Add a --gpu-robustness debug option

[spencer-lunarg]

Hey, so I am the tech lead of the Vulkan Validation Layers

We are trying to create a tool called GPU-AV which is an optional setting the Validation Layers to help detect errors that occur on the GPU.

While testing various Godot samples and open sourced games with the tool, I noticed a lot of the larger shaders have these huge SSBO structs with dozen of members. One of our checks is for making sure the bound VkBuffer was bound so that when you access the SSBO you are not going OOB

I noticed in https://github.com/godotengine/godot/blob/ce94b26de74a0a0f2463c5dc61b9b683a1e3676b/drivers/vulkan/rendering_device_driver_vulkan.cpp#L671 that robustBufferAccess is turned off (for good reasons)

The GPU-AV tool has a "force on robustness for me" option, which will help speed up us having to add this check, but I think this could be used as a proper debugging option inside Godot itself (as I know it is used in many other engines)

Basically by having a --gpu-robustness option that turns on the various robustness options, if someone is debugging a the harder GPU level issue, they can easily turn on the setting and if the problem is magically fixed, it can help lead people to narrow down the issue.

I have not tried looking into the Godot source, and seems like something I could try and prototype, but wanted to first gauge interest from those who are working on the core Vulkan rendering backend

Some links for more information:

• https://vulkan.org/user/pages/09.events/vulkanised-2025/T39-Spencer-Fricke-LunarG.pdf
• https://docs.vulkan.org/guide/latest/robustness.html

[Calinou]

Are there use cases for using robust buffer access without validation layers? If not, --gpu-validation could automatically enable robust buffer access.

[spencer-lunarg]

In theory Validation lets you know if things doing are incorrect or not, turning on robustness now "alters" what being ran

For example if we (Vulkan Validation GPU-Av) detects robustness, we won't validate that stuff (for sake of performance) because it was indicated from the user they know they expect that behavior and its not invalid

Maybe it would be good enough for 95% of uses cases to just bundle robustness with --gpu-validation