[RFC PATCH 0/3] new subsystem for compute accelerator devices

[Alex Deucher]

On Tue, Oct 25, 2022 at 7:15 AM Jason Gunthorpe <jgg at nvidia.com> wrote:
>
> On Tue, Oct 25, 2022 at 12:27:11PM +1000, Dave Airlie wrote:
>
> > The userspace for those is normally bespoke like ROCm, which uses
> > amdkfd, and amdkfd doesn't operate like most device files from what I
> > know, so I'm not sure we'd want it to operate as an accel device.
>
> I intensely dislike this direction that drivers will create their own
> char devs buried inside their device driver with no support or
> supervision.
>
> We've been here before with RDMA and it is just a complete mess.
>
> Whatever special non-drm stuff amdkfd need to do should be supported
> through the new subsystem, in a proper maintainable way.

We plan to eventually move ROCm over the drm interfaces once we get
user mode queues working on non-compute queues which is already in
progress.  ROCm already uses the existing drm nodes and libdrm for a
number of things today (buffer sharing, media and compute command
submission in certain cases, etc.).  I don't see much value in the
accel nodes for AMD products at this time.  Even when we transition,
there are still a bunch of things that we'd need to think about, so
the current kfd node may stick around until we figure out a plan for
those areas.  E.g., the kfd node provides platform level compute
topology information; e.g., the NUMA details for connected GPUs and
CPUs, non-GPU compute node information, cache level topologies, etc.

Alex

>
> Jason