Re: GPU virtualization in Spectrum

[Nathan Myers <ncm@cantrip.org>]

This is a very welcome update.

I wonder: is Vulkan itself practical as a virtualized
interface?  I can imagine that shaders running on behalf
of different VMs would need to be guarded against seeing
memory meant for other VMs. Might this mean swapping in
a different memory map when running on behalf of each VM,
one at a time? Overhead from swapping memory maps might
be better than no access at all.

The use case I am concerned for is exercising Vulkan as
a general compute acceleration engine, as in Kompute
(http://kompute.cc), running alongside ordinary graphics
rendering chores, and in place of proprietary CUDA. It
would be tragic for security details to make 90+% of the
computational power of our machines available only for
sterile graphical rendering.

If Vulkan could use a virtio-gpu backend, I guess this
might come out to much the same thing. But would it lack
access to GPU-specific optimizations implemented in e.g.
an AMD driver and made available via higher-level Vulkan
APIs not visible via virtio-gpu?