Re: [PATCH v3 3/5] media: dt-bindings: media: Add binding for the Raspberry Pi HEVC decoder

[Krzysztof Kozlowski <krzk@kernel.org>]

On 23/06/2025 19:54, Dave Stevenson wrote:
>>
>>> +          - brcm,bcm2711-hevc-dec
>>> +          - brcm,bcm2712-hevc-dec
>>>
>>>> You referred to file "raspberrypi,pisbe.yaml" before, but there is no
>>>> such file in the next.
>>>
>>> Typo.
>>> https://github.com/torvalds/linux/blob/master/Documentation/devicetree/bindings/media/raspberrypi%2Cpispbe.yaml
>>> Reviewed by Rob only just over a year ago [2]
>>
>> There were some discussions and reasons with explanations.
>> Feel free to use same arguments and accurately describe the hardware, so
>> we won't be needing to ask usual questions.
> 
> Sorry if the previous descriptions hadn't been clear. All the same
> points as for raspberrypi,pispbe.yaml apply here, which is why I'd
> tried to link to it originally.
> 
> Describe the hardware:
> Raspberry Pi designed and are the sole owners of the IP for this HEVC
> decoder block. This is *not* Broadcom IP.
> 
> That design was given to Broadcom as Verilog to wire into the
> bus/interrupt/clock fabric of BCM2711, and to manufacture it via TSMC
> on a 28nm process node.
> 
> A few years later the same design was given to Broadcom to wire into
> BCM2712, and to manufacture it on a 16nm process node.
> 
> As it is Raspberry Pi owned IP it can be used in other places than
> Broadcom SoCs.
> Broadcom does not have a licence to use the IP in any other of their chips.
> 
> It is the same situation as for raspberrypi,pispbe.yaml except the
> block already exists in 2 chips rather than just the 1. There also
> isn't a version register in the hardware that is different between
> those chips (an oversight noted for future chips).
> It could be compared to a Synopsis or Cadence IP block dropped into an
> SoC. The vendor prefix happens to be raspberrypi instead of snps or
> cdns.
> 
> Is there any part that needs to be further clarified?

Dunno, thread is way too old to be in my inbox. Anyway, it is part of
the SoC as you said, so SoC compatibles rules apply here. It's the same
for Synopsys.

> 
>> Anyway, we asked for fallback, so you need items, just like every SoC
>> compatible (see also example schema).

Just implement the feedback.

>>
>>
>>>
>>>> Before you reply that there is a binding using different rules: well,
>>>> there is always poor code. Above two comments are repeated, especially
>>>> this about specific compatible - all the time, so these are not new
>>>> rules. These are given in reviews since some years.
>>>
>>> My Google-foo is totally failing with the only directly relevant
>>> mention of "fallback compatible" I find is [3], which just says to use
>>> them.
>>>
>>> You're effectively saying I can't take anything in the kernel tree as
>>> being a valid example as it could be poor code, and a layman such as
>>> myself has no way of telling.
>>
>> No, I am saying that argument "I saw someone doing this, so I am allowed
>> to do the same" is not correct. There are good and bad examples. For
>> example in my talks I mentioned good examples. The list of good examples
>> was not accepted to the kernel so well... I just use as an example any
>> recent Qcom binding using specific compatibles.
>>
>>> Could you please point me at documentation and examples I can rely on,
>>> or educate me with what is wanted in this situation to avoid me having
>>> to guess?
>>>
>>> A further mailing list search has brought up [4] which is a thread
>>> with yourself from 2 years ago which looks to be a very similar
>>> situation. Other than missing the const on the SoC strings (although
>>> that isn't in the merged version of cnm,wave521c.yaml), and two SoC
>>> specific strings, I'm not seeing an obvious difference between there
>>> and here either.
>>
>> How is the [4] relevant? That's IP block from other vendor.
> 
> It was another example of an IP block belonging to vendor A being
> added into an SoC from vendor B, same as the situation here. Raspberry
> Pi != Broadcom.

People hide real explanation, real facts, so don't use one example for
other case. I was told *explicitly* that Raspberry ISP or whatetever it
was called, was NOT PART OF the SoC. Now you said:
"A few years later the same design was given to Broadcom to wire into
BCM2712, and to manufacture it on a 16nm process node."


Best regards,
Krzysztof