[Intel-gfx] [PATCH 6/8] drm/i915: reprogram NOA muxes on context switch when using perf
Lionel Landwerlin
lionel.g.landwerlin at intel.com
Wed Apr 25 14:35:42 UTC 2018
On 25/04/18 12:57, Chris Wilson wrote:
> Quoting Lionel Landwerlin (2018-04-25 12:45:19)
>> If some of the contexts submitting workloads to the GPU have been
>> configured to shutdown slices/subslices, we might loose the NOA
>> configurations written in the NOA muxes. We need to reprogram them at
>> context switch.
> On every single batchbuffer, no way. That's around a 33% latency hit,
> ignoring the cost of what's inside the wa_bb.
>
> If they are not context saved, userspace isn't going to be allowed to
> modify them. So why do we need this? If they not context saved, then the
> kernel needs to control them and doesn't need to reset around every
> batch, just the one's that want non-default (fancier if we handle
> preemption)?
> -Chris
>
These are neither context saved, nor power saved :(
I can't tell whether the GuC might schedule things behind our back.
But eventually that will happen anyway. So that's why I put the
reprogramming in per ctx wa.
Agreed, it sucks.
I'll look into the fancier things.
Thanks,
-
Lionel
More information about the Intel-gfx
mailing list