Re: [PATCH] GPU acceleration

[Alyssa Ross <hi@alyssa.is>]

[-- Attachment #1: Type: text/plain, Size: 4353 bytes --]

Demi Marie Obenour <demiobenour@gmail.com> writes:

> On 1/29/26 09:10, Alyssa Ross wrote:
>> Worth marking patches like this as RFC or similar in future, so
>> everybody looking at it is on the same page about its status.  (I'm
>> assuming this was meant as a proof of concept and not meant to be
>> applied as is.)
>
> With the hash change, I was hoping it could be applied.  I forgot to
> run scripts/genfiles.sh, though, and completely missed the missing
> fallback in crosvm.
>
>> Demi Marie Obenour <demiobenour@gmail.com> writes:
>> 
>>> This exposes GPU acceleration to VMs via virtio-GPU native contexts on
>>> AMD and Qualcomm GPUs.  Apple GPU support also exists but might require
>>> patches to virglrenderer that are not upstream yet.
>>>
>>> On a system with an AMD GPU, this has been tested by running vkcube in
>>> the VM.  It reports that the AMD GPU is being used for rendering.
>>>
>>> Signed-off-by: Demi Marie Obenour <demiobenour@gmail.com>
>>> ---
>>> I used an empty string for the hashes in 'gitfetch' and
>>> 'buildRustPackage'.  This is, of course, incorrect.  The correct value
>>> according to my own testing is
>>> "sha256-EOMkQ0aPRjsowdGuZjy5K1yKyKEzd5AVYxaECTz7n6k=" (git hash) and
>>> "sha256-k3dmxIuCQoOrn/VwauTdzuRw/XKQB6LPLgO5ql0rE7E=" (cargoHash).
>>> However, these should be validated before applying them, which is why I
>>> didn't include them in the patch themselves.  This is a security
>>> precaution: anyone who wants to inject malicious content must serve it
>>> to multiple people, not just one.
>>> ---
>>>  .../template/data/service/vhost-user-gpu/run            |  2 +-
>>>  img/app/Makefile                                        |  2 +-
>>>  img/app/default.nix                                     | 14 ++++++++++++--
>>>  img/app/image/etc/mdev.conf                             |  1 +
>>>  .../notification-fd                                     |  0
>>>  .../notification-fd.license                             |  0
>>>  .../{wayland-proxy-virtwl => wl-cross-domain-proxy}/run | 17 +++++------------
>>>  .../type                                                |  0
>>>  .../type.license                                        |  0
>>>  9 files changed, 20 insertions(+), 16 deletions(-)
>>>
>>> diff --git a/host/rootfs/image/etc/s6-linux-init/run-image/service/vm-services/template/data/service/vhost-user-gpu/run b/host/rootfs/image/etc/s6-linux-init/run-image/service/vm-services/template/data/service/vhost-user-gpu/run
>>> index fb9ac9971aef82dabe0b54c1299ac8c66d133eb5..87d72f55e293ea81b6f4aa12786a993bafc623e2 100755
>>> --- a/host/rootfs/image/etc/s6-linux-init/run-image/service/vm-services/template/data/service/vhost-user-gpu/run
>>> +++ b/host/rootfs/image/etc/s6-linux-init/run-image/service/vm-services/template/data/service/vhost-user-gpu/run
>>> @@ -42,4 +42,4 @@ bwrap
>>>  crosvm --no-syslog device gpu
>>>    --fd 0
>>>    --wayland-sock $WAYLAND_DISPLAY
>>> -  --params "{\"context-types\":\"cross-domain\"}"
>>> +  --params "{\"context-types\":\"cross-domain:drm\"}"
>> 
>> I suppose we'll need to get some sort of graceful fallback into
>> crosvm/rutabaga, or to somehow detect native contexts support in this
>> script before we enable it on the crosvm command line?
>> 
>> 2026-01-29 14:06:30.082434936  [2026-01-29T14:06:30.082073581+00:00 ERROR rutabaga_gfx::virgl_renderer] failed to initialize drm renderer
>> 2026-01-29 14:06:30.082842461  [2026-01-29T14:06:30.082652949+00:00 WARN  rutabaga_gfx::rutabaga_core] error initializing gpu backend=virglrenderer, falling back to 2d.
>> 2026-01-29 14:06:30.099779984  [2026-01-29T14:06:30.099379669+00:00 WARN  devices::virtio::gpu::virtio_gpu] virtio-gpu get_capset_info(index=1) failed. intentionally poisoning response
>
> Ideally crosvm would fallback automatically, but on ChromeOS it never
> needs to because it is built specifically for its target hardware.

Okay, but crosvm is open to contributions.  They're willing to accept
affordances for non-Chrome OS use.  Even Google doesn't use it
exclusively on Chrome OS…

> It should be sufficient to detect if an AMD or Qualcomm GPU is
> available and the user has enabled GPU acceleration.

That doesn't sound very good, because then we'll have to keep updating
it as new GPUs are supported.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 227 bytes --]