesEXT[1].gl_Position = transpose(proj)*transpose(cam)*vec4(1,0,0,1);
    gl_MeshVerticesEXT[2].gl_Position = transpose(proj)*transpose(cam)*vec4(0,1,0,1);

    v_out[0].uv = vec2(0, 0);
    v_out[1].uv = vec2(0, 1);
    v_out[2].uv = vec2(1, 0);
}

I ran this through superluminal, here is what the frame time looks like:

This is what one of the higher cost frames looks like in superluminal:

I haven't looked deeper into this unfortunately. However a few points stand:

1. This doesn't happen if there is no uniform.
2. This doesn't happen in traditional pipelines.
3. This doesn't happen when validation layer is disabled.

Expected behavior
Cost of validation layer being constant and not increase overtime.

[spencer-lunarg]

@arno-lunarg this seems to be leaking from that stdext::inplace_function thing you added...can you look into this

[arno-lunarg]

I think I know the issue, or at least my PR is fixing a leak!

[Sharlock93]

Thank you for the quick fix :)

[arno-lunarg]

No problem, thank you for the good info, it helped me pinpoint the issue quickly!