[Intel-gfx] [PATCH 3/3] drm/i915: Tweak RPS thresholds to more aggressively downclock

Daniel Vetter daniel at ffwll.ch
Thu Oct 3 09:41:35 CEST 2013


On Tue, Oct 01, 2013 at 03:04:01PM -0700, Jesse Barnes wrote:
> On Wed, 25 Sep 2013 17:34:57 +0100
> Chris Wilson <chris at chris-wilson.co.uk> wrote:
> 
> > After applying wait-boost we often find ourselves stuck at higher clocks
> > than required. The current threshold value requires the GPU to be
> > continuously and completely idle for 313ms before it is dropped by one
> > bin. Conversely, we require the GPU to be busy for an average of 90% over
> > a 84ms period before we upclock. So the current thresholds almost never
> > downclock the GPU, and respond very slowly to sudden demands for more
> > power. It is easy to observe that we currently lock into the wrong bin
> > and both underperform in benchmarks and consume more power than optimal
> > (just by repeating the task and measuring the different results).
> > 
> > An alternative approach, as discussed in the bspec, is to use a
> > continuous threshold for upclocking, and an average value for downclocking.
> > This is good for quickly detecting and reacting to state changes within a
> > frame, however it fails with the common throttling method of waiting
> > upon the outstanding frame - at least it is difficult to choose a
> > threshold that works well at 15,000fps and at 60fps. So continue to use
> > average busy/idle loads to determine frequency change.
> > 
> > v2: Use 3 power zones to keep frequencies low in steady-state mostly
> > idle (e.g. scrolling, interactive 2D drawing), and frequencies high
> > for demanding games. In between those end-states, we use a
> > fast-reclocking algorithm to converge more quickly on the desired bin.
> > 
> > v3: Bug fixes - make sure we reset adj after switching power zones.
> > 
> > v4: Tune - drop the continuous busy thresholds as it prevents us from
> > choosing the right frequency for glxgears style swap benchmarks. Instead
> > the goal is to be able to find the right clocks irrespective of the
> > wait-boost.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Kenneth Graunke <kenneth at whitecape.org>
> > Cc: St├ęphane Marchesin <stephane.marchesin at gmail.com>
> > Cc: Owen Taylor <otaylor at redhat.com>
> > Cc: "Meng, Mengmeng" <mengmeng.meng at intel.com>
> > Cc: "Zhuang, Lena" <lena.zhuang at intel.com>
> > ---
> 
> It's a little scary to mess with these, but we've gotten some good
> numbers so far so I guess it's ok.
> 
> As a follow up, it might be nice to expose the power, balanced,
> performance profiles to userspace via sysfs.  Since we can't solve this
> problem for all users and all needs, we can just punt it out to
> userspace. :)
> 
> Reviewed-by: Jesse Barnes <jbarnes at virtuousgeek.org>

All three merged, thanks for patches&review. I've frobbed the first two
with tiny style bikesheds:
- Dropped the typedef usage for the plain struct drm_i915_private in new
  code.
- Dropped the extern qualifier for the function prototypes in header
  files.

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch



More information about the Intel-gfx mailing list