On Wed, Mar 24, 2021 at 7:39 PM Maciej Kwapulinski
[off-list ref] wrote:

This submission is a kernel driver to support Intel(R) Gaussian & Neural
Accelerator (Intel(R) GNA). Intel(R) GNA is a PCI-based neural co-processor
available on multiple Intel platforms.

I clearly remember Olof Johansson talking about the potential need
of creating a kernel subsystem for inference engines, so I believe he
wants to be in on this discussion.

There is already misc/habanalabs, and I personally feel this is already
going in the same direction as did pin control before we standardized
it (somewhat), with vendors claiming they are all necessarily different.

So are they necessarily different? New frontiers in the Wild West
every vendor shooting from the hip without any attempts at
standardizing this thing?

Habanalabs was first at this and they made it in, has there been
any attempt to see if the two drivers could actually share code or
have anything in common? Could they share interfaces to userspace?
That kind of thing.

In the end what kernel users want is to be able to write a
userspace making use of any kind of inference/statistics engine
without having to worry about the underlying hardware, this is
what abstractions are for.

The driver works with Intel(R) libraries in user space. The Intel(R) driver
exposes a few IOCTL interfaces for use by libraries in user space. The
libraries are open sourced and are available at:
https://github.com/intel/gna

This is nice.

Have you made any attempts to cooperate with anyone else in the
world on this, or is this Intel's personal playground?

Yours,
Linus Walleij