On 5/24/23 12:40, Daniel Vetter wrote:
On Wed, May 24, 2023 at 01:27:00PM +0300, Oded Gabbay wrote:
On Wed, May 24, 2023 at 2:34 AM Kevin Hilman khilman@baylibre.com wrote:
Jeffrey Hugo quic_jhugo@quicinc.com writes:
On 5/17/2023 8:52 AM, Alexandre Bailon wrote:
This adds a DRM driver that implements communication between the CPU and an APU. The driver target embedded device that usually run inference using some prebuilt models. The goal is to provide common infrastructure that could be re-used to support many accelerators. Both kernel, userspace and firmware tries to use standard and existing to leverage the development and maintenance effort. The series implements two platform drivers, one for simulation and another one for the mt8183 (compatible with mt8365).
This looks like the 3 existing Accel drivers. Why is this in DRM?
Yes, this belongs in accel. I think Alex had some issues around the infra in accel with device nodes not appearing/opening properly, but I'll let him comment there. But either way, the right approach should be to fix any issues in accel and move it there.
[...]
.../devicetree/bindings/gpu/mtk,apu-drm.yaml | 38 ++ drivers/gpu/drm/Kconfig | 2 + drivers/gpu/drm/Makefile | 1 + drivers/gpu/drm/apu/Kconfig | 22 + drivers/gpu/drm/apu/Makefile | 10 + drivers/gpu/drm/apu/apu_drv.c | 282 +++++++++ drivers/gpu/drm/apu/apu_gem.c | 230 +++++++ drivers/gpu/drm/apu/apu_internal.h | 205 ++++++ drivers/gpu/drm/apu/apu_sched.c | 592 ++++++++++++++++++ drivers/gpu/drm/apu/simu_apu.c | 313 +++++++++ include/uapi/drm/apu_drm.h | 81 +++
"apu" seems too generic. We already have 3 "AI processing units" over in drivers/accel already...
Indeed, it is generic, but that's kind of the point for this driver since it's targetted at generalizing the interface with "AI processing units" on a growing number of embedded SoCs (ARM, RISC-V, etc.) In addition, the generic naming is intentional because the goal is bigger than the kernel and is working towards a generic, shared "libAPU" userspace[1], but also common firmware for DSP-style inference engines (e.g. analgous Sound Open Firmware for audio DSPs.)
As usual, the various SoC vendors use different names (APU, NPU, NN unit, etc.) but we'd like a generic name for the class of devices targetted by this driver. And unfortunately, it looks like the equally generic "Versatile processing unit" is already taken Intel's drivers/accel/ivpu. :)
Maybe since this is more about generalizing the interface between the CPU running linux and the APU, what about the name apu_if? But I guess that applies to the other 3 drivers in drivers/accell also. Hmmm...
Naming things is hard[2], so we're definitly open to other ideas. Any suggestions?
Maybe model it according to the tiny driver in drm display ? You can then call it tiny_apu :-) Disclosure: It was Daniel's suggestion, he can chime in with more details on the tiny driver concept.
Yeah so maybe a bit more detail on my thoughts:
First this smells like a need bypass of the entire "we want open userspace for accel drivers" rule. The rule isn't quite a strict as for drm gpu drivers (not sure we ended up documenting exactly what, but iirc the consensus was that for build-time only dependencies we're ok with downstream compilers), but it's still there.
What is letting you think that we want to bypass open source requirements ? Although the neural network firmware and userspace application are not yet opensource, our intention is to develop a full open source stack. Currently, we only support Mediatek APU (an Xtensa VP6) and we have to use closed source sotfware to execute inferences on the accelerator. As far I know, there software stack similar to mesa where we could add support of a new accelerator (this is also true for firmware). That is actually what we would like to do. But this will take a lot of time and we consider this driver as a first (small) step.
And at least from a quick look apu.ko and libapu just look like a generic accel interface, and that's not enough.
For the big training engines it's more or less "enough to run pytorch, but it can be really slow", not sure what the right standard for these inference-only drivers should be.
To be honest, I don't know what would be required for training engines. We only target accelerators for embedded device that usually only run inferences. In my opinion, this is 2 different use cases and I don't think we could address them in the same way.
So that's the first reason why I don't like this.
The other is that I think if we do end up with a pile of tiny accel drivers, we should probably look into something like simmpledrm for the tiny display drivers. Probably still IP specific ioctls (at least most) so that IP specific job knows and all that are easy, but then just pass to a framework that simplifies a drm gem driver to "write ptes" and "run job" callback, maybe with an optional "create/destroy vm/ctx" for hw which can do that.
So maybe we end up with a drivers/accel/tiny and a bunch more helpers around the existing gem ones. The rule we have for drm/tiny is "1 file, less than 1kloc", and there's a bunch of them. I do think we can achieve the same for tiny accel inference engines (but it's still a bit a road). Maybe tiny accel is more like "less than 5kloc" since you need a bit more glue for the driver specific ioctl stuff - maybe that's only needed for the submit ioctl, maybe also for buffer map/unmap and creation.
This makes sense to me.
Also note that there's an entire pile of in-flight work for adding new helpers to the gem world to make this all easier. Once we have gpuva and exec helpers there not much glue left to tie it all together with the scheduler.
I wrote this series a long time ago and just rebased it recently. I will take some time to see the in-flight work and see if that something I could start using.
But the real crux is that an accel inference driver really needs to have enough userspace to do an actual inference job with some android/cros/whatever framework for inference (there's just too many).
We are currently stuck with closed source fimrware, userspace applications and toolchains (works with android and linux). We are looking for a solution but implementing something will take some time.
Alexandre
-Daniel
Oded
Kevin
[1] https://gitlab.baylibre.com/baylibre/libapu/libapu
[2] "There are 2 hard problems in computer science: cache invalidation, naming things and off-by-1 errors." -- https://twitter.com/secretGeek/status/7269997868