[RFC 0/3] Dynamic SSEU mashup
Navik, Ankit P
ankit.p.navik at intel.com
Fri Jan 10 04:06:01 UTC 2020
Hi Tvrtko,
Thank you for taking time to work on this.
> -----Original Message-----
> From: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
> Sent: Thursday, January 9, 2020 10:24 PM
> To: intel-gfx-trybot at lists.freedesktop.org
> Cc: Navik, Ankit P <ankit.p.navik at intel.com>; Ursulin, Tvrtko
> <tvrtko.ursulin at intel.com>
> Subject: [RFC 0/3] Dynamic SSEU mashup
>
> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>
> Hi Ankit,
>
> A few points with these patches:
>
> They are rough, unfinished and altogether bad. However I promised you
> something this week so please give them a spin to see how perf and power looks
> on your system. They survived some smoke testing on my end and did not seem
> to regress gfxbench3.
>
> I got bogged down with request counting so I changed the approach completely
> and instead of counting outstanding requests I am counting completed requests
> per context and per engine class. Maybe it's the same - answers the question on
> how active is the context.
>
> I changed the heuristic to look at a sliding window weighted average of
> completed requests. This is because when I looked at the previous logic, its
> behaviour looked completely wrong to me. It was constantly selecting lowest
> load type for many heavy workloads.
>
> But I am not happy at all with my weighted past history. I am pretty sure that on
> top we would need to force periodic re-configurations (context re-pin).
>
> Even with all these changes I'm afraid to say I started to doubt in any request
> counting heuristic even more than I have before. I am very curious what your
> testing will show, but I don't know how we could do this properly without maybe
> hooking into OA so we can know (rather than guess) how wide the workloads
> are.
>
> Many opens.. but as said, I ran out of time to make this any better this week.
> Please test and we will see.
Yes, Let me test it and I'll share results soon with it.
Regards,
Ankit
>
> Ankit Navik (1):
> drm/i915: set optimum eu/slice/sub-slice configuration based on load
> type
>
> Tvrtko Ursulin (2):
> drm/i915: Track per context/class rq completion frequency
> drm/i915: Record if context has user SSEU configuration applied
>
> drivers/gpu/drm/i915/gem/i915_gem_context.c | 14 +++++
> .../gpu/drm/i915/gem/i915_gem_context_types.h | 17 ++++++
> drivers/gpu/drm/i915/gt/intel_context.c | 52 +++++++++++++++++
> drivers/gpu/drm/i915/gt/intel_context_types.h | 5 ++
> drivers/gpu/drm/i915/gt/intel_lrc.c | 22 +++++--
> drivers/gpu/drm/i915/gt/intel_sseu.c | 46 +++++++++++++++
> drivers/gpu/drm/i915/gt/intel_sseu.h | 4 ++
> drivers/gpu/drm/i915/i915_drv.h | 5 ++
> drivers/gpu/drm/i915/i915_sysfs.c | 34 +++++++++++
> drivers/gpu/drm/i915/intel_device_info.c | 57 ++++++++++++++++++-
> 10 files changed, 250 insertions(+), 6 deletions(-)
>
> --
> 2.20.1
More information about the Intel-gfx-trybot
mailing list