Looks like it's virglrenderer. This isn't like SR-IOV.
It's opengl/vulkan call interception (virgl/venus) and/or virtgpu drm native context which operates at the graphics driver level for both guest and host.
Separately, wayland memory sharing channels between guest and host need to be established. This is also done through the virtgpu pipeline.
Currently this appears to both work on Android and macOS[1]. AMDGPU also has virtgpu drm native context merged into Mesa but I believe it's still 'experimental'.
It's opengl/vulkan call interception (virgl/venus) and/or virtgpu drm native context which operates at the graphics driver level for both guest and host.
Separately, wayland memory sharing channels between guest and host need to be established. This is also done through the virtgpu pipeline.
Currently this appears to both work on Android and macOS[1]. AMDGPU also has virtgpu drm native context merged into Mesa but I believe it's still 'experimental'.
[1] <https://github.com/AsahiLinux/muvm> , <https://github.com/containers/libkrun>