[Mesa-dev] [PATCH] i965/gen8: Enable hiz for all depth levels

[Eero Tamminen]

Hi,

I ran large set of performance tests on BDW i3/GT2 & SKL i5/GT2.
Impact of the patch is within variance (variance in the test setup is
on SKL <1% for GPU limited cases, on BDW, which is TDP limited,
variance is ~3x larger).  I.e. no noticeable improvement in them,
but neither there is no regression.

Patch is so simple, that I'd recommend just pushing it in, on
average HiZ should improve performance more than regress it.


	- Eero

On 01.02.2016 20:46, Ben Widawsky wrote:
> On Thu, Jan 28, 2016 at 05:31:43PM -0800, Ben Widawsky wrote:
>> From: Jordan Justen <jordan.l.justen at intel.com>
>>
>> After modifying the hiz buffer allocation and qpitch calculation, hiz
>> appears to work in all cases on gen8.
>>
>> v2 by Ben: add GEN8_WM_HZ_FULL_SURFACE_DEPTH_CLEAR. From docs:
>> "If the depth buffer clear operation does clear the entire width and height of
>> the surface, then the “full surface clear” bit in 3DSTATE_WM_OP must be set to
>> 1."
>> I'm very surprised that no regressions were spotted originally.
>>
>> Signed-off-by: Jordan Justen <jordan.l.justen at intel.com> (v1)
>> Reviewed-by: Topi Pohjolainen <topi.pohjolainen at intel.com> (v1)
>> Signed-off-by: Ben Widawsky <benjamin.widawsky at intel.com> (v2)
>> ---
>>
>> I am currently running perf data on this patch, but we might not see an
>> improvement (Jordan didn't last time around). However, I think using more HiZ is
>> general a good thing, and we should merge the patch unless we have perf
>> regressions.
>>
>> Jenkins shows no regressions except on BXT where I have no data.
>
> I ran the perf data on this (BDW GT3). In my admittedly limited data set, I
> don't really see any difference. As you pointed out today, this is optimizing
> the less interesting case since the feature is about bandwidth savings, and
> we're already doing this for LOD0 (arguably the level with the highest savings).
>
> So, I think we should probably merge this patch given that I have no jenkins
> regressions, but, it might be nice to collect perf data on HSW, and SKL as well,
> just to make sure. *shrug*
>
> [snip]
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>