Re: 回复: kernel.org 6.5.4 , NPU driver, --not support (RFC)

From: Oded Gabbay
Date: Thu Jan 18 2024 - 09:26:52 EST

Next message: Daniel Thompson: "[GIT PULL] kgdb changes for v6.8"
Previous message: Jason Gunthorpe: "Re: [PATCH] iommu/arm-smmu-v3: Add a threshold to avoid potential soft lockup"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On 12/01/2024 8:23, Cancan Chang wrote:

Hi Oded,

Thanks for your code review. I‘m working on cleaning up the OS-wrapper code
and adjusting driver submodules. I will fix it befor sending patch series.

On the other hand, is it necessary to covert the driver to use drm? Amlogic's driver
is not only used in the visual field, but also in voice-related field, and is not very
compatible with drm. It hardly enjoy the convenience provided by drm. So，can we
do it like accel/habanalabs which has almost no use of drm, is this friendly to you?

thanks.

habanalabs is a special case as it was a couple of years already in drivers/misc, and therefore didn't use drm mechanisms at all. If we would have upstreamed a new driver, the expectation would have been that we would use some drm mechanisms.

I think that the minimal requirement is to use GEM/BOs for memory management operations. At least I see both qaic and ivpu use it.

ofc, you need to register your device through drm/accel and use the sysfs/debugfs drm infra.

If you have device memory that the user can allocate and map, then you will also need to use TTM.

I don't think you need to use drm_sched as it is tightly coupled with how GPUs work.

Oded

________________________________________
发件人: Oded Gabbay <ogabbay@xxxxxxxxxx>
发送时间: 2024年1月11日 21:14
收件人: Tomeu Vizoso; Cancan Chang
抄送: Jagan Teki; linux-media; linux-kernel; Dave Airlie; Daniel Vetter
主题: Re: kernel.org 6.5.4 , NPU driver, --not support (RFC)

[ EXTERNAL EMAIL ]

On 11/01/2024 10:04, Tomeu Vizoso wrote:

Hi Oded,

Out of curiosity, did you end up taking a look at Amlogic's driver?

Cheers,

Tomeu

Hi Tomeu,
Yes, I have looked at the driver's code. It was not an in-depth review,
but I tried to mainly understand the features the driver provide to the
user and how much complex it is.

From what I could see, this is a full-fledged accelerator which
requires command submission/completion handling, memory management,
information and debug capabilities and more.

Therefore, I do think the correct place is in the accel sub-system,
which will require you to convert the driver to use drm (we can discuss
exactly what is the level of integration required).

As I said, I didn't do a full-fledged review, but please note the driver
has a lot of OS-wrapper code, which is not acceptable in the Linux
kernel, so you will have to clean all the up.

Thanks,
Oded

On Sat, Oct 7, 2023 at 8:37 AM Cancan Chang <Cancan.Chang@xxxxxxxxxxx> wrote:

Oded,
You can get the driver code from github link： https://github.com/OldDaddy9/driver
e.g. git clone https://github.com/OldDaddy9/driver.git

________________________________________
发件人: Oded Gabbay <ogabbay@xxxxxxxxxx>
发送时间: 2023年10月3日 18:52
收件人: Cancan Chang
抄送: Jagan Teki; linux-media; linux-kernel; Dave Airlie; Daniel Vetter
主题: Re: kernel.org 6.5.4 , NPU driver, --not support (RFC)

[ EXTERNAL EMAIL ]

On Thu, Sep 28, 2023 at 11:16 AM Cancan Chang <Cancan.Chang@xxxxxxxxxxx> wrote:

“What happens if you call this again without waiting for the previous
inference to complete ?”
--- There is a work-queue in the driver to manage inference tasks.
When two consecutive inference tasks occur, the second inference task will be add to
the "pending list". While the previous inference task ends, the second inference task will
switch to the "scheduled list", and be executed.
Each inference task has an id, "inferece" and "wait until finish" are paired.

thanks

Thanks for the clarification.
I'll wait for your driver's code link. It doesn't have to be a patch
series at this point. A link to a git repo is enough.
I just want to do a quick pass.

