Re: [PATCH 10/16] drm/i915: Improve watermark dirtyness checks

["Ville Syrjälä" <ville.syrjala@linux.intel.com>]

On Fri, Oct 11, 2013 at 12:02:53PM -0300, Paulo Zanoni wrote:
> 2013/10/9  <ville.syrjala@linux.intel.com>:
> > From: Ville Syrjälä <ville.syrjala@linux.intel.com>
> >
> > Currently hsw_write_vm_values() may write to certain watermark
> > registers needlessly. For instance it will disable LP1+ watermarks
> > around WM_PIPE changes even though that's not needed. Also if only,
> > say, LP3 changes, the current code will again disable all LP1+
> > watermarks even though only LP3 needs to be reconfigured.
> >
> > Add an easy to read function that will compute the dirtyness of the
> > watermarks, and use that information to further optimize the watermark
> > programming.
> 
> Clever solution to the problem!
> 
> Some comments below.
> 
> 
> >
> > Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
> > ---
> >  drivers/gpu/drm/i915/intel_pm.c | 95 +++++++++++++++++++++++++++++++++--------
> >  1 file changed, 77 insertions(+), 18 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > index bed96fb..5bd8c73 100644
> > --- a/drivers/gpu/drm/i915/intel_pm.c
> > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > @@ -2776,6 +2776,63 @@ static struct intel_pipe_wm *hsw_find_best_result(struct drm_device *dev,
> >         }
> >  }
> >
> > +/* dirty bits used to track which watermarks need changes */
> > +#define WM_DIRTY_PIPE(pipe) (1 << (pipe))
> > +#define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
> > +#define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
> > +#define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
> > +#define WM_DIRTY_FBC (1 << 24)
> > +#define WM_DIRTY_DDB (1 << 25)
> > +
> > +static unsigned int ilk_compute_wm_dirty(struct drm_device *dev,
> > +                                        const struct hsw_wm_values *old,
> > +                                        const struct hsw_wm_values *new)
> > +{
> > +       unsigned int dirty = 0;
> > +       enum pipe pipe;
> > +       int wm_lp;
> > +
> > +       for_each_pipe(pipe) {
> > +               if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
> > +                       dirty |= WM_DIRTY_LINETIME(pipe);
> > +                       /* Must disable LP1+ watermarks too */
> > +                       dirty |= WM_DIRTY_LP_ALL;
> > +               }
> > +
> > +               if (old->wm_pipe[pipe] != new->wm_pipe[pipe])
> > +                       dirty |= WM_DIRTY_PIPE(pipe);
> 
> I think this one also needs WM_DIRTY_LP_ALL. I don't think changing
> level zero without stopping the LP levels is safe.

Dunno. Spec doesn't tell me to do it. Although I haven't even bothered
to check how hard it would be to cook up a configuration where just
WM_PIPE changes but WM_LPx stays the same.

Actually the spec doesn't even say that we should disable a particular
LP1+ watermark if we want to change it. It just says that you have to
disable and enable them in order.

