Re: [PATCH v2 00/20] Add minimal Tensor/GS101 SoC support and Oriole/Pixel6 board

Maksym Holovach <maksym.holovach.an.2022@xxxxxxx> · Sat, 4 Nov 2023 01:05:35 +0200

Hi William,

On 11/3/23 19:36, William McVicker wrote:
Hi Maksym, Krzysztof, Peter,

On 11/03/2023, Krzysztof Kozlowski wrote:
On 03/11/2023 14:56, Maksym Holovach wrote:
Hi Peter,

On 11/3/23 15:11, Peter Griffin wrote:
Hi Maksym,

Thanks for your feedback.

On Thu, 2 Nov 2023 at 22:32, Maksym Holovach
<maksym.holovach.an.2022@xxxxxxx> wrote:
Hi, all

I wanted to inquire about how do you all feel about calling this SoC by
the Google "gs101" name.
Interesting question, I think calling it gs101 is the correct approach see
below for my rationale.

I believe the proper name for it should be the actual Samsung name,
written in the silicon and reported in the Chip ID hardware: Exynos9845.
This also touches the Tensor G2 (Exynos9855), Tensor G3 (Exynos9865),
and possibly the "Tesla" SoCs.

I do not think the Linux kernel should be a marketing material: it
should reflect reality. The chip is almost 100% composed of Samsung
Exynos IP blocks and should be called that way.
As you alluded to Tesla fsd and Axis artpec8 SoCs are also based on
Exynos designs and support upstream uses the axis,artpec8* or tesla,fsd*
compatibles.

So using google,gs101 is consistent with the existing upstream naming
scheme, for customized ASICs that were based off a Exynos design. But
it also reflects the reality that this SoC is not a Exynos9845 as there is
also a lot of Google owned and other third party IP integrated that is not
found in Exynos9845.
A quick question: Do you imply Exynos9845 exists outside of the context
of Tensor G1? I used to believe Exynos9845 **is** Tensor G1.
Yes, the gs101 SoC is *not* equivalent to the Exynos9845. Similar to how Tesla
FSD licenses Exynos IP blocks, gs101 does not only comprise of Exynos IP
blocks. The final design is unique to Google and comprises of several different
vendor IP blocks (not only Exynos).

Also, what kind of Google IP are you talking about? I believe only the
neural accelerator should be custom-ish.

Additionally, I believe it having or not having Google IP is irrelevant:
for example, the new Raspberry Pi 5 Broadcom SoC has a lot of
Raspberry's own IP, but it's still called Broadcom as it's the real
manufacturer and designer of the chip.
That's a good argument. Indeed BCM2712 contains "New Raspberry
Pi-developed ISP".
https://www.raspberrypi.com/documentation/computers/processors.html

There aren't many patches but GPU is still called brcm,2712.

For Tesla FSD, there was discussion and output was not very consisting.
First, the name itself was used for everything - SoC architecture, one
given SoC and eventually the board.
https://lore.kernel.org/all/5ab62673-8d46-ec1d-1c80-696421ab69ca@xxxxxxxxxxxxx/

Eventually the last part - board - was renamed to "Evaluation board",
but I don't know how true or real it is.

See also:
"I would argue that if this SoC shares the pinctrl, clock, spi, adc,
and timer implementation
with Exynos, we should consider it part of the Exynos family,"
https://lore.kernel.org/all/CAK8P3a31bCHNcNWrLX+QW+4RuK=DBpxLA_j5BFKxXxXKCT8PFQ@xxxxxxxxxxxxxx/

However it was also claimed:

"AFA architecture is concerns both Exynos and FSD has completely
different architecture (at least at HW level)."
https://lore.kernel.org/all/07ce01d8091e$9a6fd9c0$cf4f8d40$@samsung.com/

I guess the same is also true for `axis,artpec8` and `tesla,fsd` SoCs.
IMO the SoC compatible string should be uniquely identifying the actual
SoC, not a close relative.

Regarding product_id you are correct this reads 0x09845000 but even
within Samsung Exynos family there are examples where the register
value does not match the SoC compatible. For example Exynos850 SoC
has a product ID value of "E3830". Where the Linux compatible is
matching the Samsung marketing name, not the internal/outdated name.
I did not know Exynos 850 is also not going under it's real name.
Ultimately, I believe all of those SoCs should go under their technical
name in the exynos/ directory.
The initial technical name does not exist outside of vendor sources and
part name. E.g. Winlink E850 board hardware manual calls it:
"Samsung Exynos 850, S5E3830"
and everywhere else Exynos 850 SoC is used.

