On Wed, May 10, 2017 at 04:09:34PM +0200, Michał Winiarski wrote:
> On Wed, May 10, 2017 at 04:47:50PM +0300, Mika Kuoppala wrote:
> > Oscar Mateo <oscar.ma...@intel.com> writes:
> >
> > > This allows userspace to shutdown slices at will for performance/power
> > > reasons
> > > (because it doesn't have a use for more slices).
> > >
> > > Cc: Dmitry Rogozhkin <dmitry.v.rogozh...@intel.com>
> > > Cc: Chris Wilson <ch...@chris-wilson.co.uk>
> > > Signed-off-by: Oscar Mateo <oscar.ma...@intel.com>
> > > ---
> > > drivers/gpu/drm/i915/intel_engine_cs.c | 11 +++++++++++
> > > 1 file changed, 11 insertions(+)
> > >
> > > diff --git a/drivers/gpu/drm/i915/intel_engine_cs.c
> > > b/drivers/gpu/drm/i915/intel_engine_cs.c
> > > index 402769d..17ff88d 100644
> > > --- a/drivers/gpu/drm/i915/intel_engine_cs.c
> > > +++ b/drivers/gpu/drm/i915/intel_engine_cs.c
> > > @@ -628,6 +628,7 @@ static int wa_ring_whitelist_reg(struct
> > > intel_engine_cs *engine,
> > > static int gen8_init_workarounds(struct intel_engine_cs *engine)
> > > {
> > > struct drm_i915_private *dev_priv = engine->i915;
> > > + int ret;
> > >
> > > WA_SET_BIT_MASKED(INSTPM, INSTPM_FORCE_ORDERING);
> > >
> > > @@ -673,6 +674,11 @@ static int gen8_init_workarounds(struct
> > > intel_engine_cs *engine)
> > > GEN6_WIZ_HASHING_MASK,
> > > GEN6_WIZ_HASHING_16x4);
> > >
> > > + /* Allow the UMD to configure their own power clock state */
> > > + ret = wa_ring_whitelist_reg(engine, GEN8_R_PWR_CLK_STATE);
> > > + if (ret)
> > > + return ret;
> > > +
> >
> > How will the different userspace clients coordinate with the slice
> > state? I guess in another words, is this part of the context?
> >
> > -Mika
> >
>
> Spec says (see IHD-OS-SKL-Vol2c-05.16 page 614):
> "This register must be initialized correctly when the context is submitted for
> the first time. This register is context save/restored as part of Exec-List
> context image in both Exec-List and Ring-Buffer mode of scheduling."
>
> So we're good, right?
>
> Except there's also:
> "This register must not be programmed using MI_LOAD_REGISTER_IMM command in
> ring
> buffer or in batch buffer, however programming "NON - SLM Indication" field
> through MI_LOAD_REGISTER_IMM is an exception defined below. If a need arises
> to
> change the render configuration for a context being executed in HW, Scheduler
> must preempt the context and update the desired render configuration in the
> logical render context image in memory and resubmit the context"
>
> So... are we sure that we're fine with giving userspace control over this
> register? If so - it would still be a good idea to mention that the info in
> the
> spec is bogus (commit message or just a comment).
Doesn't sound like it. So we are back to doing the preempt and changing
the context image as suggested in the other patch. Still we have the
issue that OA config conflicts with changing the slice eu config.
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
