[Mesa-dev] [PATCH 27/27] i965: Enable the AMD_performance_monitor extension on Gen5+.
Kenneth Graunke
kenneth at whitecape.org
Mon Nov 18 20:10:32 PST 2013
On 11/18/2013 05:22 PM, Eric Anholt wrote:
> Kenneth Graunke <kenneth at whitecape.org> writes:
>
>> On 11/18/2013 10:33 AM, Eric Anholt wrote:
>>> Kenneth Graunke <kenneth at whitecape.org> writes:
>>>
>>>> XXX: Gen6+ needs to be predicated on register writes.
>>>>
>>>> our register write checking function doesn't work on Gen6.
>>>
>>> Even if you can just enable it on gen7, this series is:
>>>
>>> Reviewed-by: Eric Anholt <eric at anholt.net>
>>
>> Now I'm confused. I thought you and Carl found regressions in patch 3
>> (the tri-state ring enum patch), and that you basically NAK'd patch 04
>> because it adds code to BEGIN_BATCH.
>>
>> I had thought I needed to rewrite patch 4 before I could upstream this.
>> Please clarify.
>
> We found some slight flushing behavior change in patch 3, which we
> talked over and I thought you'd squashed in the fix for already (the
> missed true/false -> *_RING).
Right. I don't remember hearing conclusively that it was resolved.
> As far as patch 4: I'd almost always rather avoid BEGIN_BATCH overhead
> since we call it so much, but the last other solution we talked about
> (explicit ring switching) seemed like a scary maintenance problem
> because you wouldn't notice when you forgot to add a switch to render,
> since the ring's almost always in render already anyway.
Okay, I thought you NAK'd both solutions and were hoping I would come up
with something else. I guess we can go with this for now, and improve
later...
I did run OpenArena with INTEL_NO_HW=1 on my Ivybridge with master vs.
explicit ring switches, and saw no difference (n=9). Hardly conclusive,
but removing the existing implicit-flush overhead didn't seem to be much
of win.
--Ken
More information about the mesa-dev
mailing list