RobustToolbox
RobustToolbox copied to clipboard
space-wizards
Consider strategy for GPU device loss.
This can happen on ANGLE and Vulkan for various reasons but we might want to handle it. TL;DR is that we'll need to be able to re-load everything onto the GPU from scratch, at any time.
Biggest consideration is texture reloading. Most textures can be reloaded from disk, but anything uploaded by content or, say, the font system, may not. Do we just cache these in host RAM?
As someone who has no clue what this means, are you seriously prepping for someone unplugging their card while the game is running? And if so, why should we even support this?
As someone who has no clue what this means, are you seriously prepping for someone unplugging their card while the game is running? And if so, why should we even support this?
AFAIK, this can happen for multiple reasons other than that such as:
- Graphics driver being updated
- Going from an from igpu to dgpu
- GPU just stops responding and has to be reset
I think one of our contributors ran into this problem while just playing SS14 normally
As someone who has no clue what this means, are you seriously prepping for someone unplugging their card while the game is running? And if so, why should we even support this?
You know that fancy Microsoft Surface where if you join the keyboard to the tablet it starts using the nvidia GPU? There you go.
I think one of our contributors ran into this problem while just playing SS14 normally
Shadow was having a problem where his GPU was being reset while resizing the window for some reason. Still sounds like a driver bug but oh well.
Oh yeah, I've had problems on Linux where unsuspending the system causes GPU memory to be cleared. In this case the driver would report a GPU device loss and require us to re-initialize everything instead, buuut desktop OpenGL traditionally had no concept of this so the GL driver is forced to keep a copy of every texture/buffer in host RAM, which is incredibly wasteful. It also can't do this for any GPU-generated textures (anything rendered on the GPU to a framebuffer) which means that many programs have broken screens upon unsuspend.
My understanding is that if we handle GPU device loss ourselves (via GL_ARB_robustness and platform-specific extensions) then the GL driver does not have to keep a host-memory copy and this would reduce memory usage drastically (we can reload textures from disk in this rare scenario instead of keeping them in memory).
A problem is that GLFW requires you to re-create the whole window. This is mildly annoying to say the least.