Thanks,
Oded

________________________________________
发件人: Oded Gabbay <ogabbay@xxxxxxxxxx>
发送时间: 2023年9月28日 15:40
收件人: Cancan Chang
抄送: Jagan Teki; linux-media; linux-kernel; Dave Airlie; Daniel Vetter
主题: Re: kernel.org 6.5.4 , NPU driver, --not support (RFC)

[ EXTERNAL EMAIL ]

On Thu, Sep 28, 2023 at 10:25 AM Cancan Chang <Cancan.Chang@xxxxxxxxxxx> wrote:

“Could you please post a link to the driver's source code ?
In addition, could you please elaborate which userspace libraries
exists that work with your driver ? Are any of them open-source ?”
--- We will prepare the adla driver link after the holiday on October 6th.
It's a pity that there is no open-source userspace library.
But you can probably understand it through a workflow, which can be simplified as:
1. create model context
ret = ioctl(context->fd, ADLAK_IOCTL_REGISTER_NETWORK, &desc);
2. set inputs
3. inference
ret = ioctl(context->fd, ADLAK_IOCTL_INVOKE, &invoke_dec);

What happens if you call this again without waiting for the previous
inference to complete ?
Oded

4. wait for the inference to complete
ret = ioctl(context->fd, ADLAK_IOCTL_WAIT_UNTIL_FINISH, &stat_req_desc);
5. destroy model context
ret = ioctl(context->fd, ADLAK_IOCTL_DESTROY_NETWORK, &submit_del);

thanks

________________________________________
发件人: Oded Gabbay <ogabbay@xxxxxxxxxx>
发送时间: 2023年9月28日 13:28
收件人: Cancan Chang
抄送: Jagan Teki; linux-media; linux-kernel; Dave Airlie; Daniel Vetter
主题: Re: kernel.org 6.5.4 , NPU driver, --not support (RFC)

[ EXTERNAL EMAIL ]

On Wed, Sep 27, 2023 at 10:01 AM Cancan Chang <Cancan.Chang@xxxxxxxxxxx> wrote:

“Or do you handle one cmd at a time, where the user sends a cmd buffer
to the driver and the driver then submit it by writing to a couple of
registers and polls on some status register until its done, or waits
for an interrupt to mark it as done ?”
--- yes， user sends a cmd buffer to driver, and driver triggers hardware by writing to register,
and then, waits for an interrupt to mark it as done.

My current driver is very different from drm, so I want to know if I have to switch to drm？

Could you please post a link to the driver's source code ?
In addition, could you please elaborate which userspace libraries
exists that work with your driver ? Are any of them open-source ?

Maybe I can refer to /driver/accel/habanalabs.

That's definitely a possibility.

Oded

thanks

________________________________________
发件人: Oded Gabbay <ogabbay@xxxxxxxxxx>
发送时间: 2023年9月26日 20:54
收件人: Cancan Chang
抄送: Jagan Teki; linux-media; linux-kernel; Dave Airlie; Daniel Vetter
主题: Re: kernel.org 6.5.4 , NPU driver, --not support (RFC)

[ EXTERNAL EMAIL ]

On Mon, Sep 25, 2023 at 12:29 PM Cancan Chang <Cancan.Chang@xxxxxxxxxxx> wrote:

Thank you for your reply from Jagan & Oded.

It is very appropritate for my driver to be placed in driver/accel.

My accelerator is named ADLA(Amlogic Deep Learning Accelerator).
It is an IP in SOC,mainly used for neural network models acceleration.
It will split and compile the neural network model into a private format cmd buffer,
and submit this cmd buffer to ADLA hardware. It is not programmable device.

