[Intel-gfx] [PATCH] drm/i915/bxt: Broxton doesn't use gen9 scaling for rps frequencies.

[Imre Deak]

On Wed, 2015-11-11 at 13:36 -0800, Bob Paauwe wrote:
> On Tue, 10 Nov 2015 11:04:22 +0200
> Mika Kuoppala <mika.kuoppala at linux.intel.com> wrote:
> 
> > Bob Paauwe <bob.j.paauwe at intel.com> writes:
> > 
> > > Signed-off-by: Bob Paauwe <bob.j.paauwe at intel.com>
> > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92768
> > > ---
> > >  drivers/gpu/drm/i915/intel_pm.c | 4 ++--
> > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > > index 647c0ff..fc5097f 100644
> > > --- a/drivers/gpu/drm/i915/intel_pm.c
> > > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > > @@ -7138,7 +7138,7 @@ static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
> > >  
> > >  int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
> > >  {
> > > -	if (IS_GEN9(dev_priv->dev))
> > > +	if (IS_GEN9(dev_priv->dev) && !IS_BROXTON(dev_priv->dev))
> > >  		return (val * GT_FREQUENCY_MULTIPLIER) /
> > >  GEN9_FREQ_SCALER;
> > 
> > Documentation disagrees with this patch. The units are 16.67Mhz
> > thus we should use 50/3.
> > 
> > -Mika
> 
> I'm not sure I trust the documentation in this case.  Elsewhere, in
> gen6_init_rps_frequencies() we use GEN9_FREQ_SCALER for SKL only, not for BXT.

On SKL the frequency _capability_ register uses 50MHz units, but the
frequency _request_ register uses 16.67MHz units. We store the frequency
limits in 16.67MHz units in rps.rp*_freq, hence the use of
GEN9_FREQ_SCALER in gen6_init_rps_frequencies(). When requesting a
frequency we only use GEN9_FREQ_SCALER to get the 16.67MHz constant
value.

On BXT both the capability and request registers use 16.67MHz units, so
we don't need to convert in gen6_init_rps_frequencies(), but we still
use GEN9_FREQ_SCALER when requesting the frequency to get the 16.67MHz
constant value.

Confusing like hell, but it works.

> I tried changing gen6_init_rps_frequencies() to use it for both SKL and
> BXT but that didn't resolve the problem when running the ps_rps igt
> test. So if we really should be using the 50/3 scale for BXT, there's
> another bug somewhere.

The problem is that atm we can't set the min/max frequencies via sysfs
to the same value that we read out from these same files, due to the
rounding down we do in the driver. Using round-to-closest is one way to
fix it I posted a patch to do this to the bug report.

> > >  	else if (IS_CHERRYVIEW(dev_priv->dev))
> > >  		return chv_gpu_freq(dev_priv, val);
> > > @@ -7150,7 +7150,7 @@ int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
> > >  
> > >  int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
> > >  {
> > > -	if (IS_GEN9(dev_priv->dev))
> > > +	if (IS_GEN9(dev_priv->dev) && !IS_BROXTON(dev_priv->dev))
> > >  		return (val * GEN9_FREQ_SCALER) / GT_FREQUENCY_MULTIPLIER;
> > >  	else if (IS_CHERRYVIEW(dev_priv->dev))
> > >  		return chv_freq_opcode(dev_priv, val);
> > > -- 
> > > 2.4.3
> > >
> > > _______________________________________________
> > > Intel-gfx mailing list
> > > Intel-gfx at lists.freedesktop.org
> > > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
> 
> 
>