[PATCH v4 1/7] accel/ivpu: Introduce a new DRM driver for Intel VPU

[Jeffrey Hugo]

On 1/5/2023 5:57 AM, Daniel Vetter wrote:
> On Thu, Dec 08, 2022 at 12:07:27PM +0100, Jacek Lawrynowicz wrote:
>> +static const struct drm_driver driver = {
>> +	.driver_features = DRIVER_GEM | DRIVER_COMPUTE_ACCEL,
> 
> So I was wondering whether this is a bright idea, and whether we shouldn't
> just go ahead and infuse more meaning into accel vs render nodes.
> 
> The uapi relevant part of render nodes is that they're multi-user safe, at
> least as much as feasible. Every new open() gives you a new private
> accelerator. This also has implications on how userspace drivers iterate
> them, they just open them all in turn and check whether it's the right
> one - because userspace apis allow applications to enumerate them all.
> Which also means that any devicie initialization at open() time is a
> really bad idea.
> 
> A lot of the compute accelerators otoh (well habanalabs) are single user,
> init can be done at open() time because you only open this when you
> actually know you're going to use it.
> 
> So given this, shouldn't multi-user inference engines be more like render
> drivers, and less like accel? So DRIVER_RENDER, but still under
> drivers/accel.
> 
> This way that entire separate /dev node would actually become meaningful
> beyond just the basic bikeshed:
> - render nodes are multi user, safe to iterate and open() just for
>    iteration
> - accel nodes are single user, you really should not ever open them unless
>    you want to use them
> 
> Of course would need a doc patch :-)
> 
> Thoughts?
> -Daniel

Hmm.

I admit, I thought DRIVER_ACCEL was the same as DRIVER_RENDER, except 
that DRIVER_ACCEL dropped the "legacy" dual node setup and also avoided 
"legacy" userspace.

qaic is multi-user.  I thought habana was the same, at-least for 
inference.  Oded, am I wrong?

So, if DRIVER_ACCEL is for single-user (training?), and multi-user ends 
up in DRIVER_RENDER, that would seem to mean qaic ends up using 
DRIVER_RENDER and not DRIVER_ACCEL.  Then qaic ends up over under 
/dev/dri with both a card node (never used) and a render node.  That 
would seem to mean that the "legacy" userspace would open qaic nodes by 
default - something I understood Oded was trying to avoid.

If there really a usecase for DRIVER_ACCEL to support single-user?  I 
wonder why we can't default to multi-user, and if a particular 
user/driver has a single-user usecase, it enforces that in a driver 
specific manner?

-Jeff