Re: Qualcomm Kryo core compatibles

From: Konrad Dybcio
Date: Fri Mar 10 2023 - 19:03:22 EST




On 10.03.2023 23:39, Rob Herring wrote:
> On Wed, Mar 8, 2023 at 4:44 AM Konrad Dybcio <konrad.dybcio@xxxxxxxxxx> wrote:
>>
>> Hi!
>>
>> I was recently debating what to do about Qualcomm Kryo compatibles.
>>
>> There are basically 3 cases:
>>
>> 1. Falkor/"real Kryo" - the (never shipped?) server platform & MSM8996
>>
>> This one's easy, it's actually Kryo so it should stay Kryo.
>>
>>
>> 2. Fake Kryo ("customized" Arm Cortex cores) (MSM8998-SM8x50)
>>
>> This one's tough.. Qualcomm marketing material seems to sometimes say
>> Cortex, sometimes Kryo, sometimes "customized Cortex".. They do use
>> their own arm IMPLEMENTER_ID in the MIDR_EL1 register and their
>> PART_NUM values are not Arm-stock, but these cores don't seem to be
>> any special.. Maybe some irq lines are routed differently? Not sure.
>>
>> My proposition here is to do:
>>
>> "qcom,kryoXXX", "arm,cortex-ABC"
>>
>> or
>>
>> "qcom,kryoXXX-PQR", "arm,cortex-ABC"
>
> I don't see much value in the fallback here. We don't do much with the
> values anyways as everything uses ID registers anyways. Do you know
> the level of modification?
Sadly no..

>
>> where PQR is one of:
>> - silver (LITTLE cores)
>> - gold (big cores)
>> - gold_plus (prime core(s))
>>
>>
>> 3. Arm cores modified within Arm implementation-defined allowance (SC8280XP+)
>>
>> These cores report Arm IMPLEMENTER_IDs and actual Arm PART_NUMs, which would
>> suggest they're bone stock Arm Cortex cores, with some Qualcomm-iness coming
>> as part of implementation details which are.. expected since Cortex allows for
>> some IMPLEMENTATION DEFINED things. The only non-obvious part here is that
>> the REVISION field they report does not always seem covered by the Arm TRMs.
>>
>> In this case I think going with
>>
>> "arm,cortex-ABC"
>>
>> is fine.. I already did this for 8550 and 8280xp and Rob seems to have liked it.
>>
>> So, I suppose the real question is what to do about 2., should they stay as
>> they are, or maybe my proposition seems attractive?
>
> What about the generic 'qcom,kryo' strings?
As I pointed out in 1., IMO it'd be most logical to only use them for
"true kryo" cores, which are a custom armv8.1-compliant design not
based on Cortex, used on MSM8996 and some dead server platform. Any other
"kryo" seems to either fall under 2. or 3.

Konrad

>
> Rob