[Intel-gfx] [PATCH 1/1] i915/query: Correlate engine and cpu timestamps with better accuracy

[Jason Ekstrand]

On April 30, 2021 23:01:44 "Dixit, Ashutosh" <ashutosh.dixit at intel.com> wrote:
> On Fri, 30 Apr 2021 19:19:59 -0700, Umesh Nerlige Ramappa wrote:
>>
>> On Fri, Apr 30, 2021 at 07:35:41PM -0500, Jason Ekstrand wrote:
>>> On April 30, 2021 18:00:58 "Dixit, Ashutosh" <ashutosh.dixit at intel.com>
>>> wrote:
>>>
>>> On Fri, 30 Apr 2021 15:26:09 -0700, Umesh Nerlige Ramappa wrote:
>>>
>>> Looks like the engine can be dropped since all timestamps are in sync.
>>> I
>>> just have one more question here. The timestamp itself is 36 bits.
>>> Should
>>> the uapi also report the timestamp width to the user OR should I just
>>> return the lower 32 bits of the timestamp?
>>>
>>> Yeah, I think reporting the timestamp width is a good idea since we're
>>> reporting the period/frequency here.
>>
>> Actually, I forgot that we are handling the overflow before returning the
>> cs_cycles to the user and overflow handling was the only reason I thought
>> user should know the width. Would you stil recommend returning the width in
>> the uapi?
>
> The width is needed for userspace to figure out if overflow has occured
> between two successive query calls. I don't think I see this happening in
> the code.

Right... We (UMDs) currently just hard-code it to 36 bits because that's 
what we've had on all platforms since close enough to forever. We bake in 
the frequency based on PCI ID. Returning the number of bits, like I said, 
goes nicely with the frequency. It's not necessary, assuming sufficiently 
smart userspace (neither is frequency), but it seems to go with it. I guess 
I don't care much either way.

Coming back to the multi-tile issue we discussed internally, I think that 
is something we should care about. Since this works by reading the 
timestamp register on an engine, I think leaving the engine specifier in 
there is fine. Userspace should know that there's actually only one clock 
and just query one of them (probably RCS). For crazy multi-device cases, 
we'll either query per logical device (read tile) or we'll have to make 
them look like a single device and sync the timestamps somehow in the UMD 
by carrying around an offset factor.

As is, this patch is

Reviewed-by: Jason Ekstrand <jason at jlekstrand.net>

I still need to review the ANV patch before we can land this though.

--Jason
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-gfx/attachments/20210501/96d61ed7/attachment.htm>