> 
> 
> > +       }
> > +
> > +       if (old->enable_fbc_wm != new->enable_fbc_wm) {
> > +               dirty |= WM_DIRTY_FBC;
> > +               /* Must disable LP1+ watermarks too */
> > +               dirty |= WM_DIRTY_LP_ALL;
> > +       }
> > +
> > +       if (old->partitioning != new->partitioning) {
> > +               dirty |= WM_DIRTY_DDB;
> > +               /* Must disable LP1+ watermarks too */
> > +               dirty |= WM_DIRTY_LP_ALL;
> > +       }
> > +
> > +       /* LP1+ watermarks already deemed dirty, no need to continue */
> > +       if (dirty & WM_DIRTY_LP_ALL)
> > +               return dirty;
> > +
> > +       /* Find the lowest numbered LP1+ watermark in need of an update... */
> > +       for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
> > +               if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
> > +                   old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
> > +                       break;
> > +       }
> > +
> > +       /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
> > +       for (; wm_lp <= 3; wm_lp++)
> > +               dirty |= WM_DIRTY_LP(wm_lp);
> > +
> > +       return dirty;
> > +}
> > +
> >  /*
> >   * The spec says we shouldn't write when we don't need, because every write
> >   * causes WMs to be re-evaluated, expending some power.
> > @@ -2784,6 +2841,7 @@ static void hsw_write_wm_values(struct drm_i915_private *dev_priv,
> >                                 struct hsw_wm_values *results)
> >  {
> >         struct hsw_wm_values previous;
> > +       unsigned int dirty;
> >         uint32_t val;
> >
> >         previous.wm_pipe[0] = I915_READ(WM0_PIPEA_ILK);
> > @@ -2804,31 +2862,32 @@ static void hsw_write_wm_values(struct drm_i915_private *dev_priv,
> >
> >         previous.enable_fbc_wm = !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
> >
> > -       if (memcmp(results, &previous, sizeof(*results)) == 0)
> 
> Oh, you've just removed the memcmp I complained about... Still, I
> believe it had a bug.
> 
> 
> > +       dirty = ilk_compute_wm_dirty(dev_priv->dev, &previous, results);
> > +       if (!dirty)
> >                 return;
> >
> > -       if (previous.wm_lp[2] != 0)
> > +       if (dirty & WM_DIRTY_LP(3) && previous.wm_lp[2] != 0)
> >                 I915_WRITE(WM3_LP_ILK, 0);
> > -       if (previous.wm_lp[1] != 0)
> > +       if (dirty & WM_DIRTY_LP(2) && previous.wm_lp[1] != 0)
> >                 I915_WRITE(WM2_LP_ILK, 0);
> > -       if (previous.wm_lp[0] != 0)
> > +       if (dirty & WM_DIRTY_LP(1) && previous.wm_lp[0] != 0)
> >                 I915_WRITE(WM1_LP_ILK, 0);
> >
> > -       if (previous.wm_pipe[0] != results->wm_pipe[0])
> > +       if (dirty & WM_DIRTY_PIPE(PIPE_A))
> >                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
> > -       if (previous.wm_pipe[1] != results->wm_pipe[1])
> > +       if (dirty & WM_DIRTY_PIPE(PIPE_B))
> >                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
> > -       if (previous.wm_pipe[2] != results->wm_pipe[2])
> > +       if (dirty & WM_DIRTY_PIPE(PIPE_C))
> >                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
> >
> > -       if (previous.wm_linetime[0] != results->wm_linetime[0])
> > +       if (dirty & WM_DIRTY_LINETIME(PIPE_A))
> >                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
> > -       if (previous.wm_linetime[1] != results->wm_linetime[1])
> > +       if (dirty & WM_DIRTY_LINETIME(PIPE_B))
> >                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
> > -       if (previous.wm_linetime[2] != results->wm_linetime[2])
> > +       if (dirty & WM_DIRTY_LINETIME(PIPE_C))
> >                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
> >
> > -       if (previous.partitioning != results->partitioning) {
> > +       if (dirty & WM_DIRTY_DDB) {
> >                 val = I915_READ(WM_MISC);
> >                 if (results->partitioning == INTEL_DDB_PART_1_2)
> >                         val &= ~WM_MISC_DATA_PARTITION_5_6;
> > @@ -2837,7 +2896,7 @@ static void hsw_write_wm_values(struct drm_i915_private *dev_priv,
> >                 I915_WRITE(WM_MISC, val);
> >         }
> >
> > -       if (previous.enable_fbc_wm != results->enable_fbc_wm) {
> > +       if (dirty & WM_DIRTY_FBC) {
> >                 val = I915_READ(DISP_ARB_CTL);
> >                 if (results->enable_fbc_wm)
> >                         val &= ~DISP_FBC_WM_DIS;
> > @@ -2846,18 +2905,18 @@ static void hsw_write_wm_values(struct drm_i915_private *dev_priv,
> >                 I915_WRITE(DISP_ARB_CTL, val);
> >         }
> >
> > -       if (previous.wm_lp_spr[0] != results->wm_lp_spr[0])
> > +       if (dirty & WM_DIRTY_LP(1) && previous.wm_lp_spr[0] != results->wm_lp_spr[0])
> >                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
> 
> As far as I understood, if previous.wm_lp_spr[X] !=
> results->wm_lp_spr[X], then for sure "dirty & WM_DIRTY_LP(X)", right?
> So the "dirty" check would be redundant here. And we can't remove the
> second part because we can have "dirty & WM_DIRTY_LP(X) while the
> sprite values don't change.

Right, because there's no separate enable bit for the sprite LP1+
watermarks, the dirty check is not really needed here. You want me to
remove it?

> 
> > -       if (previous.wm_lp_spr[1] != results->wm_lp_spr[1])
> > +       if (dirty & WM_DIRTY_LP(2) && previous.wm_lp_spr[1] != results->wm_lp_spr[1])
> >                 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
> > -       if (previous.wm_lp_spr[2] != results->wm_lp_spr[2])
> > +       if (dirty & WM_DIRTY_LP(3) && previous.wm_lp_spr[2] != results->wm_lp_spr[2])
> >                 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
> >
> > -       if (results->wm_lp[0] != 0)
> > +       if (dirty & WM_DIRTY_LP(1) && results->wm_lp[0] != 0)
> >                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
> > -       if (results->wm_lp[1] != 0)
> > +       if (dirty & WM_DIRTY_LP(2) && results->wm_lp[1] != 0)
> >                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
> > -       if (results->wm_lp[2] != 0)
> > +       if (dirty & WM_DIRTY_LP(3) && results->wm_lp[2] != 0)
> >                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
> >
> >         dev_priv->wm.hw = *results;
> > --
> > 1.8.1.5
> >
> > _______________________________________________
> > Intel-gfx mailing list
> > Intel-gfx@lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
> 
> 
> 
> -- 
> Paulo Zanoni

-- 
Ville Syrjälä
Intel OTC