[Intel-gfx] [PATCH] drm/i915: Enable provoking vertex fix on Gen9 systems.

[Chris Wilson]

Quoting Joonas Lahtinen (2018-06-18 10:03:24)
> Quoting Chris Wilson (2018-06-15 22:06:05)
> > From: Kenneth Graunke <kenneth at whitecape.org>
> > 
> > The SF and clipper units mishandle the provoking vertex in some cases,
> > which can cause misrendering with shaders that use flat shaded inputs.
> > 
> > There are chicken bits in 3D_CHICKEN3 (for SF) and FF_SLICE_CHICKEN
> > (for the clipper) that work around the issue.  These registers are
> > unfortunately not part of the logical context (even the power context),
> > and so we must reload them every time we start executing in a context.
> > 
> > Bugzilla: https://bugs.freedesktop.org/103047
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> 
> Reviewed-by: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>

Thanks, fingers crossed we don't discover a reason why we shouldn't do
this for all contexts... Pushed.
 
> One note below.
> 
> > +++ b/drivers/gpu/drm/i915/intel_lrc.c
> > @@ -1575,11 +1575,21 @@ static u32 *gen9_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
> >         /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt,glk */
> >         batch = gen8_emit_flush_coherentl3_wa(engine, batch);
> >  
> > +       *batch++ = MI_LOAD_REGISTER_IMM(3);
> > +
> >         /* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl,glk */
> > -       *batch++ = MI_LOAD_REGISTER_IMM(1);
> >         *batch++ = i915_mmio_reg_offset(COMMON_SLICE_CHICKEN2);
> >         *batch++ = _MASKED_BIT_DISABLE(
> >                         GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE);
> > +
> > +       /* BSpec: 11391 */
> > +       *batch++ = i915_mmio_reg_offset(FF_SLICE_CHICKEN);
> > +       *batch++ = _MASKED_BIT_ENABLE(FF_SLICE_CHICKEN_CL_PROVOKING_VERTEX_FIX);
> > +
> > +       /* BSpec: 11299 */
> > +       *batch++ = i915_mmio_reg_offset(_3D_CHICKEN3);
> > +       *batch++ = _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_PROVOKING_VERTEX_FIX);
> 
> I'm almost betting that somebody will take one of these 3 off without
> noticing the distant LOAD_REGISTER_IMM(). To perfect this, const table
> of pairs and use ARRAY_SIZE()? Then we're also one step closer to the
> const W/A tables...

I'll be back later to tidy this up.
-Chris