CUDA in win/linux guest not available when device override used about vgpu_unlock-rs

[rliwoch]

BTW.
cuda_enabled seems to be doing it's job when the device id is not spoofed. Taking M40 24G as an example, that gets spoofed on the mdev level as a M60 - NVIDIA in their docs says only the 8Q profile is capable of CUDA:

but with the cuda_enabled flag I was able to happily run CUDA apps on the 2Q profile (providing there's no device spoofing for the guest OS as decribed above).

Also, going back to the guest spoofing - it doesn't matter where it's spoofed - the same behaviour when I spoof the device using your config.toml or that's an override of device id on the KVM level.

EDIT: Goes without saying - I'm happy to test any ideas you might have.

[mbilker]

CUDA is not supported when spoofing the PCI device ID. So far, it only works with the GRID guest driver.

[rliwoch]

Have you ever investigated why is that?
I haven't gone through the nvidia's open source kernel modules myself, you must have, so I'm just curious what is so specific about spoofing and cuda

it only works with the GRID guest driver.

Also - that's the thing, I'm not changing the driver in the guest to be the GRID one. I'm using the standard driver, and cuda works when dev id is not spoofed

EDIT: Also OpenCL doesn't fly either - so it's not specific to CUDA API rather it's a compute capabilities thing.

[mbilker]

There is some code path done with the GRID guest driver that does not happen when doing PCI-e ID spoofing for some reason.