What exactly does it mean to "submit this cmd buffer to ADLA hardware" ?

Does your h/w provides queues for the user/driver to put their
workloads/cmd-bufs on them ? And does it provide some completion queue
to notify when the work is completed?

Or do you handle one cmd at a time, where the user sends a cmd buffer
to the driver and the driver then submit it by writing to a couple of
registers and polls on some status register until its done, or waits
for an interrupt to mark it as done ?

ADLA includes four hardware engines:
RS engines : working for the reshape operators
MAC engines : working for the convolution operators
DW engines : working for the planer & Elementwise operators
Activation engines : working for activation operators(ReLu,tanh..)

By the way, my IP is mainly used for SOC, and the current driver registration is through the platform_driver,
is it necessary to switch to drm?

This probably depends on the answer to my question above. btw, there
are drivers in drm that handle IPs that are part of an SOC, so
platform_driver is supported.

Oded

thanks.

________________________________________
发件人: Oded Gabbay <ogabbay@xxxxxxxxxx>
发送时间: 2023年9月22日 23:08
收件人: Jagan Teki
抄送: Cancan Chang; linux-media; linux-kernel; Dave Airlie; Daniel Vetter
主题: Re: kernel.org 6.5.4 , NPU driver, --not support (RFC)

[你通常不会收到来自 ogabbay@xxxxxxxxxx 的电子邮件。请访问 https://aka.ms/LearnAboutSenderIdentification，以了解这一点为什么很重要;]

[ EXTERNAL EMAIL ]

On Fri, Sep 22, 2023 at 12:38 PM Jagan Teki <jagan@xxxxxxxxxx> wrote:

On Fri, 22 Sept 2023 at 15:04, Cancan Chang <Cancan.Chang@xxxxxxxxxxx> wrote:

Dear Media Maintainers:
Thanks for your attention. Before describing my problem，let me introduce to you what I mean by NPU.
NPU is Neural Processing Unit, It is designed for deep learning acceleration, It is also called TPU, APU ..

The real problems:
When I was about to upstream my NPU driver codes to linux mainline, i meet two problems:
1. According to my research, There is no NPU module path in the linux (base on linux 6.5.4) , I have searched all linux projects and found no organization or comany that has submitted NPU code. Is there a path prepared for NPU driver currently?
2. If there is no NPU driver path currently, I am going to put my NPU driver code in the drivers/media/platform/amlogic/ , because my NPU driver belongs to amlogic. and amlogic NPU is mainly used for AI vision applications. Is this plan suitabe for you?

If I'm correct about the discussion with Oded Gabby before. I think
the drivers/accel/ is proper for AI Accelerators including NPU.

+ Oded in case he can comment.

Thanks,
Jagan.

Thanks Jagan for adding me to this thread. Adding Dave & Daniel as well.

Indeed, the drivers/accel is the place for Accelerators, mainly for
AI/Deep-Learning accelerators.
We currently have 3 drivers there already.

The accel subsystem is part of the larger drm subsystem. Basically, to
get into accel, you need to integrate your driver with the drm at the
basic level (registering a device, hooking up with the proper
callbacks). ofc the more you use code from drm, the better.
You can take a look at the drivers under accel for some examples on
how to do that.

Could you please describe in a couple of sentences what your
accelerator does, which engines it contains, how you program it. i.e.
Is it a fixed-function device where you write to a couple of registers
to execute workloads, or is it a fully programmable device where you
load compiled code into it (GPU style) ?

For better background on the accel subsystem, please read the following:
https://docs.kernel.org/accel/introduction.html
This introduction also contains links to other important email threads
and to Dave Airlie's BOF summary in LPC2022.

Thanks,
Oded

Next message: Daniel Thompson: "[GIT PULL] kgdb changes for v6.8"
Previous message: Jason Gunthorpe: "Re: [PATCH] iommu/arm-smmu-v3: Add a threshold to avoid potential soft lockup"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]