Hi,
On 23.12.2021 18:11, Alyssa Rosenzweig wrote:
>> The kernel driver itself can't guess which jobs need a such a strict
>> affinity, so setting proper requirements is the responsibility of
>> the userspace (Mesa). However the userspace is not smart enough [yet].
>> Therefore this patch applies the above affinity rule to all jobs on
>> dual core group GPUs.
>
> What does Mesa need to do for this to work "properly"?
I don't know.
The blob restricts affinity of jobs with JD_REQ_COHERENT_GROUP requirement.
In theory jobs without such a requirement can run on any core, but in
practice all jobs in slots 0, 1 are assigned to core group 0 (with workloads
I've run - i.e. weston, firefox, glmark2, perhaps it's also SoC dependent).
So I've forced all jobs in slots 0, 1 to core group 0. Surprisingly this
(and memory attributes adjustment) appeared to be enough to get panfrost
working with T628 (on some SoCs). Without these patches GPU locks up in
a few seconds.
> What are the limitations of the approach implemented here?
Suboptimal performance.
1) There might be job chains which don't care about affinity
(I haven't seen any of these yet on systems I've got).
2) There might be dual core group GPUs which don't need such a strict affinity.
(I haven't seen any dual core group T[78]xx GPUs yet. This doesn't mean such
GPUs don't exist).
> If we need to extend it down the line with a UABI change, what would that look like?
I have no idea. And I'm not sure if it's worth the effort (since most jobs
end up on core group 0 anyway).
Best regards,
Alexey
