[PATCH] drm/xe/oa/uapi: Allow OA config's to have arbitrary names/id's
Umesh Nerlige Ramappa
umesh.nerlige.ramappa at intel.com
Fri Jul 12 19:48:38 UTC 2024
On Tue, Jul 09, 2024 at 06:09:27PM -0700, Dixit, Ashutosh wrote:
>On Tue, 09 Jul 2024 18:17:41 -0700, Ashutosh Dixit wrote:
>>
>> Added OA configs were previously identified by means of a uuid. Rather than
>> forcing userspace to use uuid, here we generalize this scheme to allow
>> userspace the flexibility to use whatever naming scheme (including uuid)
>> they may wish to use to identify and track OA config's. Names must still be
>> unique or will be rejected.
>
>To provide some additional context for the need for this patch: at present
>multiple UMD's can add OA configs. But these UMD's don't have a centralized
>place to manage OA config uuid's. Therefore, OA configs added by one UMD
>are incomprehensible to the other UMD (since they both use their own
>private uuid, even for the same OA config).
>
>So one solution to this problem is of course to have the UMD's coordinate
>and use global uuid's across all UMD's.
>
>If this is somehow not possible, originating OA config's do have another
>globally unique identifier, "metricset name" (such as RenderBasic,
>ComputeBasic, TestOA etc.). So if we allow UMD's to use this "metricset
>name" when adding OA configs to Xe KMD, UMD's will be able to identify an
>OA config added by a different UMD. So the "metricset name" can serve as a
>globally unique identifier for OA configs.
>
>This patch allows this second solution to be possible, without preventing
>the first solution.
Looks like this is related to vulkan performance query as well as some
support required in containers (for the same query or in general?). Can
we get some more details on how this Vulkan API uses the perf-configs
query and what's the relation with containers?
I am not sure we have any use case where different UMDs would want to
query the configs though, so not sure if we should spend any effort
thinking about that.
Regards,
Umesh
>
>Thanks.
>--
>Ashutosh
More information about the Intel-xe
mailing list