[Intel-gfx] [PATCH] drm/i915: Make sample_c messages go faster on Haswell.

[Ville Syrjälä]

On Thu, Oct 30, 2014 at 10:32:38AM -0700, Kenneth Graunke wrote:
> On Thursday, October 30, 2014 01:01:30 PM Ville Syrjälä wrote:
> > On Thu, Oct 30, 2014 at 02:32:40AM -0700, Kenneth Graunke wrote:
> > > On Thursday, October 30, 2014 11:00:51 AM Ville Syrjälä wrote:
> > > > On Thu, Oct 30, 2014 at 10:50:03AM +0200, Ville Syrjälä wrote:
> > > > > On Wed, Oct 29, 2014 at 03:12:43PM -0700, Kenneth Graunke wrote:
> > > > > > Haswell significantly improved the performance of sampler_c 
> messages,
> > > > > > but the optimization appears to be off by default.  Later platforms
> > > > > > remove this bit, and apparently always enable the optimization.
> > > > > > 
> > > > > > Improves performance in "Counter Strike: Global Offensive" by 18%
> > > > > > at default settings on Iris Pro.  No Piglit regressions.
> > > > > 
> > > > > Nice. We need more bits like this ;)
> > > > > 
> > > > > > 
> > > > > > Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
> > > > > > ---
> > > > > >  drivers/gpu/drm/i915/i915_reg.h | 1 +
> > > > > >  drivers/gpu/drm/i915/intel_pm.c | 4 ++++
> > > > > >  2 files changed, 5 insertions(+)
> > > > > > 
> > > > > > diff --git a/drivers/gpu/drm/i915/i915_reg.h 
> > > b/drivers/gpu/drm/i915/i915_reg.h
> > > > > > index 77fce96..340821a 100644
> > > > > > --- a/drivers/gpu/drm/i915/i915_reg.h
> > > > > > +++ b/drivers/gpu/drm/i915/i915_reg.h
> > > > > > @@ -5952,6 +5952,7 @@ enum punit_power_well {
> > > > > >  #define  HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE    (1 << 6)
> > > > > >  
> > > > > >  #define HALF_SLICE_CHICKEN3		0xe184
> > > > > > +#define   HSW_SAMPLE_C_PERFORMANCE	(1<<9)
> > > > > >  #define   GEN8_CENTROID_PIXEL_OPT_DIS	(1<<8)
> > > > > >  #define   GEN8_SAMPLER_POWER_BYPASS_DIS	(1<<1)
> > > > > >  
> > > > > > diff --git a/drivers/gpu/drm/i915/intel_pm.c 
> > > b/drivers/gpu/drm/i915/intel_pm.c
> > > > > > index 7a69eba..50c72a7 100644
> > > > > > --- a/drivers/gpu/drm/i915/intel_pm.c
> > > > > > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > > > > > @@ -5736,6 +5736,10 @@ static void haswell_init_clock_gating(struct 
> > > drm_device *dev)
> > > > > >  	I915_WRITE(GEN7_GT_MODE,
> > > > > >  		   GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
> > > > > >  
> > > > > > +	/* Make sample_c messages faster. */
> > > > > 
> > > > > I found a name for it in the w/a database.
> > > > > 
> > > > > WaSampleCChickenBitEnable:hsw
> > > > > 
> > > > > Reviewed-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > > 
> > > > Oh actually it says palette won't work when this bit is on. I'm assuming
> > > > that's the texture palette. Do we have any use of that anywhere?
> > > 
> > > That's a good point.  3DSTATE_SAMPLER_PALETTE_LOAD and the A8P8/indexed 
> > > formats aren't used by Mesa or xf86-video-intel, but it looks like they 
> might 
> > > be used by libva.
> > > 
> > > Can someone confirm that libva does use the sampler palette?
> > > 
> > > If they do, what do we do about it?
> > 
> > I suppose the best option then would be to use an LRI from a batch,
> > which means the register would need to be added to the cmd parser
> > white list. This is one of the context saved registers so doing the
> > LRI just once per context should be enough.
> 
> I don't like that solution.  For one, it's impossible - you can't LRI from 
> userspace batches, even if you add it to the kernel command parser's 
> whitelist, because the hardware scanner is still enabled.  Given that I've 
> been waiting two years for this capability, I want to find a more immediate 
> solution.

Ah. I've somehow convinced myself the cmd parser might actually be doing
something besides just eating CPU cycles these days. But I guess not.

> 
> Another option is to have some sort of execbuf flag...maybe a 3D/Media "usage" 
> flag.  If set to 3D, write 0x6000200...if media, write 0x6000000.  Or 
> something specific.  I do hate adding more junk to the execbuf path, though.
> 
> Other ideas?

Fast vs. slow flag? :)

More seriously, one somewhat crappy option would be to initialize that
bit to 1 for all explicit contexts, and then have the kernel always turn
it off before executing something with the default context. It's not
unlike how we imagined the RS stuff would work since old userspace
doesn't know to turn RS off when using the default context.

But if we would do something like this, I think it would be nice if we
could just add it as a temporary hack and potentially drop it if we
manage to lose the restore inhibit flag and someone finishes the
cmd parser. So maybe Mesa could also try to set it from a batch,
which currently would be a nop or rejected by the cmd parser, but
when the cmd parser starts to work it might actually do something?

But always initializing it to 1 for all explicit context in the kernel
obviously requires that libva doesn't use an explicit context itself.
In case it does, another idea would be to just add a hack to the cmd
parser to emit the LRI on the ring when it enconters this register.
Not very pretty, but should work somewhat. It wouldn't allow mid-batch
changes to the register value though, but enough for setting it once
for each context in Mesa. And of course we'd still need the kernel to
turn it off before any default context will see it.

> 
> > Well that's assuming libva doesn't use the default context. I'm getting
> > another itch to drop the restore inhibit flag for default contexts.
> > That would actually make it possible to do these sort of things without
> > risking breakage to existing userspace. But I think Chris is going
> > scream unless the patch comes with performance data that shows it
> > doesn't hurt too much.
> 
> I suppose it wouldn't affect Mesa much, since we never use the default context 
> on Gen6+.  But otherwise I'd probably want to see the data, like Chris...

Yeah. I just have this feeling we're going to have to go there
eventually since there are plenty of registers in the context. Some of
those registers might have other interesting bits that certain users
might want to frob from batches. But I suppose we could always have the
kernel reset those too once they get added to the cmd parser white
list. Except on new platforms where the hw scanner itself has a whitelist,
so we'd have to reset them all basically, unless we want the cmd parser
to still scan every batch on those platforms.

-- 
Ville Syrjälä
Intel OTC