On 04/02/2019 13:22, Chris Wilson wrote:
SW PMU reports semaphore time as busy, HW PMU reports semaphore time as
idle. Who is correct?
[It's not really HW PMU, it's a different implementation of the SW PMU. :)]
As an additional data point, HW tracking of accumulated total context
runtime as stored in the PPHWSP also reports semaphore spin time
(polling mode) as context running.
So overall from the point of view of busy being opposite of idle, it is
kind of correct. Regardless of whether engine is doing something useful
or not. It is unavailable for other contexts due some action of the
currently executing context.
In this light we could view busy as aggregate of busy and semaphore
want. (MI_WAIT_EVENT is an open.) But there is indeed an inconsistency
on platforms which cannot do context tracking.
Therefore solution a) add semaphore wait time to busy when reporting
busy on those platforms.
Advantage - PMU sampling timer is already running on these platform so
additional cost is small.
From the point of view of wanting to make busy mean "useful" work, that
seems much harder.
Option b) could be subtract semaphore wait time from busy, on the other
set of platforms.
Disadvantage - this would mean running the PMU sampling timer when it
today doesn't need to.
So I am leaning towards option a). Engine busy time semantics would
therefore be defined as engine not being idle = occupied by a context
doing something.
Regards,
Tvrtko
---
drivers/gpu/drm/i915/intel_lrc.c | 1 -
1 file changed, 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index ae90ce034252..d00b268ed6ee 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -2308,7 +2308,6 @@ void intel_execlists_set_default_submission(struct intel_engine_cs *engine)
engine->unpark = NULL;
engine->flags |= I915_ENGINE_HAS_SEMAPHORES;
- engine->flags |= I915_ENGINE_SUPPORTS_STATS;
if (engine->i915->preempt_context)
engine->flags |= I915_ENGINE_HAS_PREEMPTION;
}
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
