Get info of memory usage of resources

[torsteingrindvik]

Hi, the context is I was running a Bevy app and also watch nvidia-smi, and my GPU was using a few GBs of memory.
Is there a good way to enumerate which textures, buffers, etc. take up space?
In a simple world some tool would be able to list them in order from largest to smallest in order to prioritize where to put effort.

I saw there was something called VK_EXT_device_memory_report which might have helped, but it seems to not be implemented (EDIT: Not in wgpu nor ash, probably because it isn't ratified?)

Are there other tools which can help out?

[Wumpf]

On Windows and Linux RenderDoc is usually a great tool to debug this sort of thing

[torsteingrindvik]

I've used RenderDoc for frame debugging but I don't think it can aggregate mem usage?

I open a capture of my app and looked around, and I can see sizes of individual textures and buffers in use for a given frame but I'm looking for something like a CPU heap profiler but for the GPU.

[Wumpf]

hum, you're right. Thought RenderDoc can do that, but it doesn't :(

Looke like VK_EXT_device_memory_report is only implemented by a bunch of mobile chips today.

I think the only reliable thing you can try is to use a vendor specific tool: Actual memory used by textures etc. is actually quite tricky since gpus often do loss-less compression on one hand and device specific padding on the other, so only a vendor specific tool like NSight will give reliable numbers.

Otherwise I reckon you might still try to use RenderDoc to capture all buffers & textures and then use that information to do an estimate of the amount of memory used by processing the data (e.g. by using the python api to get all the necessary data out 🤔 )