[Intel-gfx] [PATCH] drm/i915: Boost GPU clocks if we miss the pageflip's vblank

[Chris Wilson]

Quoting Ville Syrjälä (2017-08-22 18:02:04)
> On Mon, Aug 21, 2017 at 04:54:21PM +0100, Chris Wilson wrote:
> > Quoting Chris Wilson (2017-08-17 13:37:06)
> > > If we miss the current vblank because the gpu was busy, that may cause a
> > > jitter as the frame rate temporarily drops. We try to limit the impact
> > > of this by then boosting the GPU clock to deliver the frame as quickly
> > > as possible. Originally done in commit 6ad790c0f5ac ("drm/i915: Boost GPU
> > > frequency if we detect outstanding pageflips") but was never forward
> > > ported to atomic and finally dropped in commit fd3a40242e87 ("drm/i915:
> > > Rip out legacy page_flip completion/irq handling").
> > > 
> > > References: https://bugs.freedesktop.org/show_bug.cgi?id=102199
> > > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > > Cc: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> > > Cc: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > Cc: Daniel Vetter <daniel.vetter at intel.com>
> > 
> > Either of you like to ack the return of this code to the display
> > subsystem? It's still reactionary and will one day be replace by a pony,
> > or perhaps supplemented by one.
> 
> It looks reasonable enough to me.
> 
> For the pony part I was wondering if a blind donkey would be enough.
> Something like "boost to rpe as soon as a flip is queued" is what
> I was thinking. But I suppose it ought to be likely that we're
> already >= rpe if we have something running on the gpu. So maybe
> rpe just isn't fast enough for these cases?

The counterpoint is that even byt can decode a 1080p mp4 and show it at
near minimal clocks. So I feel any arbitrary boosting will run afoul of
power efficient hw (or at least fixed purpose doing just that). For the
interactivity detection, Ray was suggesting we listen to input events,
but at least we should push that coupling to userspace. My current
favourite remains granting boost privileges to a context so that when
such an interactive workload comes in, we boost (or we generalize that
with "desired clocks" on a context). We are not far then from having a
budget + deadline and the building blocks of a singing and dancing pony.
-Chris