Re: [PATCH v4 4/8] soc: mediatek: mtk-cmdq: Add pa_base parsing for unsupported subsys ID hardware

[AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>]

Il 05/03/25 16:58, Jason-JH Lin (林睿祥) ha scritto:
> On Wed, 2025-03-05 at 12:03 +0100, AngeloGioacchino Del Regno wrote:
>>
>> External email : Please do not click links or open attachments until
>> you have verified the sender or the content.
>>
>>
>> Il 05/03/25 10:46, Jason-JH Lin (林睿祥) ha scritto:
>>> On Tue, 2025-03-04 at 10:35 +0100, AngeloGioacchino Del Regno
>>> wrote:
>>>>
>>>> External email : Please do not click links or open attachments
>>>> until
>>>> you have verified the sender or the content.
>>>>
>>>>
>>>> Il 18/02/25 06:41, Jason-JH Lin ha scritto:
>>>>> When GCE executes instructions, the corresponding hardware
>>>>> register
>>>>> can be found through the subsys ID. For hardware that does not
>>>>> support
>>>>> subsys ID, its subsys ID will be set to invalid value and its
>>>>> physical
>>>>> address needs to be used to generate GCE instructions.
>>>>>
>>>>> This commit adds a pa_base parsing flow to the cmdq_client_reg
>>>>> structure
>>>>> for these unsupported subsys ID hardware.
>>>>>
>>>>
>>>> Does this work only for the MMINFRA located GCEs, or does this
>>>> work
>>>> also for
>>>> the legacy ones in MT8173/83/88/92/95 // MT6795/6893/etc?
>>>>
>>>> In order to actually review and decide, I do need to know :-)
>>>>
>>>
>>> Yes, it's for the SoCs without subsys ID, it's not related to the
>>> MMINFRA.
>>>
>>> This can also work on MT8173/83/92/95 // MT6795/6893/etc.
>>> You can remove the `mediatek,gce-client-reg` properties in their
>>> dtsi
>>> and cherry-pick this series to verify it. :-)
>>>
>>
>> This is curious - and that brings more questions to the table (for
>> curiosity
>> more than anything else at this point).
>>
>> Since this is a way to make use of the CMDQ for address ranges that
>> are not tied
>> to any subsys id (hence no gce-client-reg and just physical address
>> parsing for
>> generating instructions), do you know what are the performance
>> implications of
>> using this, instead of subsys IDs on SoCs that do support them?
>>
> 
> The main advantage of using subsys ID is to reduce the number of
> instruction.
> Without subsy ID, you will need one more `ASSIGN` instruction to assign
> the high bytes of the physical address.
> 
> E,g. In mt8195-gce.h: #define SUBSYS_1c00XXXX 3
> 
> If you want GCE to write the value 0x0000000f to 0x1c00_002c.
> 
> With subsys ID, you can use only one instruction to achieve it:
> 1. WRITE value: 0x000000f to subsys: 0x3 + offset: 0x0002c
> - OP code: WRTIE = 0x90
> - subsys ID: 0x1c00XXXX = 0x03
> - offset: 0x002c
> - value: 0x0000000f
> 
> Without subsys ID, you will need 2 instructions to achieve it:
> 1. ASSIGN address high bytes: 0x1c00 to GCE temp register: SPR0
> - OP code: LOGIC = 0xa0
> - arg_type: register, value, value = (0x8)
> - sub OP: ASSIGN = 0x0
> - register index to store the assign value: SPR0 = 0x0
> - value to assign: 0x1c00
> 2. WRITE value: 0x0000000f to temp register: SPR0 + offset:0x002c
> - OP code: WRITE = 0x90
> - sub OP(temp register index): SPR0 = 0x0
> - offset for temp register: 0x002c
> - value: 0x0000000f
> 
>> Being clear: if we were to migrate a SoC like MT8195 to using this
>> globally
>> instead of using subsys ids, would the performance be degraded?
>> And if yes, do you know by how much?
>>
> 
> E,g. If the inst number with subsys ID is N.
> 1. If CMDQ is implement like this, then inst number will be (N * 2):
> assign SPR0 = 0x1c00
> write A to SPR0 + offset: 0x2c
> assign SPR0 = 0x1c00
> write B to SPR0 + offset: 0x3c
> assign SPR0 = 0x1c00
> write C to SPR0 + offset: 0x4c
> ...
> 
> 2. If CMDQ is implement like this, the inst number will be (N + 1 * n):
> assign SPR0 = 0x1c00
> write A to SPR0 + offset: 0x2c
> write B to SPR0 + offset: 0x3c
> write C to SPR0 + offset: 0x4c
> ...
> 
> When the same cmd buffer changes the base address for n times:
> assign SPR0 = 0x1c00
> write A to SPR0 + offset: 0x2c
> assign SPR0 = 0x1c01
> write B to SPR0 + offset: 0x2c
> assign SPR0 = 0x1c02
> write C to SPR0 + offset: 0x2c
> assign SPR0 = 0x1c00
> write D to SPR0 + offset: 0x3c
> ...
> 
> So you can imagine the performance will increase, but maybe not too
> much if we use it in the right way...
> Except the old SoC that didn't support SPR and CPR. The reason will be
> addressed in the next paragraph.
> 
>> What you're proposing almost looks like being too good to be true -
>> and makes
>> me wonder, at this point, why the subsys id was used in the first
>> place :-)
>>
> 
> That's because of the old GCE version in the old SoC only support GPR,
> it didn't support SPR and CPR.
> 
> GPR:
> All 32 GCE threads share the same GPR0~GPR15, GPR will be affected by
> other GCE threads if they use it at the same time.
> 
> SPR:
> Each GCE thread has 4 SPR, SPR won't be affected by another GCE thread.
> 
> CPR:
> All 32 GCE threads share the same CPR, there are over 1000 CPR can be
> used. It need to be managed properly to avoid the resource conflicting.
> 
> Due to the GPR resource restriction in the old GCE version, the usage
> of subsys ID can avoid GPR conflicting issues when multiple GCE threads
> are using GPR to physical assign high bytes all the time.
> 
> 
> I have simplified some complicate instruction rules, so the description
> above may not be 100% matched to the CMDQ helper driver code.
> But I think the main concept is correct.
> Hope these explanation can help well :-)
> 

Jason, that's simply wonderful. Thanks a lot for this precious description.

Yes, this has clarified even more than I was asking for, and besides,
yeah I know that there are some rules to follow, some of which I know,
some of which I imagine - and describing all of that would need lots
and lots of text - again, I know, and no worries about that! :-D

Thanks again!
Angelo

> Regards,
> Jason-JH Lin
> 
>> Cheers!
>> Angelo
>>
>>> Regards,
>>> Jason-JH Lin
>>>
>>>> Thanks,
>>>> Angelo
>>>>
>>>>
>>>
>>
>>
>>
>