If you start calling it Exynos 3830, only me and Sam (who mainlined it)
would know what is it. Everyone else, all users of kernel, would be
confused.

Therefore using well known final product name is for Exynos850 reasonable.
I agree with this. By using the final (well known) product SoC name -- gs101 --
other developers will be able to easily identify the particular SoC.

Another concern is that Google could in the future license other SoC: be
it Qualcomm, Nvidia or anything. If we put completely different hw under
google/ directory, does it really make sense? In that case, who'll
maintain the google/ directory? Exynos people? Qualcomm people if they
license it? Some other people?
I don't understand why the architecture of the SoC would dictate which folder
to put the device tree files under. It makes more sense to group board DT files
together based on who distributes them. Having all the Pixel DT board files
together allows Google to create a single device tree binary per SoC coupled
with the set of device tree overlays per board variant (this is the dtbo.img)
to ship to all their devices. If you look at all the in-market Pixel devices
with Tensor SoCs, you will find that you could create one dtb (concatenate
gs101.dtb, gs201.dtb, and zuma.dtb) and one dtbo image for 10 devices which
significantly simplifies the maintenance, testing, and software distribution
for all 10 of those products.

How is that relevant?

I believe it is none of the kernel concerns, it's up to the user to do 
whatever with the built .dtb files.

Also I do not see an issue in having a file list of all the .dtbo files 
you might want.

That's indeed a problem. Future Tesla SoC might have just few pieces
similar to FSD. There would be no common SoC part, except the actual
Tesla IP.

Same for Google. Future GSXXX, if done by Qualcomm, will be absolutely
different than GS101 and the only common part would be the TPU (Tensor).

So now let's decide what is the common denominator:
1. Core SoC architecture, like buses, pinctrl, clocks, timers, serial,
and many IP blocks, which constitute 95% of Devicetree bindings and drivers,
2. The one, big piece made by Samsung's customer: TPU, NPU or whatever.
As mentioned above, I think this should be based on how the DTBs and DTBOs are
used and distributed. What is the benefit of adding the gs101 board files under
the exynos folder?

One clear benefit would be the ease of maintaining all the SoC files at 
once. It's not that it is a benefit of having it in the Exynos folder, 
it's more like that there's no benefit in having a separate folder, and 
that also comes with some additional issues.

As I said earlier, it's pretty similar to the Raspberry Pi 5 example: It 
contains Raspberry's in-house IP, but it's still called properly 
Broadcom. The only difference is that Raspberry does not want its name 
on the chip, but Google does, despite it being just as custom as the 
Raspberry SoC is. The company's policy should not be a factor for this 
decision, in my opinion.

However as you've added, gs101 is the same thing as Exynos9845, so I 
believe there's no question that the Exynos name should be specified 
somewhere too, because this is what's literally wired in hardware, and 
not just a "well-known name that is used by Google in the Pixel factory 
kernel".

I agree though that just specifying the internal E9845 name could 
mislead some people, but GS101 is a similarly obscure name, and not even 
the real name of the hardware.

Thanks,
Will

Then, I don't think Tensor G3 has a proper "GS" name, it goes by "Zuma"
in decompiled kernel modules as far as I see.

Finally, Tesla people already tried to submit drivers called by Tesla
name, but which basically copied the functionality of the Exynos
drivers. We would want to avoid that, ideally.

My opinion is that all the Tesla and Google SoCs should be in the
exynos/ directory, not only because they are basically Samsung Exynos,
but also because they don't really need a separate directory: neither
Google nor Tesla didn't neither manufacture or design those SoCs from
scratch. The only reason I can think of for them to have it in a
separate directory is maybe because Google and Tesla actually paid
Samsung money for the right to call Exynos "Google designed" SoCs, but I
believe the kernel should be left out of that.
For some reason, although I know which, Cc-list is here trimmed and
misses Alim...

So standard reply follow (it makes me really, really grumpy, because it
means you develop on some crazy old kernel or do not use tools which
automate the process):

Please use scripts/get_maintainers.pl to get a list of necessary people
and lists to CC (and consider --no-git-fallback argument). It might
happen, that command when run on an older kernel, gives you outdated
entries. Therefore please be sure you base your patches on recent Linux
kernel.

Best regards,
Krzysztof

Yours,

Maksym